From Italy and the European Union to the United Kingdom, the United States, China and Russia: artificial intelligence, administrative efficiency, institutional accountability and the political economy of time.
Scope: A comparative assessment of how governments introduce artificial intelligence into public administration, who exercises authority over the resulting efficiency gains, and whether technological acceleration produces measurable improvements in public services, institutional capacity and fiscal performance.
Executive Summary — Bottom Line Up Front
The decisive question in the adoption of artificial intelligence by governments is no longer simply how much work an algorithm can perform, but who has the authority to allocate the human capacity that automation releases, how that allocation is governed, and whether citizens receive a demonstrable benefit.
The available official evidence establishes five findings.
- Italy has entered a more operational phase of AI governance. Law No. 132 of 23 September 2025, the 2024–2026 public-administration digital plan and AgID’s AI adoption guidelines of 9 November 2026 provide a developing institutional framework. The unresolved economic question concerns measurement of end-to-end service performance rather than formal adoption alone.
- France is testing AI within government workflows. A 2026 evaluation reports that 75% of participating public employees considered its AI assistant useful, while earlier experiments documented reductions in response times in specific services.
- Germany is pursuing AI-assisted administrative processes alongside digital sovereignty and workforce development. Federal reporting describes applications intended to accelerate permitting and other procedures, but their results require careful distinction between local pilots and system-wide productivity.
- The United Kingdom has produced unusually concrete public-sector evidence: a cross-government trial involving more than 20,000 civil servants reported average savings of 26 minutes per working day. A separate Department for Work and Pensions evaluation estimated approximately 19 minutes, under a different methodology.
- The United States, China and Russia illustrate different institutional approaches to AI governance, ranging from federal agency autonomy within centrally established safeguards to more centrally coordinated digital transformation.
The central analytical judgment is that time saved at the task level is a potential input to public-sector productivity, not evidence that productivity has already increased. The conversion depends on workflow redesign, staffing and managerial authority, data governance, legal obligations, demand, service quality and the availability of capacity at subsequent stages.
Governments that cannot measure this conversion risk reporting technological success without demonstrating a corresponding improvement in administrative outcomes. Conversely, governments that can document it gain a stronger basis for budgeting, organisational reform, workforce development and democratic accountability.
Who Controls the Time Saved by Artificial Intelligence? The New Economics of Government Productivity
From Italy’s administrative reforms to Washington’s federal agencies and Beijing’s centralised infrastructure, the decisive question is no longer how quickly AI performs government work, but who controls the capacity it releases—and whether taxpayers receive the benefit.
Artificial intelligence is exposing a contradiction at the centre of government modernisation: public employees can complete tasks faster while administrative services remain just as slow and public expenditure continues to rise. Italy’s AI Law No. 132 of 23 September 2025 and the adoption guidelines issued by the Agenzia per l’Italia Digitale on 9 November 2026 give that contradiction immediate institutional significance. Across Europe, the United States and China, governments are investing in systems capable of reducing routine work without necessarily possessing the organisational mechanisms to convert those savings into completed services. The decisive economic question is who has the authority to allocate the working time AI releases, and whether that authority can produce measurable public value rather than another layer of technological expenditure. The answer will determine the fiscal credibility of government AI investment and the future organisation of public employment.
Italy Has Regulated AI. It Must Now Account for the Time It Saves
Italy has moved from general principles toward an operational framework for artificial intelligence in government. Law No. 132/2025, effective from 10 October 2025, established national provisions governing AI, while AgID Determination No. 169 of 9 November 2026 adopted guidelines for AI deployment in public administration. The guidance addresses organisational maturity, risk management, cybersecurity, data quality and professional competencies. It creates a basis for more disciplined implementation, but does not demonstrate that administrative procedures have become faster.
That distinction exposes the weakness of conventional digitalisation reporting. An AI system may reduce the time required to prepare a legal opinion, examine documentation or draft an administrative response. Yet the file may remain in a queue awaiting verification, consultation or formal authorisation. The employee becomes more efficient; the administration does not necessarily become more productive. Italy’s implementation framework can govern how technology enters public offices, but the economic return depends on subsequent managerial decisions concerning staffing, workflows and the allocation of recovered capacity.
The constraint is also legal. Article 14 of Law No. 132/2025 preserves the centrality of human responsibility in public administration. A faster AI-generated assessment cannot replace the professional judgement and procedural safeguards required for a consequential decision. Under Legislative Decree No. 165/2001 and the applicable public-employment arrangements, changing duties and reallocating personnel involves institutional authority, not simply the installation of software. Italy’s productivity dividend will therefore be determined as much by administrative management as by technological performance.
Europe’s Numbers Reveal the Gap Between Faster Tasks and Better Services
The United Kingdom provides the clearest illustration of both the potential and the limits of AI productivity claims. A Government Digital Service experiment conducted between September and December 2024 involved approximately 20,000 civil servants across 12 organisations and reported average self-assessed savings of 26 minutes per participant per working day. A separate Department for Work and Pensions evaluation, published on 29 January 2026, examined a trial involving 3,549 employees and estimated an adjusted saving of 19 minutes daily, with a 95% confidence interval of 17–22 minutes.
These are significant operational findings, but they are not interchangeable. The studies used different populations and evaluation methods, and neither established that every minute recovered became additional public-service output or reduced Treasury expenditure. The DWP evaluation also found estimated savings of 26 minutes in information searches, 25 minutes in email writing and 24 minutes in summarisation. Such task-specific effects cannot simply be added together to calculate an employee’s total daily productivity gain.
France provides a second perspective. Its government evaluated an AI assistant introduced to 10,000 public employees across six ministerial groups, reporting in June 2026 that 75% of surveyed participants considered the tool useful and 65% would recommend it. Experimental exercises identified approximately 12% time savings in writing tasks and up to 16% in document summarisation. Another 56% of respondents reported reducing their professional reliance on non-sovereign AI alternatives, connecting employee productivity to the government’s interest in controlling its technological environment.
Germany is pursuing a more process-oriented approach. Federal initiatives reported in November 2026 included AI-assisted planning and permitting applications and an expansion phase involving 38 municipalities and 17 technology companies. Early pilot results indicated substantial acceleration in selected administrative procedures. Their significance lies in the attempt to improve workflows rather than isolated employee activities, although results from individual municipalities cannot establish nationwide productivity.
The distinction matters for public accounts. A 20-minute daily saving has economic value only when its subsequent use can be identified. It may reduce backlogs, expand service capacity, improve verification or create avoidable expenditure. Without that evidence, converting employee time savings into national fiscal forecasts risks confusing potential labour capacity with money available to the government.
Washington’s AI Expansion Exposes the Cost of Weak Institutional Capacity
The United States demonstrates how rapidly AI adoption can expand before the corresponding administrative benefits become measurable. On 3 April 2025, the Office of Management and Budget issued Memoranda M-25-21, governing accelerated federal AI use and institutional safeguards, and M-25-22, addressing efficient government acquisition, competition, interoperability and supplier dependence. The framework assigns substantial implementation responsibility to federal agencies while establishing common expectations for procurement and governance.
The Government Accountability Office reported on 29 July 2025 that AI use cases across 11 selected federal agencies increased from 571 in 2023 to 1,110 in 2024. Generative-AI use cases rose from 32 to 282. Yet 84 of the 282 generative-AI cases remained in their initial phase, and another 75 were in acquisition or development. The inventory documented technological expansion, not an equivalent increase in completed public services.
A subsequent GAO investigation into the Internal Revenue Service, published on 24 March 2026, exposed a more consequential constraint. The agency’s June 2025 inventory contained 126 active AI use cases, but the audit identified deficiencies in workforce planning, inventory quality and strategic management. One relevant research and analytics group had lost 63 employees working full- or part-time on AI, contributing to concerns about the technical capacity required to develop and sustain applications.
The contradiction is financially important. Government can purchase software intended to reduce labour requirements while simultaneously losing the specialist personnel needed to integrate, verify and maintain it. AI procurement therefore creates a continuing obligation to fund computing, data management, cybersecurity, technical expertise and supplier support. Without those complementary capabilities, a larger technology inventory can coexist with weaker institutional readiness.
Beijing Centralises Infrastructure; Moscow Sets Digital Targets. Neither Proves an AI Dividend
China has chosen a different implementation structure. Its State Council’s AI+ Action, issued in 2025, established development objectives extending to 2027, 2030 and 2035. The programme included adoption ambitions exceeding 70% by 2027 and 90% by 2030 for specified intelligent terminals and agents. These are broad technology-diffusion targets, not measurements of government productivity.
More directly relevant to public administration were the Guidelines for the Deployment and Application of Large AI Models in Government Affairs, issued by the Cyberspace Administration of China and the National Development and Reform Commission on 10 October 2025. The document promotes shared computing infrastructure, common government AI platforms and the reuse of higher-level resources by subordinate administrations. County-level authorities and lower levels are encouraged, in principle, to avoid unnecessary independent development of large government models.
The economic logic is straightforward: shared infrastructure can reduce duplicated investment, concentrate expertise and support coordinated security arrangements. But centralisation introduces its own dependencies. Administrative offices may become reliant on common platforms that do not fully accommodate specialised requirements, while interruption of a shared service can affect multiple institutions simultaneously. China’s guidance explicitly warns against duplicative systems and ineffective digital formalism, recognising that technological deployment can generate additional bureaucracy rather than reduce it.
Russia’s Government Order No. 637-r of 16 March 2024 establishes a broader digital-administration strategy extending to 2030. Its targets include 100 government services capable of producing immediate online results and full electronic document-management coverage across specified local-government and public institutions. The strategy emphasises common information infrastructure, administrative interoperability and technological independence. Yet these are digital-government targets, not verified AI-specific productivity gains.
The comparison reveals different approaches to the same economic constraint. Washington relies extensively on agency implementation within federal guidance; Beijing promotes shared deployment infrastructure; Moscow sets central digital-transformation objectives. None of these institutional models eliminates the need to measure the actual cost, reliability and public-service consequences of automation.
The Fiscal Dividend Is Not the Value of Employees’ Saved Minutes
The fiscal consequences become clearer when the accounting categories are separated. Gross time saved, usable administrative capacity and cash expenditure avoided are three different economic quantities. A government employee who saves 20 minutes per working day does not automatically generate an equivalent reduction in payroll expenditure. The employee may instead process additional applications, assist more citizens or perform more demanding verification. Those outcomes can represent real public value, but they do not release the same amount of cash for other budgetary purposes.
The United Kingdom’s Digital and Data Benefits Framework, published on 7 April 2026, illustrates the scale of prospective gains and the risks of interpreting them prematurely. It discusses an estimated £23.7 billion in potential benefits across healthcare, education and policing, comprising £13.6 billion in healthcare, £7.5 billion in education and £2.6 billion in policing under specified task-automation assumptions. These figures represent estimated possibilities, not realised Treasury savings.
HM Treasury’s Green Book 2026, published on 28 October, reinforces the distinction between public-sector financial effects and wider social value. Its appraisal framework requires attention to costs, benefits, uncertainty, distributional consequences and optimism bias. For government AI, this means accounting for licences, integration, computing infrastructure, employee training, security, maintenance and eventual supplier replacement alongside the benefits attributed to automation.
The workforce question is inseparable from this accounting problem. Recovered time can be reinvested in additional services, professional development or improved decision quality. Alternatively, managers can attempt to convert it into higher output requirements or reduced expenditure. Each choice has different consequences for employees, institutional capacity and citizens. Premature staffing reductions can eliminate competencies that future AI deployments require, as the IRS audit illustrates. Retaining every existing procedure unchanged, however, can allow technological expenditure to accumulate without improving public services.
The European AI Act adds a further constraint. Following the 2026 amendments, specified Annex III high-risk requirements are scheduled to apply from 2 December 2027, while requirements concerning specified product-related high-risk systems follow on 2 August 2028. Compliance, human oversight and information governance impose operational requirements that cannot be dismissed as avoidable friction when they protect legally consequential decisions.
The economic objective is therefore not to minimise human involvement indiscriminately. It is to reduce unnecessary administrative work while preserving the professional judgement, technical resilience and legal safeguards that determine whether a public decision is valid.
By 2028, the Cost of Inaction Will Fall on Budgets, Employees and Service Users
Over the 12–24 months following November 2026, governments will face increasingly concrete choices about which AI deployments to expand, which processes to redesign and which claimed benefits to recognise in public accounts. Italy must translate its November 2026 AgID adoption framework into operational evaluation. France and Germany must determine whether experimental improvements can survive wider deployment. The United Kingdom must connect its quantified employee-level savings with documented service and fiscal outcomes. American agencies must reconcile expanding AI inventories with the technical skills and governance capacity required to operate them.
The immediate cost of inaction is not necessarily technological obsolescence. It is the accumulation of recurring procurement expenditure without a corresponding improvement in completed services. Departments may continue purchasing AI capabilities while retaining the same queues, approval bottlenecks and fragmented information systems. Employees may face higher production expectations without adequate training or clear responsibilities. Citizens and businesses may continue bearing the economic cost of administrative delays even after the public sector reports substantial time savings.
The alternative is measurable but institutionally demanding. Public managers must identify where AI reduces active work, determine whether that improvement affects the capacity of the complete procedure and document the allocation of recovered time. Financial authorities must distinguish avoided expenditure from additional service capacity. Audit and supervisory functions must assess the reliability of reported outcomes without substituting for managerial responsibility.
The decisive benchmark for 2027–2028 will not be the number of government AI licences, models or pilot projects. It will be whether administrations can demonstrate shorter completion times, more valid services, sustainable operating costs and accountable use of the capacity released by automation.
If they cannot, taxpayers will finance the technology, employees will absorb its organisational consequences and service users will continue paying for delays that artificial intelligence was supposed to reduce.
Navigational Index
I – Institutional
Pillar I — The Governance of Time and Public-Sector Productivity
- Chapter 1. Italy: From Individual Efficiency to Administrative Performance
- Chapter 2. European Governance: EU Law, France, Germany and the United Kingdom
- Chapter 3. Who Owns Administrative Productivity? Authority, Labour and Accountability
II – Comparative
Pillar II — International Models and the Economics of AI Deployment
- Chapter 4. The United States: Federal AI Governance and Agency Performance
- Chapter 5. China and Russia: State Coordination, Administrative Modernisation and Technological Sovereignty
- Chapter 6. Shadow AI, Human Capital, Cybersecurity and the Institutional Value of Time
III – Decisions
Pillar III — Measurement, Strategic Choices and the 2031 Outlook
- Chapter 7. Measuring the Conversion of Time Savings into Public Value
- Chapter 8. Fiscal Implications, Workforce Reallocation and Institutional Risks
- Chapter 9. Strategic Scenarios, Early-Warning Indicators and the Final Net Assessment
Master Abstract
The productivity paradox inside the modern state
Artificial intelligence is creating a new distinction in the economics of public administration: the difference between accelerating an activity and improving an institution’s ability to fulfil its statutory responsibilities. A government employee may complete a document, analyse a regulation or prepare a briefing significantly faster with an AI assistant, without reducing the time required for a permit, benefit determination, judicial support process or administrative response to reach its final recipient. The reason is structural. Public services operate through connected stages, with formal authorisations, information dependencies, professional responsibilities and legally mandated controls. Improving one stage does not automatically increase the capacity of the entire process.
This creates an increasingly consequential governance issue. The economic benefit of AI is determined not only by model performance but by decisions about the use of recovered working time. Management may allocate that capacity to backlogs, additional cases, better quality assurance, more demanding investigations, professional training or shorter service queues. It may also absorb the gain through additional reporting, higher output targets or unchanged administrative routines. None of these outcomes follows inevitably from the technology itself. Each reflects choices about public value, institutional priorities, labour organisation and risk.
The question is particularly significant in Europe, where AI adoption intersects with administrative law, data-protection obligations, public procurement, workforce rights and a developing European regulatory framework. Italy’s newly adopted operational guidance makes this institutional problem immediate rather than theoretical. France and the United Kingdom provide evidence from employee trials; Germany is examining workflow automation and public-sector technological autonomy. The United States, China and Russia offer additional perspectives on central coordination, agency-level discretion and the relationship between state capacity and control over digital infrastructure.
The OECD’s September 2025 assessment of 200 public-sector AI examples across 11 core government functions provides an important comparative baseline. It recognises opportunities in service delivery, administrative processes, fraud detection and decision support, while identifying legacy systems, skills shortages, data limitations, trust and accountability as significant constraints. It does not establish a universal conversion rate between employee time savings and improved public services. OECD+1
The principal conclusion is therefore institutional: the authority to use time recovered through AI should be made explicit, while the effects of that decision should be measured at service level. By 2031, the meaningful distinction between effective and ineffective AI adoption will depend less on the number of deployed tools than on demonstrable changes in service completion times, administrative capacity, decision quality, accessibility and the public resources required to produce those outcomes.
Key Evidence Table — What Governments Have Actually Established
The following indicators distinguish actual measurements from administrative targets, experimental findings and formal policy commitments. They should not be combined into a single international productivity ranking because they concern different activities, populations, observation periods and evaluation methods.
| Indicator | Value or status | Reference date | Definition and scope | Issuer | Official source |
|---|---|---|---|---|---|
| Italian AI legislation | Law No. 132/2025, effective 10 October 2025 | 2025 | National AI legal framework | Italian Republic | Law No. 132/2025 — Normattiva |
| Italian administrative AI guidelines | Determination No. 169/2026 | 9 Nov 2026 | Adoption of AI guidance for public administrations | AgID | Determination No. 169 — AgID |
| Italian Digital Decade roadmap | 67 measures; €62.3 billion total planned budget | 2025 report | Overall national digital roadmap, not a dedicated public-sector AI budget | European Commission | Italy Digital Decade Report — European Commission |
| French public-sector AI trial | 10,000 participating employees across six ministries | Jun 2026 | AI assistant pilot | DITP | L’Assistant IA Evaluation — DITP |
| French AI task savings | Approximately 12% in writing and up to 16% in document summarisation | Jun 2026 | Experimental focus-group measurements, not whole-service gains | DITP | L’Assistant IA Evaluation — DITP |
| UK cross-government trial | 26 minutes saved per participant per working day | Jun 2025 | Trial involving over 20,000 civil servants; self-reported outcomes | Government Digital Service | Cross-Government Findings — GOV.UK |
| UK Department for Work and Pensions | 19 minutes per day estimated | Jan 2026 | Adjusted comparative analysis of the Copilot trial | DWP | DWP Copilot Evaluation — GOV.UK |
| US federal generative-AI use cases | 32 in 2023; 282 in 2024 | Jul 2025 | Reported use cases across 11 selected agencies | GAO | GAO-25-107653 — GAO |
| German administrative AI | Operational pilots and federal AI-modernisation initiatives | Nov 2026 | Workflow automation and administrative decision support | Federal Government | Federal AI Policy FAQ — Bundesregierung |
| Chinese public-sector AI | Government deployment guidance issued | Oct 2025 | Deployment and governance of large language models in government | CAC and NDRC | Government LLM Deployment Guidance — CAC |
| Russian digital administration | Strategic transformation direction to 2030 | Mar 2024 | Digital government services, administrative infrastructure and automation | Russian Government | Government Order No. 637-r — Government of Russia |
Sources are official legislative, governmental or intergovernmental records. The figures represent distinct measures of deployment, employee experience, task efficiency or policy commitments. They do not constitute comparable national productivity rates.
The International Evidence: Different States, Different Definitions of Success
Italy — An operational framework without an established national productivity dividend
Italy’s institutional development is particularly important because its regulatory and digital-transformation arrangements increasingly address AI as an organisational instrument rather than an isolated technological experiment. The adoption of AgID’s guidelines on 9 November 2026 provides a new reference point for public administrations implementing AI systems. The guidelines sit alongside the national legislation of September 2025 and the updated Three-Year Plan for ICT in Public Administration.
The consequences for administrative governance extend beyond compliance. Public administrations require defined responsibilities for selecting applications, authorising data use, supervising outputs and evaluating whether new systems improve established procedures. These arrangements also affect the distribution of efficiency gains. An administrative director can introduce AI-assisted drafting, but improving the resulting service may require changes to the allocation of personnel, coordination between departments, the design of digital case-management systems or the operation of subsequent approval stages.
Italy’s Digital Decade roadmap provides useful context, but its stated overall budget of €62.3 billion across 67 measures must not be interpreted as expenditure on AI adoption in government. Similarly, improvements in digital service maturity cannot automatically be attributed to artificial intelligence. The European Commission’s June 2025 country assessment recognised progress in Italy’s digital public services and infrastructure while identifying continuing difficulties in AI adoption. These findings concern different areas of digital development and cannot establish the productivity impact of public-sector AI deployment. Shaping Europe’s digital future
The immediate Italian governance question is consequently whether administrations can produce paired measurements of internal workload and completed service outcomes. A credible assessment would examine the time required to prepare individual documents, the time a case spends waiting for approval, total service completion time, the proportion of cases requiring correction and the ability to process additional demand without reducing the quality of decisions.
For Italy, the strategic significance of AI is therefore inseparable from administrative organisation. The technology can create capacity, but responsibility for converting it into service performance remains with the institutions authorised to govern the process.
France — Moving from employee adoption to measured administrative performance
France provides a more developed experimental record concerning how public employees interact with AI assistants. Its June 2026 evaluation covered a programme involving 10,000 public-sector employees across six ministries. According to the Direction interministérielle de la transformation publique, 75% of respondents considered the assistant useful for their work, 65% would recommend it to colleagues and 56% reported reducing their professional use of non-sovereign alternatives.
The experimental focus groups found approximately 12% time savings on writing tasks and up to 16% on document summarisation, accompanied by improvements in output quality under blind assessment. Those figures refer to measured task-level effects, not equivalent reductions in the duration of entire administrative procedures. DITP+1
France also has relevant evidence linking AI-supported work to a service-facing indicator. In a programme reported in December 2023, public-service employees used AI to assist responses to citizens’ online feedback. The government reported that average response time in participating services fell from seven days to three days after the initial experimental period. That observation is more directly relevant to citizen experience than a measure of document preparation time, although the published institutional account does not establish that AI alone caused the entire reduction. DITP
These two sets of findings illustrate why different performance measures must coexist. Employee acceptance establishes whether a system is usable. Task-level measurement establishes whether it improves particular activities. Service-level measurement establishes whether citizens or businesses experience the benefit. A government requires all three before drawing strong conclusions about administrative productivity.
Germany — Administrative acceleration and technological sovereignty
Germany’s policy approach places substantial emphasis on the organisation of AI-enabled administrative processes and the technological conditions under which government systems operate.
In its official explanation published on 2 November 2026, the Federal Government described AI applications designed to accelerate planning and permitting processes, including the SPARK initiative and the development of agentic AI applications with municipalities and technology firms. The government also reported local pilot results involving care-related applications and data-protection requests, while indicating a further phase involving 38 municipalities and 17 AI start-ups and scale-ups. These figures demonstrate the scale of the programme’s next phase, not nationwide achieved productivity. Bundesregierung
Germany’s approach also raises an important second-order consideration: the freedom to reallocate working time may depend on the freedom to control the systems that create it. Where public institutions cannot inspect important system components, replace suppliers, govern sensitive data or maintain administrative continuity, operational efficiency can be accompanied by increased technological dependence.
A December 2025 statement by the federal digital ministry, referring to research on large language models in federal administration, emphasised internally operated, replaceable and open-source-based solutions as a means of supporting digital sovereignty. This describes the federal assessment of the examined systems, rather than proving technological independence throughout German public administration. bmds.bund.de
The German experience therefore broadens the productivity question from labour time alone to the infrastructure and institutional control necessary to make AI-assisted services reliable, legally sustainable and adaptable.
United Kingdom — Stronger task-level evidence and a clear measurement challenge
The United Kingdom provides unusually detailed public evidence concerning employee time savings. Its cross-government Microsoft 365 Copilot experiment involved over 20,000 civil servants and reported an average reduction of 26 minutes per participant per working day. More than 70% of participants agreed that the system reduced time spent searching for information or performing routine tasks and allowed greater attention to strategic work. GOV.UK+1
A separate evaluation by the Department for Work and Pensions, published on 29 January 2026, used comparison-group data and statistical adjustments. It estimated approximately 19 minutes of daily savings in its fully adjusted specification, with a reported 95% confidence interval of 17–22 minutes. These results should not be averaged with the cross-government findings: the populations, study designs and underlying methods differ. GOV.UK+1
The UK’s April 2026 Digital and Data Benefits Framework also provides an institutional mechanism for distinguishing benefits. Its estimates of wider public-sector AI and automation potential concern prospective economic opportunities, not cash savings already delivered to the Treasury. In particular, the framework discusses a previous estimate of £23.7 billion across health, education and policing under specified automation assumptions. These figures should not be converted into established fiscal returns. GOV.UK
The United Kingdom illustrates both the value and the limits of quantitative evidence. Minutes saved are measurable. Their conversion into greater institutional output, better decisions, shorter queues or lower expenditure remains a further proposition requiring separate verification.
European Union — Common obligations, nationally organised productivity
The European Union provides a common regulatory framework while member states retain their own administrative structures, employment arrangements and service-delivery responsibilities.
Article 4 of the EU AI Act establishes AI-literacy obligations for providers and deployers. The European Commission’s current guidance explains the application of these obligations and their subsequent amendments in 2026. It also distinguishes general organisational literacy measures from requirements concerning human oversight of high-risk AI systems. Shaping Europe’s digital future+1
This legal architecture does not itself prescribe how every administration must reinvest time saved by AI. The allocation of staff, modification of operating procedures, implementation of service targets and adoption of workforce measures remain dependent on the applicable public-law and employment frameworks.
Consequently, a common European legal regime does not imply identical productivity outcomes. An Italian municipality, a French ministry and a German federal agency can operate under related EU requirements while having materially different staffing arrangements, digital infrastructure, procurement capacity, operational autonomy and decision-making procedures.
The principal opportunity for European coordination lies in establishing comparable performance definitions and auditable evaluation methods without treating radically different administrative functions as interchangeable.
United States — Agency deployment, federal safeguards and fragmented outcomes
The United States offers evidence of rapid growth in reported federal AI adoption, though adoption counts are not equivalent to productivity gains.
The Government Accountability Office reported in July 2025 that, across 11 selected agencies, total AI use cases rose from 571 in 2023 to 1,110 in 2024, while generative-AI use cases increased from 32 to 282. These are agency-reported use cases within the selected sample, not a comprehensive measure of every federal institution or the economic value generated. U.S. Government Accountability Office (GAO)
The Office of Management and Budget’s Memorandum M-25-21, issued on 3 April 2025, established guidance for accelerating federal AI use while retaining applicable safeguards concerning privacy, civil rights and civil liberties. It placed agency innovation within a government-wide framework of responsibilities and risk management. The White House+1
The American model creates a distinction between central policy and agency implementation. The federal government can set priorities and governance requirements, while individual agencies must determine where AI fits their missions, operating processes and statutory obligations.
The question of time allocation therefore becomes a matter of managerial authority, congressional appropriations, service objectives and accountability for public funds. Increasing the number of AI deployments demonstrates technological diffusion; it does not determine whether the gains are converted into additional service capacity, improved enforcement, reduced administrative burdens or measurable fiscal savings.
China — Central coordination and administrative capacity
China’s State Council issued its AI+ Action policy in August 2025, setting out an extensive programme for the integration of artificial intelligence across major economic and social sectors, including public governance. The document establishes strategic objectives extending to 2027, 2030 and 2035, with explicit ambitions for AI-enabled services and administrative processes. cac.gov.cn
In October 2025, the Cyberspace Administration of China and the National Development and Reform Commission issued guidance for deploying large language models in government. The guidance addressed public services, administrative work, social governance and decision support. It also called for shared infrastructure, clearer operational boundaries, lifecycle management, safeguards for sensitive information and the avoidance of isolated or duplicative government AI systems. cac.gov.cn
China’s policy framework places substantial weight on coordinated infrastructure, administrative standardisation and the reuse of technical capabilities across departments and levels of government.
This approach could reduce some forms of duplication and facilitate coordinated deployment. Nevertheless, central coordination alone does not establish realised efficiency improvements, nor does it demonstrate how recovered labour capacity is distributed between service improvements, additional administrative tasks and other institutional priorities.
The official record examined supports a strong conclusion about policy direction and governance design, but not a comparable national estimate of public-sector working time saved.
Russia — Digital sovereignty, central infrastructure and the limits of published productivity evidence
Russia’s government approved a strategic direction for the digital transformation of public administration through Order No. 637-r of 16 March 2024, extending to 2030.
The framework addresses electronic public services, administrative information exchange, shared government platforms, document management, automated processing and technological independence. It contains numerical targets for digital services and administrative systems, including the availability of results from 100 public services online by the end of 2030. These are strategic targets, not evidence that the corresponding outcomes have already been achieved. static.government.ru
The Russian framework is relevant because it treats administrative digitalisation as part of a wider programme of state infrastructure, security and information management. Its emphasis differs from employee-centred experimental evidence of the kind reported in France and the United Kingdom.
The reviewed official strategic document does not establish a nationwide causal estimate of how much working time AI has saved Russian public employees, what proportion of that capacity has been reassigned, or how much end-to-end service duration has declined because of AI specifically.
The defensible comparative conclusion is therefore limited: Russia has an official digital-government transformation framework, but the evidence reviewed is insufficient for a direct numerical comparison with Western European AI productivity trials.
Competing Pathways — How Governments May Use the Productivity Dividend
The official evidence supports several distinct pathways for the allocation of capacity recovered through AI. They are not mutually exclusive: an administration may pursue several simultaneously, and the appropriate combination depends on its statutory responsibilities, organisational structure and operational constraints.
| Pathway | Institutional mechanism | Potential public benefit | Principal risk | Decisive indicator |
|---|---|---|---|---|
| Service acceleration | Reallocate capacity to delayed cases and downstream processing | Shorter waiting and completion times | New bottlenecks at later stages | End-to-end service duration |
| Capacity expansion | Process additional cases with existing staffing | Greater service throughput | Excessive workloads or quality deterioration | Completed cases per staff-hour |
| Quality improvement | Allocate time to verification, complex cases and individual assistance | Fewer errors and better decisions | Quality gains are poorly measured | Error, appeal and correction rates |
| Workforce development | Reinvest capacity in training and professional adaptation | Stronger institutional capabilities | Training fails to translate into changed operations | Demonstrated competencies and subsequent performance |
| Fiscal efficiency | Reduce avoidable operating costs while preserving service obligations | Greater budgetary flexibility | Premature workforce or expenditure reductions | Verified net cost per completed service |
| Administrative intensification | Increase reporting, monitoring or task requirements | Potentially greater management information | Efficiency gains absorbed by additional bureaucracy | Administrative overhead relative to service output |
These pathways also clarify why the governance of recovered time cannot be assigned exclusively to software developers or individual employees. The technical system determines which operations can be accelerated. Public managers determine how work is organised. Human-resources and budgetary authorities govern staffing and expenditure within the applicable legal framework. Legislatures establish statutory obligations and exercise fiscal oversight. Audit institutions and other competent supervisory bodies evaluate legality and performance. Citizens ultimately experience the consequences through the accessibility, timeliness and quality of public services.
The conversion problem: from individual efficiency to institutional value
A useful analytical distinction involves four successive levels of measurement.
Task efficiency records how much less time is required to complete a defined activity while maintaining acceptable quality. Workflow efficiency measures whether the improvement changes the processing capacity of a sequence of activities. Service performance examines whether final outputs improve in quantity, timeliness, accessibility or quality. Public value considers whether those changes achieve legitimate policy objectives without creating unacceptable legal, economic or social costs.
These levels cannot be treated as equivalent.
For example, a 50% reduction in the time necessary to prepare an administrative report is a substantial improvement in that task. If the report then remains in a queue awaiting review, the final service completion time may change very little. Conversely, if document preparation is the principal capacity constraint, the same improvement may contribute to significantly greater throughput.
The difference depends on the location of the bottleneck, the availability of downstream capacity, the arrival rate of new cases and the organisation’s ability to modify its procedures.
A related distinction concerns financial accounting. Hours saved are not automatically cash savings. Public expenditure declines only when the relevant costs can actually be avoided or reduced, net of AI procurement, infrastructure, training, oversight, maintenance and transition expenses. Alternatively, an administration may retain expenditure at its existing level and use the released capacity to deliver additional or better services.
Both outcomes can represent economic value, but they must be described and measured differently.
The OECD’s examination of AI in government reinforces the relevance of organisational design, governance, skills and infrastructure to the realisation of benefits. Its evidence does not support applying a universal productivity multiplier across jurisdictions or public functions. OECD
Principal Gaps and Watch Indicators
The central evidence gap in the international record is not the absence of AI programmes. It is the limited availability of independently interpretable, end-to-end measurements connecting AI use to public-service outcomes.
The following monitoring priorities are particularly consequential for institutional decision-makers.
Administrative completion times
Monitor the median and upper-percentile time required to complete an entire service, not merely the time spent on AI-assisted tasks. Case-management records should distinguish active processing from waiting, approval and transfer times.
Allocation of recovered capacity
Require records showing whether time savings are used for additional cases, backlog reduction, professional development, accessibility improvements, quality assurance or other institutional functions. This is essential to determining who actually controls the productivity dividend.
Authorised and unauthorised AI use
Examine approved-tool inventories, data-handling controls, recorded incidents, employee training and accessible alternatives to unapproved systems. Undocumented AI use creates uncertainty about both security exposure and the reliability of reported productivity gains.
Decision quality and human oversight
Track errors, reversals, complaints, appeals and the resources devoted to verification. A shorter initial processing time may be offset by additional corrective work, and apparent productivity gains should not conceal deterioration in substantive decision quality.
Net fiscal and economic benefits
Compare fully accounted AI operating and transition costs with verified changes in service capacity, expenditure and outcomes. Distinguish forecast benefits from observed improvements and from expenditure actually avoided.
Institutional scalability and resilience
Determine whether pilots can be extended across departments without weakening data governance, introducing vendor lock-in or shifting congestion to other administrative stages. The decisive evidence is sustained performance after deployment at operational scale.
Decision thresholds and institutional consequences
The clearest justification for expanding a public-sector AI application would be evidence that it maintains legal and professional standards while improving one or more final-service outcomes at an acceptable total cost. Where only employee-level efficiency has been demonstrated, wider deployment remains a proposition to be evaluated rather than an established productivity success.
A different decision is required when a system saves staff time but leaves processing queues unchanged. In such circumstances, the priority is to examine the workflow and identify whether the recovered capacity can be deployed at the constrained stage. Purchasing additional licences without examining that dependency may increase expenditure without addressing the operational cause of delay.
When AI-assisted outputs generate significant rework, the authority responsible for deployment must assess whether better data, training, model configuration, human review or a narrower application could resolve the problem. An increase in the number of automated tasks has little public value when it is accompanied by a deterioration in decision reliability.
Finally, the allocation of recovered capacity raises questions that cannot be resolved by measurement alone. Decisions concerning staffing, service priorities and the distribution of benefits among citizens, employees and public budgets are matters of institutional governance. Performance evaluation can inform those choices, but it cannot replace the authorities legally responsible for making them.
No decision-useful quantitative visualisation is supportable from a harmonised international productivity series in the verified record.
Strategic Assessment — The 2027–2031 Outlook
Three broad developments warrant monitoring over the five-year horizon.
The first is the movement from experimental adoption toward institution-wide integration. France’s government trials, Germany’s administrative applications, Britain’s quantified evaluations and Italy’s evolving implementation framework establish that the debate is increasingly concerned with operational adoption rather than AI capabilities in isolation. The principal uncertainty is whether institutions can reorganise complete administrative processes at the same pace as individual employees adopt new tools.
The second concerns the distribution of authority over automated work. AI assistants are presently used extensively for support activities such as drafting, research and summarisation. More integrated applications create a different organisational challenge because they can influence the sequencing, preparation and execution of administrative operations. This strengthens the importance of clearly defined human authority, auditability and responsibility for outcomes.
The third is the development of a more demanding public-sector productivity standard. Evidence of time saved will remain relevant, but governments seeking to establish durable economic benefits will need to show how those savings affect completed services, public expenditure, quality and administrative accessibility. The British evaluation record illustrates the value of explicit statistical methods; the French record demonstrates the benefit of combining employee experience, task measurements and service-related indicators. Italy’s implementation phase offers an opportunity to institutionalise such distinctions from the outset.
The United States, China and Russia demonstrate why this evolution will not produce a uniform international governance model. National administrative structures differ in the distribution of authority, treatment of government data, technological infrastructure, fiscal arrangements and mechanisms of public accountability. These differences affect how AI is deployed and how its outcomes can be assessed. They do not, by themselves, demonstrate superior or inferior productivity.
The decisive strategic question is whether governments can establish an accountable relationship between technological capability and the public value created from it.
Artificial intelligence can reduce the labour required for a government task. It cannot independently determine the legitimate use of the institutional capacity that becomes available. That determination belongs to public authorities acting within their legal mandates, informed by professional expertise, workforce arrangements, fiscal constraints and the needs of the people they serve.
For Italy and its European partners, the immediate opportunity is to connect AI governance with administrative performance measurement. For other major powers, the same underlying conversion problem appears within different institutional frameworks. Across all jurisdictions, the strongest evidence of progress will not be the volume of AI adoption, but sustained improvements in public-service delivery that can be attributed to identifiable organisational and technological changes.
Who Controls the Time Saved by Artificial Intelligence?
A comparative strategic dashboard on AI, government productivity, public value and institutional control.
What the official evidence shows
From time savings to public value
Four different levels of measurement must not be confused.
Administrative Time Conversion Simulator
Change the assumptions to see why a faster task does not necessarily produce an equally faster public service.
Simulation parameters
Interactive efficiency curve
X-axis: AI task-time reduction. Y-axis: total service-time reduction.
Illustrative model, not observed government data. Assumes sequential task, wait and review times; only the selected task duration changes. No extra rework or demand effects.
3D-Style Evidence Comparison
Different methods and units are presented in separate charts. Bars show observed indicators, not a country ranking.
UK: Daily time-saving estimates
Minutes per working day
Graphic scale: 0–30 minutes. Different populations and designs; the estimates must not be averaged.
France: Experimental task savings
Percentage of task duration
Graphic scale: 0–20%. Focus-group task results, not reductions in complete administrative procedures.
Comparative Government Intelligence
Select a jurisdiction to explore its documented governance model and the remaining productivity evidence gap.
Italy
Italy’s AI implementation framework is developing. The unresolved question is whether faster internal work produces measurable improvements in services.
Evidence gap
International Evidence Matrix
Official records, differentiated by the nature of their evidence.
| Jurisdiction | Official record | Indicator / status | Evidence type | Source |
|---|---|---|---|---|
| Italy | AgID guidelines, November 2026 | Adoption framework | Regulatory | AgID |
Six Controls for Converting AI Time into Public Value
Map the workflow
Identify the task, handoffs, queues and bottlenecks before claiming a service-level benefit.
Assign authority
Define who may authorise AI use and reallocate the working capacity it releases.
Protect accountability
Preserve human responsibility, lawful data handling and review of consequential decisions.
Measure final outcomes
Track completed services, waiting times, quality, appeals and error correction.
Audit fiscal benefits
Separate staff-hours saved from avoidable cash costs and fully account for AI expenditure.
Monitor deployment
Review security, unauthorised AI, workforce effects, vendor dependence and sustained performance.
OPEN-SOURCE ANALYTICAL ASSESSMENT | 20 NOVEMBER 2026
Pillar I — The Governance of Time and Public-Sector Productivity
Institutional authority, administrative transformation, European legal frameworks, workforce governance and accountability for the productivity gains generated by artificial intelligence.
Geographic coverage: Italy, France, Germany, the United Kingdom and the European Union. Strategic assessment horizon: 2027–2031.
Chapter 1. Italy: From Individual Efficiency to Administrative Performance
The central institutional challenge: transforming technological efficiency into measurable state capacity
Italy’s adoption of artificial intelligence in public administration has reached an important institutional transition. The decisive issue is no longer whether the technology can accelerate individual administrative activities, but whether public authorities possess the organisational capacity, legal safeguards and measurement systems required to transform those improvements into demonstrable changes in public-service performance. The central analytical judgment is that Italy now has a developing framework for governing the adoption of AI, but the existence of that framework does not establish that individual productivity gains have been converted into shorter administrative procedures, additional completed services or lower public expenditure. The distinction is fundamental because administrative productivity is not determined exclusively by the speed of the employee performing a task. It depends on the functioning of an interconnected institution in which responsibilities, controls, information flows, authorisations and statutory obligations determine when an administrative activity becomes a completed public service.
A modern administration performs activities that are only partially visible to the citizens and enterprises it serves. Preparing a legal opinion, reviewing an application, checking financial documentation, drafting correspondence and analysing administrative records may be necessary components of a procedure, but their completion does not necessarily constitute the final administrative outcome. An AI system can accelerate one or several of these operations without modifying the time required for the remaining stages. Consequently, the relevant unit of economic analysis cannot be limited to the individual employee or software application. It must extend to the entire process through which a public institution transforms resources, information and professional judgement into legally valid outputs. For Italy, this requires linking the deployment of AI to the organisation of public offices, the distribution of managerial responsibilities, the administration of public employment and the measurement of services delivered to citizens.
The Law of 23 September 2025, No. 132, published in the Gazzetta Ufficiale on 25 September and effective from 10 October 2025, established a national legislative framework for artificial intelligence. Its significance for the public sector lies in the relationship between technological development, fundamental rights, transparency, human autonomy and institutional responsibilities. The legislation must be interpreted alongside the European AI regulatory framework and existing domestic provisions governing administrative action, data protection and public employment. It should not be understood as an instrument that independently authorises the automated exercise of every administrative power or removes responsibilities already assigned to competent officials. Law No. 132 of 23 September 2025 — Gazzetta Ufficiale, September 2025.
A further development occurred on 9 November 2026, when the Agenzia per l’Italia Digitale adopted the guidelines for introducing artificial intelligence into public administration through Determination No. 169/2026. This was an important operational step because it translated elements of the national and European regulatory framework into guidance relevant to administrative adoption, organisational maturity, risk classification, cybersecurity, data quality and personnel competencies. AgID’s published explanation also identifies a Plan-Do-Check-Act approach to adoption, supported by the principles of continuous improvement and organisational assessment. Notably, the guidelines for developing AI systems and those concerning AI procurement were still undergoing their respective adoption procedures at the date of the agency’s published overview. This distinction matters: the adoption framework had been formally issued, while the broader guidance architecture was not yet uniformly complete. Determination No. 169/2026 — AgID, November 2026 and Artificial Intelligence: Guidelines and Implementation Status — AgID, November 2026.
The institutional significance of this development extends beyond technology governance. A formal adoption programme creates an opportunity to define what an administration intends to achieve, which activities will change, what risks must be controlled, who possesses decision-making authority and how performance will be evaluated. Nevertheless, the connection between organisational maturity and measurable administrative productivity must be established through actual implementation. A department can satisfy important governance requirements while leaving its principal service-delivery bottlenecks largely unchanged. Equally, an office may achieve substantial task-level efficiency without possessing the authority to redistribute staff, change approval sequences or redesign the service. The distinction between these possibilities is essential to any assessment of whether the public sector has obtained an economic return from AI.
The legal foundation: human authority, administrative responsibility and the boundaries of automation
Italian administrative law places particular importance on the lawful exercise of public powers and the identification of responsibility for administrative action. Artificial intelligence enters this institutional environment as a technical instrument whose outputs may influence the preparation, assessment and organisation of public decisions. The central legal issue is whether the system performs a supporting function or whether its operation amounts, in substance, to the exercise of a decision-making function subject to specific legal requirements. The distinction cannot be resolved simply by describing an application as an assistant. A system that generates a draft for independent human consideration raises different questions from one whose output is normally accepted without substantive examination, or from one that automatically determines how an application will be treated.
Article 14 of Law No. 132/2025 directly addresses the use of artificial intelligence in public administration. Its legislative setting places public-sector AI within the principles of lawful administrative action and responsibility, rather than establishing a general transfer of decision-making authority from public officials to automated systems. Article 14, Law No. 132/2025 — Gazzetta Ufficiale, September 2025.
This creates a critical operational requirement. Human oversight must represent an effective capacity to understand, question, correct or reject relevant machine-generated contributions, not merely a formal sign-off added to an otherwise automated decision. A public employee asked to approve an AI-generated administrative assessment must possess appropriate information about its evidentiary basis and limitations, sufficient time to conduct meaningful review and the institutional authority to intervene. Where these conditions are absent, the administration risks creating nominal human responsibility without corresponding practical control.
The economic consequence is equally important. Human review is not necessarily an inefficiency to eliminate. In high-consequence administrative functions, verification can form part of the service itself because it supports legality, protects rights and reduces downstream corrections. The appropriate productivity objective is therefore not the unrestricted reduction of human intervention, but the improvement of the overall process while preserving the scrutiny required by law and by the consequences of the decision.
The relationship between AI and the rights of individuals is further governed by European data-protection law. Article 22 of the General Data Protection Regulation establishes protections concerning decisions based solely on automated processing that produce legal effects or similarly significant effects, subject to the specific conditions and exceptions contained in the Regulation. The Court of Justice of the European Union, in Case C-634/21, interpreted the provision in relation to automated creditworthiness assessment and clarified the significance of the concept of an automated decision. These legal principles do not prohibit every use of algorithms in administration; they require analysis of the actual processing, its effects, the applicable legal basis and relevant safeguards. General Data Protection Regulation, Article 22 — European Parliament and Council, April 2016 and Judgment in Case C-634/21, SCHUFA Holding — Court of Justice of the European Union, December 2023.
For Italy, the resulting administrative challenge is to distinguish between functions that can be safely assisted, functions requiring structured human validation and activities in which automated processing encounters additional legal restrictions. This distinction is particularly relevant to procedures involving individual rights, public benefits, sensitive records, employment, enforcement or other consequential administrative determinations. It also demonstrates why the time recovered through AI cannot be allocated independently of the legal requirements governing the work that remains.
The organisational bottleneck: why faster employees do not automatically produce faster administration
The economic mechanism is best understood through the difference between individual task duration, operational capacity and total case completion time. Consider a hypothetical administrative procedure involving document preparation, verification, legal assessment and final authorisation. If AI reduces document preparation from 60 to 20 minutes, the administration has recovered 40 minutes of active working time for that case. If the file subsequently remains in an authorisation queue for several days, the total completion time may decline only marginally. The recovered capacity is real at the task level, but its value to the citizen depends on whether it can be used to reduce a binding constraint, process additional demand or improve the quality of subsequent work.
This distinction becomes more significant when AI is deployed across large organisations. If several departments accelerate their document preparation simultaneously while one legally required authorisation function remains unchanged, the volume of files awaiting that function may increase. In such circumstances, the local improvement can produce higher upstream output without proportionate gains in completed cases. An administration may consequently report greater employee productivity while citizens continue experiencing the same delays. The correct interpretation is not that AI has failed to accelerate the activities for which it was introduced, but that the organisation has not yet converted that acceleration into a corresponding improvement in the capacity of the complete process.
The relevant managerial task is therefore to identify which activities determine overall throughput and which are subordinate to other constraints. Administrative procedures should be examined through their constituent stages, the distribution of staff time, the duration of queues, the frequency of rework and the rules governing the transfer of files between units. Where bottlenecks are caused by insufficient professional capacity, AI may be most valuable when it reduces preparatory workloads and releases qualified personnel for activities requiring judgement. Where delays arise from fragmented information systems or procedural dependencies, the more effective intervention may involve integrating data and redesigning workflows rather than providing additional AI assistants to individual employees.
Italy’s AgID adoption guidance is relevant to this problem because the agency identifies organisational assessment, appropriate governance arrangements and continuous improvement as elements of responsible adoption. The guidance provides an institutional basis for treating AI deployment as a managed organisational process, although the precise productivity indicators and allocation rules adopted by individual administrations require separate examination. Guidelines for the Adoption of AI in Public Administration — AgID, November 2026.
The Italian administrative productivity measurement framework
The following table is an analytical measurement framework proposed for evaluating public-sector AI deployment. It is not a claim that the listed indicators have already been adopted or measured by all Italian administrations.
| Measurement dimension | Required indicator | Administrative significance | Principal limitation |
|---|---|---|---|
| Individual task efficiency | Minutes of active work per completed task | Establishes direct operational acceleration | Does not measure a completed public service |
| Workflow performance | Active processing time and waiting time at each stage | Identifies whether gains reach downstream units | Requires reliable process-level records |
| Service completion | Median and 90th-percentile completion time | Measures outcomes experienced by applicants | Changes may reflect demand or other reforms |
| Administrative throughput | Completed cases per period and staff-hour | Measures the capacity of the complete service | Must account for case complexity |
| Decision quality | Correction, reversal, appeal and error rates | Tests whether acceleration compromises reliability | Appeals may reflect factors beyond AI |
| Public expenditure | Fully loaded cost per completed service | Connects operational changes to fiscal outcomes | Must include implementation and operating costs |
| Workforce capacity | Hours reallocated by activity and professional role | Establishes how productivity gains are used | Reallocation may be informal or undocumented |
| Institutional safeguards | Recorded reviews, incidents and unauthorised uses | Evaluates accountable deployment | Detection practices affect incident counts |
| Accessibility | Successful service completion by relevant user groups | Tests equitable access to improvements | Requires appropriately disaggregated data |
| Sustainability | Performance after extended operational deployment | Distinguishes durable gains from pilot effects | Long observation periods delay evaluation |
An important methodological distinction is required between the mean and the distribution of administrative completion times. An average may improve while the most difficult cases remain severely delayed. In procedures involving vulnerable citizens, complex legal questions or incomplete documentation, the upper end of the completion-time distribution may be more relevant to public value than the average alone. Accordingly, any institutional productivity assessment should distinguish routine cases from complex cases, measure changes in demand and disclose whether altered case selection contributes to the apparent improvement. An administration that improves its average processing time by concentrating on straightforward applications while allowing difficult cases to accumulate has not necessarily improved its overall performance.
A credible assessment should also examine the rate at which completed outputs require subsequent correction. If AI reduces initial drafting time but increases the number of inaccurate legal references, incomplete assessments or unsuitable recommendations, the apparent gain may be offset by rework. Such costs can occur in different departments and at different points in time, making them difficult to identify through employee surveys alone. The appropriate measurement framework therefore connects task-level efficiency with error correction, administrative appeals, final service outcomes and resources used across the entire procedure.
Italian administrative productivity control architecture
Who Governs the AI Productivity Dividend?
Proposed allocation of responsibilities across technology, administrative management and institutional oversight.
Government and competent authorities
Policy priorities, regulatory obligations, public funding and cross-administration coordination.
Agency leadership and process owners
Authorised deployment, staffing decisions, service targets and workflow redesign.
ICT, legal, data and operational personnel
Security, implementation, validation, professional judgement and documented review.
Verified staff time released from defined tasks, with responsibility for its subsequent allocation.
Public employment: the distribution of recovered time is an institutional and contractual issue
Public-sector productivity also involves the organisation of employment, an area in which the technical ability to reduce task duration does not automatically confer authority to change staff duties, performance targets or working arrangements. In Italy, the employment framework established by Legislative Decree No. 165/2001 and the applicable system of collective bargaining must be considered when technological changes affect the conditions under which employees perform their work. The national framework agreement signed on 28 October 2025 for the 2025–2027 bargaining period confirmed the principal bargaining sectors of central government functions, local government functions, healthcare, and education and research. The existence of these separate sectors is relevant because workforce structures, functions and applicable contractual provisions cannot be treated as uniform throughout the Italian public sector. National Framework Collective Agreement for Public Employment Bargaining Sectors 2025–2027 — ARAN, October 2025.
The managerial question is not simply whether an employee can perform an established task more quickly. It is whether the administration possesses the authority and practical capacity to redirect the recovered working time toward additional files, more complex assessments, professional development or other authorised responsibilities. Different forms of reallocation can produce different economic effects. Increasing the volume of completed cases may expand effective service capacity without reducing expenditure. Assigning staff to more demanding legal or technical work may improve quality and institutional resilience. Reducing avoidable overtime or external support expenditure could produce measurable fiscal savings, provided the reduction actually occurs and is not offset by the cost of AI systems. The available capacity could also be absorbed by additional internal reporting, expanded supervision or administrative activities whose contribution to public value is limited.
These possibilities demonstrate why workforce governance must be incorporated into AI implementation before productivity expectations become embedded in formal operating plans. If management measures only the volume of AI-assisted output, employees may face pressure to increase production even where legal complexity, insufficient data quality or verification requirements make such increases inappropriate. Conversely, if employees retain all recovered time without an agreed organisational mechanism for sharing improvements, successful practices may remain isolated rather than spreading across the institution. An effective arrangement must therefore connect professional competencies, authorised management decisions, applicable labour rules and transparent performance evaluation.
Shadow AI, procurement and institutional visibility
Unauthorised AI use introduces another layer of governance complexity because the operational benefits may appear before the administration possesses reliable information about their technical origin, security implications or reproducibility. Employees may use personal subscriptions or external services for drafting, summarisation or research even when the organisation has not approved the tools for the relevant information. The immediate individual benefit may be reduced effort, but the institution may not know whether confidential records were processed, whether generated material was verified or whether information was subsequently retained by an external provider. This is not merely a cybersecurity consideration. It is also a productivity-measurement problem because the administration may attribute improvements to individual performance without identifying the technological process responsible for them.
An appropriate institutional response requires a distinction between identifying useful employee practices and authorising the technologies through which those practices are performed. The existence of unauthorised experimentation does not prove that the underlying activity is unsuitable for AI assistance. It indicates that the organisation needs an effective mechanism for assessing the use case, classifying the data involved, evaluating suppliers, testing accuracy and providing an approved alternative where appropriate. Excessively restrictive arrangements can discourage disclosure and leave informal practices outside institutional visibility. Conversely, unrestricted use can expose sensitive information and weaken accountability. AgID’s guidance on data quality, risk management, cybersecurity and organisational competencies provides relevant components for constructing an authorised adoption process. AI Adoption Guidelines — AgID, November 2026.
Procurement creates a related challenge. Purchasing licences and recording their distribution is administratively straightforward compared with establishing whether the resulting systems improve public-service performance. The procurement decision should therefore be connected to measurable use cases, interoperability requirements, security responsibilities, evaluation procedures and the administration’s ability to modify or terminate an arrangement when expected benefits do not materialise. Where procurement and operational assessment are separated, institutions may acquire technologically capable systems without sufficient incentives or organisational capacity to use them effectively.
Italy’s immediate strategic options and decision thresholds
Italy’s institutional framework supports three practical courses of action that can be distinguished by their scope and implementation requirements. An administration can first undertake a controlled measurement programme within existing procedures, recording task duration, waiting time, completed cases and quality indicators before and after AI adoption. This approach requires relatively limited organisational change, but it will not necessarily eliminate the underlying bottlenecks. A more substantial intervention would redesign processes around measurable service outcomes, which may require coordination among multiple offices, changes to case-management arrangements and adjustments to the distribution of personnel. A third option would institutionalise a broader performance and accountability model, integrating operational indicators with workforce governance, procurement evaluation, cybersecurity and financial reporting. This offers more comprehensive oversight but creates greater reporting and implementation burdens.
| Course of action | Competent decision level | Expected effect | Implementation burden | Principal risk |
|---|---|---|---|---|
| Controlled task-level evaluation | Authorised administrative management | Identifies genuine local efficiency gains | Limited to moderate | Results mistaken for service-wide gains |
| End-to-end process redesign | Service and agency leadership, with other competent authorities where required | Converts recovered capacity into improved throughput or service quality | Moderate to substantial | Shifting bottlenecks and disruption during transition |
| Institution-wide performance governance | Competent strategic, managerial and oversight bodies | Connects deployment with service, workforce and fiscal accountability | Substantial and continuing | Excessive reporting or compliance burdens |
| Approved-use and security framework | Management, ICT, security and data-protection functions within their mandates | Reduces unmanaged tool use and improves reproducibility | Moderate and continuing | Controls that are impractical for employees |
| Integrated workforce development | Management and competent personnel functions under applicable employment arrangements | Supports meaningful validation and new professional capabilities | Continuing investment | Training without corresponding organisational change |
These options should not be regarded as mutually exclusive. Their implementation requires sequencing and proportionality, rather than uniform application to every public office. An administration handling highly consequential individual decisions will require a different combination of controls from one using AI primarily for public communications or internal document retrieval. The relevant threshold for expansion is a demonstrated improvement in an appropriately defined service or organisational function, supported by evidence that legal compliance, information security and decision quality remain acceptable.
Chapter 1 — Key judgments
Italy has established important legal and administrative foundations for responsible AI adoption, including Law No. 132/2025 and AgID’s adoption guidelines of November 2026. These developments create institutional conditions for more systematic implementation, but they do not establish a national productivity dividend. The decisive remaining task is to connect approved deployment with records showing how working time is recovered, who possesses authority to reallocate it, which administrative constraints are affected and whether final public services improve.
The assessment would materially change if coordinated Italian administrative reporting established reliable before-and-after measurements of completed service outcomes, adjustment for changing case volumes and complexity, and the treatment of additional verification and implementation costs. Such evidence would make it possible to distinguish technological adoption from demonstrable administrative transformation.
Open official record: comparable service-level AI evaluation reports, documented capacity-reallocation arrangements and full-cost implementation assessments across Italian administrative bodies.
Chapter 2. European Governance: EU Law, France, Germany and the United Kingdom
European convergence in regulation, divergence in administrative performance
European governments are entering a phase in which AI adoption is increasingly judged against operational results, but their institutional arrangements remain substantially different. The principal judgment is that Europe is developing common rules for trustworthy AI while retaining nationally distinct systems for deciding how administrative capacity is organised, measured and redistributed. This separation has far-reaching implications. The European Union can establish harmonised obligations concerning certain AI systems, data governance and fundamental rights, but these obligations do not automatically align national budgetary processes, public-employment structures, administrative procedures or the managerial decisions through which recovered working time becomes improved service delivery.
This distinction makes comparisons between European countries both necessary and methodologically difficult. France has documented employee adoption and task-specific savings through an extensive government trial. Germany has concentrated part of its administrative modernisation activity on AI-assisted workflows and complex procedures, including planning and permitting. The United Kingdom has published government-wide experimental findings and a separate evaluation using an adjusted comparison-group methodology. Italy has recently formalised adoption guidance that connects technology implementation with organisational considerations. These developments concern related institutional objectives, but they are not equivalent stages of deployment or interchangeable measures of productivity.
The European challenge is therefore twofold. Governments must establish whether AI tools function effectively within their own administrative environments, and they must determine whether the resulting efficiency improvements are compatible with the rules governing public authority, individual rights and the use of public resources. These are complementary rather than competing requirements. An AI deployment that accelerates administrative work while generating unlawful decisions, unreliable assessments or unmanageable information-security risks does not constitute a sustainable improvement in state capacity. Equally, a compliant deployment that produces no measurable improvement in the relevant public function cannot be considered economically successful merely because its governance arrangements are formally satisfactory.
European Union: the regulatory framework and its operational consequences
The European AI Act, Regulation (EU) 2024/1689, created a risk-based legal framework governing the development, deployment and use of artificial intelligence. For public administrations, its significance derives from the nature of particular AI applications rather than the simple fact that a public institution uses the technology. Systems employed for routine drafting or information retrieval may raise different obligations from systems used in sensitive areas involving employment, education, public benefits, migration, law enforcement or justice. Relevant classification depends on the applicable legal definitions, the system’s intended purpose and the circumstances in which it is used, including the provisions governing high-risk applications and their exceptions.
The European framework must also be assessed in its November 2026 legal form, rather than relying on earlier descriptions of implementation deadlines. The AI Omnibus entered into force on 27 July 2026 following legislative changes to the AI Act. According to the European Commission’s implementation information, the rules for certain Annex III high-risk applications are scheduled to apply from 2 December 2027, while those concerning high-risk AI embedded in specified regulated products are scheduled to apply from 2 August 2028. Earlier provisions, including important prohibitions, governance requirements and transparency obligations, follow their own application schedules. The 2026 amendments also simplified elements of the AI-literacy framework and strengthened the role of the Commission and member states in promoting relevant competencies. AI Omnibus Enters into Force — European Commission, July 2026 and AI Act: Regulatory Framework and Application Timeline — European Commission, 2026.
This development is particularly relevant to the economic governance of public-sector AI because regulatory requirements influence the organisation and cost of deployment. Technical documentation, risk assessment, logging, information quality, transparency and human oversight can require resources that are not visible in headline estimates of employee time savings. Some of these activities represent new costs, while others formalise controls that public authorities should already maintain under existing administrative and data-protection requirements. The net economic effect depends on which obligations actually apply, how they are implemented and whether the administrative system already possesses the necessary capabilities.
It would therefore be incorrect to treat compliance expenditure uniformly as a loss of productivity. In some applications, well-designed controls can reduce costly errors, appeals, security incidents and disputes over automated decisions. In others, excessively fragmented or duplicative compliance procedures could increase administrative overhead without proportionate improvement in outcomes. The policy problem is to distinguish safeguards that materially reduce risk from processes that simply generate additional documentation. That requires coordination between legal professionals, technical specialists, procurement authorities, security functions and operational managers.
European data-protection law creates a further boundary. The GDPR governs the processing of personal data, including data used in AI-enabled administrative systems, and the specific rules on automated individual decisions must be examined wherever a system’s outputs produce legally significant consequences. The safeguards required for processing personal data cannot be replaced by the assumption that a human official is nominally responsible for the final outcome. The real function of that official, the legal basis for processing, the information made available and the practical ability to intervene remain relevant. General Data Protection Regulation — European Parliament and Council, April 2016 and AI Act Enforcement Framework — European Commission, 2026.
The resulting European institutional architecture contains an unavoidable allocation of responsibilities. EU law establishes applicable obligations and the conditions for their enforcement. National governments and competent authorities retain responsibilities for organising public administration within their legal systems. Public managers determine how authorised applications are integrated into services. Data-protection, audit and other supervisory bodies exercise their respective mandates. Personnel must possess the competencies necessary to use and evaluate the systems. These functions overlap operationally, but they cannot be consolidated into a single technological decision.
France: institutionalised experimentation and the distinction between adoption, performance and sovereignty
France provides particularly useful evidence because its government has conducted a relatively extensive evaluation of an AI assistant intended for public employees. The L’Assistant experiment, described in the Direction interministérielle de la transformation publique’s evaluation published on 16 June 2026, involved 10,000 employees across six ministerial groups over a ten-month experimental programme. The participating institutions included areas of government dealing with justice, finance, education, culture, higher education and research, and the Prime Minister’s services. The programme examined everyday activities such as document summarisation, rewriting, information extraction and translation, combining employee surveys, interviews and experimental focus groups.
The findings contained several distinct indicators. 75% of surveyed employees considered the assistant useful for their professional activity, while 65% indicated that they would recommend it to colleagues. A further 56% reported reducing their professional use of non-sovereign AI alternatives, a finding relevant to institutional control over the technology used in government. Experimental focus groups identified approximately 12% savings in writing activities and savings of up to 16% in document summarisation, accompanied by improvements in output quality under blind assessment. These findings support the proposition that an authorised public-sector AI assistant can produce measurable benefits in selected activities. They do not establish an equivalent reduction in the time required to complete entire administrative services. L’Assistant IA: Evaluation of the Experiment Among Public Employees — DITP, June 2026.
The significance of the French evidence is not limited to the size of the pilot. Its evaluation design recognises that adoption, perceived usefulness, measurable task efficiency and output quality are distinct dimensions of performance. A system can be well received without producing substantial time savings; it can generate faster drafts without improving accuracy; or it can provide valuable assistance to particular professional categories while offering little benefit in other functions. France’s combination of survey information, qualitative observations and controlled exercises provides a more informative foundation than licence counts or satisfaction measurements alone, although it remains necessary to examine the external validity of findings when extending them to different administrative functions.
The sovereignty dimension deserves separate treatment. The reported reduction in employees’ use of non-sovereign tools is an indication of changed behaviour among respondents, not a verified measure of all unauthorised use across the French public sector. Nevertheless, it demonstrates a relevant institutional mechanism: providing an accessible, authorised tool can change the incentives that lead employees to use external systems independently. This may improve visibility over data handling and technical dependencies while preserving the practical advantages that motivated adoption. The effectiveness of such an approach depends on whether approved tools meet real operational needs, whether employees understand their permitted uses and whether institutional security arrangements remain proportionate to the tasks involved.
France’s next productivity question concerns organisational integration. A centrally supported assistant may improve consistency in access and governance, but the conversion of recovered time into service outcomes still occurs within ministries, directorates and individual services. The government must therefore distinguish the administrative benefits attributable to a general-purpose assistant from those requiring specialised tools, structured data access and changes to professional workflows. An evaluation of text production cannot automatically establish the performance of a complex administrative decision, especially where that decision requires consultation, legal interpretation or several independent approvals.
Germany: redesigning complex procedures and the importance of local implementation evidence
Germany’s recent administrative AI initiatives provide a different perspective because they explicitly address the acceleration of complex operational procedures rather than concentrating solely on general-purpose employee assistance. The Federal Government’s official account published in November 2026 identifies the SPARK project, which uses AI-assisted modules within planning and approval processes, and describes initiatives that connect municipalities with companies developing agentic AI applications. According to the government’s account, early projects undertaken in spring 2026 achieved substantial reductions in processing times for certain care-related applications and requests associated with data-protection procedures. The government reported that care applications could be processed almost twice as quickly in the cited examples, while describing a 90% improvement in the processing of certain data-protection-related requests. The programme was also entering a further phase involving 38 municipalities and 17 AI start-ups or scale-ups. These results are official descriptions of particular pilot outcomes, not a nationally representative evaluation of all German administrative services. Artificial Intelligence as a Driver of Growth and Innovation: Federal Government FAQ — Bundesregierung, November 2026.
The German evidence is institutionally important because many high-value opportunities for administrative AI lie in the relationships between procedural stages rather than in isolated document-generation tasks. Planning applications, authorisations and other complex administrative processes can require numerous documents, technical checks, consultations and formal decisions. AI-assisted extraction, classification and preparation may reduce the time professionals spend obtaining relevant information or checking documentation, but the final economic result depends on whether these changes shorten the process without weakening legally required review.
The federal and municipal dimensions add complexity. Applications developed through local cooperation can produce useful evidence under conditions that are not necessarily present elsewhere. Municipalities differ in their technical infrastructure, professional staffing, administrative demand and ability to integrate AI systems with existing software. A result obtained in a well-supported pilot may therefore diminish when implemented in a less prepared institution. This is a standard external-validity problem, not evidence against the usefulness of the technology. The appropriate response is to document the implementation conditions associated with each outcome and determine whether they can be reproduced elsewhere.
Germany’s emphasis on digital sovereignty also affects the distribution of economic benefits. Public authorities that depend heavily on proprietary technical arrangements may face limitations in changing suppliers, transferring models, integrating systems or preserving long-term access to operational data. These factors can increase switching costs even when initial deployment produces measurable gains. Conversely, retaining greater control over technical infrastructure may require additional internal expertise, investment and maintenance capacity. The optimal balance cannot be determined from sovereignty objectives alone; it depends on the operational requirements, security risks, financial costs and available alternatives for a given administrative function.
Germany therefore highlights an important analytical distinction: the authority to assign recovered staff time and the capacity to control the technology generating those gains are connected but not identical. Administrative leaders may possess legal authority to redesign work while remaining constrained by technical systems that are difficult to modify. Technological sovereignty can reduce certain dependencies, but it does not guarantee that public institutions possess the organisational capacity needed to reallocate resources effectively.
United Kingdom: stronger evidence on the productivity of individual tasks, but incomplete evidence on final administrative outcomes
The United Kingdom provides detailed evidence that makes it possible to examine not only the magnitude of reported time savings but also the methodological limitations of attributing those savings to AI. The Government Digital Service conducted a cross-government Microsoft 365 Copilot experiment between 30 September and 31 December 2024, involving approximately 20,000 government employees across 12 participating organisations. Its report, published on 2 June 2025, described an average self-reported saving of 26 minutes per participant per working day. More than 70% of respondents agreed that the system reduced time spent on routine activities or searching for information and supported greater attention to strategic work. The report also identified limitations in complex and data-intensive tasks and recognised that security concerns influenced usage in some circumstances. Microsoft 365 Copilot Experiment: Cross-Government Findings Report — Government Digital Service, June 2025.
The most consequential limitation of this study concerns the allocation of recovered time. The Government Digital Service explicitly recognised that the experimental design did not establish precisely how time saved was subsequently used. That qualification is essential because employees’ perceptions of having more time for valuable work do not, by themselves, demonstrate how much additional work was completed, whether downstream delays declined or whether expenditure was reduced. Self-reported time savings can provide useful information about user experience and perceived efficiency, but they are not identical to independently observed productivity changes in complete administrative processes.
The Department for Work and Pensions subsequently published a separate evaluation on 29 January 2026, examining a trial conducted between October 2024 and March 2025 involving 3,549 staff members. Its methodology included surveys of 1,716 Copilot users and 2,535 non-users, alongside econometric analysis and qualitative evidence. Unlike the broader cross-government report, the DWP evaluation incorporated a comparison group and adjusted its estimates for characteristics associated with tool use. The fully adjusted model estimated approximately 19 minutes saved per working day, with a 95% confidence interval of 17–22 minutes. The study also acknowledged that the user group was not randomly selected and that self-selection remained a relevant methodological concern. An Evaluation of DWP’s Microsoft 365 Copilot Trial — Department for Work and Pensions, January 2026.
The DWP study is particularly informative because it disaggregates estimated effects by task. In its fully specified analysis, the estimated treatment effects included 26 minutes for searching existing information, 25 minutes for writing emails, 24 minutes for summarising information and 20 minutes for producing or editing written materials. Lower estimates were associated with other tasks, including 9 minutes for transcribing or summarising meetings. These are task-specific estimates within the study’s methodology; they should not be added together to infer the total savings achieved by every employee. DWP Copilot Evaluation, Table 4.4 — Department for Work and Pensions, January 2026.
The same evaluation identified changes in employee perceptions of work quality and job satisfaction. Its adjusted analysis reported an estimated 0.56-point improvement in job satisfaction and a 0.49-point improvement in perceived work quality, measured on seven-point scales. These results are relevant to workforce policy because productivity gains may be accompanied by effects on the experience of work and employees’ capacity to handle demanding activities. However, perceived output quality and job satisfaction are separate outcomes; they cannot be converted directly into additional completed cases or fiscal savings.
The United Kingdom consequently offers a useful methodological lesson for European governments. A credible evaluation should distinguish between direct time savings, employee-reported behavioural changes, adjusted estimates of productivity, final operational outcomes and financial consequences. The greater sophistication of an evaluation improves the interpretation of its findings, but even statistically significant estimates of employee task efficiency do not establish that an administration has recovered an equivalent quantity of budgetary resources.
European comparative evidence table: what can be measured and what remains unknown
| Jurisdiction | Official programme or framework | Verified figure or status | Evidence category | Principal inference permitted | Inference not established |
|---|---|---|---|---|---|
| Italy | AgID adoption guidelines, November 2026 | Determination No. 169/2026 | Operational guidance | A formal AI adoption framework has been issued | National service-productivity gains |
| France | L’Assistant, June 2026 | 10,000 participants across six ministerial groups | Deployment and evaluation | Substantial-scale experimentation occurred | National administrative productivity rate |
| France | Employee evaluation | 75% perceived usefulness; 65% recommendation; 56% reported reduced use of non-sovereign alternatives | Survey results | Reported acceptance and behaviour | Independently verified institutional time allocation |
| France | Experimental task evaluation | Approximately 12% writing savings; up to 16% summarisation savings | Task-level experiment | Specific task-duration improvements | Equivalent service-wide reductions |
| Germany | AI-supported municipal and federal initiatives | 38 municipalities and 17 technology firms in the next phase | Programme scope | Expansion of supported experimentation | Nationwide achieved productivity |
| Germany | Early operational pilots | Substantial reductions reported for selected procedures | Local pilot results | Potential for workflow-level acceleration | Reproducibility across all administrations |
| United Kingdom | Cross-government Copilot experiment | 20,000 participants; 26 minutes per day | Self-reported experimental outcome | Reported employee-level efficiency | Exact subsequent allocation of recovered time |
| United Kingdom | DWP Copilot evaluation | 3,549 trial staff; adjusted estimate of 19 minutes per day | Comparison-group econometric analysis | Evidence of task-efficiency effects in the evaluated setting | Equivalent fiscal savings |
| European Union | AI Act and 2026 amendments | Phased obligations, including 2027 and 2028 high-risk milestones | Legal framework | Regulatory requirements and implementation timetable | Uniform productivity outcomes across member states |
Interpretation: The available figures belong to different measurement categories. Participant counts are deployment measures; percentages of usefulness are perception measures; minutes saved are efficiency measures; process-duration reductions in specific pilots are operational outcomes. Combining them into a single national performance score would produce an artificial comparison unsupported by the evidence.
Sources: AgID Determination No. 169/2026, French DITP Evaluation, June 2026, German Federal Government AI Assessment, November 2026, UK Cross-Government Experiment, June 2025, UK DWP Evaluation, January 2026, and European Commission AI Act Framework, 2026.
Comparative institutional evidence
European Public-Sector AI: Verified Evidence Matrix
Official indicators, preserved in their original units. Different populations, methods and outcomes are not combined.
| Jurisdiction | Programme | Verified indicator | Evidence type | Official source |
|---|---|---|---|---|
| Italy | AgID adoption guidelines | 09 Nov 2026
Formal determination |
Regulatory implementation | AgID |
| France | L’Assistant | 10,000
Pilot participants |
Programme coverage | DITP |
| France | Writing-task experiments | ≈12%
Time saving |
Task experiment | DITP |
| Germany | Agentic AI Hubs | 38
Municipalities in next phase |
Programme expansion | Federal Government |
| United Kingdom | Cross-government Copilot trial | 26 min/day
Self-reported average |
Employee survey | GDS |
| United Kingdom | DWP Copilot evaluation | 19 min/day
Adjusted estimate |
Econometric evaluation | DWP |
| European Union | AI Act high-risk rules | 02 Dec 2027
Annex III application |
Legal timetable | European Commission |
Evidence cut-off: 20 November 2026. The table is descriptive, not a ranking. A deployment count, a task saving and a legal milestone cannot be represented on a common productivity scale.
What European administrations can learn from the differences
The most useful comparison is not a ranking of countries but an examination of which institutional questions their evidence addresses. The French evaluation provides information about employee adoption, task efficiency and the availability of an approved public-sector assistant. The German initiatives demonstrate the potential significance of applying AI to operational workflows where multiple processing stages can be accelerated. The United Kingdom provides a more detailed basis for assessing the statistical credibility of employee time savings. Italy’s recent guidance offers a framework through which performance assessment can be incorporated into AI adoption as the implementation process develops. The European Union provides common legal obligations that shape deployment without prescribing a uniform public-employment or service-management model.
These differences imply that policy learning should occur at the level of evaluation design and operational mechanisms rather than through direct transfer of headline figures. A French writing-task saving cannot be applied automatically to German permitting procedures, and a British estimate derived from Microsoft 365 Copilot use cannot be treated as the expected result of deploying a different system in an Italian municipality. Countries can instead compare the conditions under which benefits were observed: task definition, user selection, training, quality control, access to relevant data, system integration, baseline measurement and downstream process capacity.
The same principle applies to public expenditure. A government may achieve substantial time savings without reducing staff costs, particularly where employment arrangements and statutory service requirements remain unchanged. Another administration may use AI to reduce external contracting or avoid future capacity expansion. Those fiscal outcomes would need separate documentation and could not be inferred from time savings alone. Furthermore, differences in public-sector wage structures, infrastructure costs and procurement arrangements would affect comparisons even where task-level gains appeared numerically similar.
European coordination could improve the evidence base through shared definitions, standardised reporting templates and clearly described evaluation methods. Such coordination would not require every administration to adopt the same software or staffing arrangements. Instead, it could make the results of different national programmes more interpretable while preserving the legal and organisational differences that remain under national authority. This would be especially valuable for cross-border administrative cooperation, procurement evaluation and the assessment of applications deployed by multiple levels of government.
Chapter 2 — Key judgments
European AI governance is moving toward more operationally consequential implementation, but regulatory convergence should not be confused with uniform institutional productivity. France provides evidence of employee acceptance and measured savings in selected activities; Germany reports targeted improvements in some administrative processes; the United Kingdom offers more detailed quantitative evaluation of employee-level efficiency; and Italy has recently formalised its adoption guidance. None of these records establishes a directly comparable national productivity rate.
The assessment would materially change if administrations published service-level measurements using compatible definitions of processing time, quality, staffing inputs and implementation costs. It would also change if the expansion of experimental applications demonstrated that measured gains were sustained after deployment across materially different administrative environments.
Open official record: harmonised public-service AI evaluation results, implementation-cost studies, comparable quality indicators and documented national arrangements for redistributing the capacity released by automation.
Chapter 3. Who Owns Administrative Productivity? Authority, Labour and Accountability
The institutional ownership of time: separating the creation of efficiency from authority over its use
The economic consequences of artificial intelligence cannot be understood without addressing the distribution of decision-making authority inside public institutions. The principal judgment is that no single actor automatically acquires exclusive control over the benefits generated by public-sector AI. The employee who develops a more efficient working method, the administration that purchases the technology, the manager responsible for service delivery, the authority controlling public expenditure and the citizens entitled to receive public services occupy different positions in the governance of productivity. Their interests may converge, but they do not always produce identical priorities, and the relationship among them is determined by statutory responsibilities, employment arrangements, organisational decisions and democratic accountability.
The idea that AI-generated efficiency belongs exclusively to the employee who discovers it would misunderstand the institutional setting in which public services are delivered. Public employees generally carry out duties within organisations whose resources, information and obligations are established through law and public financing. Improvements in the performance of those duties can therefore have consequences beyond the individual worker. Nevertheless, the opposite assumption—that an administration may convert every minute saved into an immediate increase in output requirements without considering contractual obligations, professional standards or the quality of work—would also oversimplify public employment. The deployment of AI changes the technical conditions of work; it does not eliminate the legal and organisational arrangements governing that work.
The relevant problem is therefore one of allocation rather than proprietary ownership. Working time recovered from one activity becomes organisationally valuable when it can be reassigned to another legitimate function or when it reduces resources actually required to deliver a service. Whether that happens depends on the authority of the relevant manager, the availability of suitable tasks, employee competencies, applicable employment protections and the priorities established for the institution. An administration cannot assume that a reduction in document-production time creates an equal reduction in staff expenditure, because employees continue to perform contractual duties and the organisation may still require their expertise for other functions.
In this context, a productivity dividend should be understood as the net improvement in an institution’s ability to deliver its responsibilities from a given set of resources, or to deliver equivalent outcomes with fewer resources, without unacceptable deterioration in legality, quality, accessibility or security. This definition prevents the concept from being reduced to minutes saved on a computer screen. It also recognises that public value can arise from several outcomes other than increased production volume, including fewer erroneous decisions, more accessible services, improved professional capacity and greater institutional resilience.
Public managers: the immediate authority to translate efficiency into organisational performance
Public managers occupy a central position because they govern the relationship between individual activities and complete administrative processes. Their authority is not unlimited, and its scope differs across legal systems and administrative levels, but management decisions are normally necessary to translate task improvements into changes in workflow, staffing allocations and service priorities. A software application may reduce the time required to prepare a report; it cannot independently determine whether the released capacity should be assigned to a backlog, a complex investigation, additional quality assurance or professional training. Those choices require knowledge of the institution’s obligations, operational demands and available resources.
This responsibility becomes more demanding as AI tools expand from supporting isolated activities to assisting with connected workflows. When several operations are partially automated, managers must determine how human roles change, where review is necessary and which parts of the process require additional capacity. A technically successful deployment can fail to improve service delivery if the administration has not addressed the organisational constraints that remain. For example, increased output from an initial processing unit may generate additional work for a downstream legal department. If the second department lacks sufficient capacity, the organisation can experience a growing queue despite greater productivity in the first unit.
Such effects are particularly important in the public sector because administrative procedures are frequently structured around legally required checks rather than discretionary commercial workflows. A manager cannot simply remove a review stage because automation has made it inconvenient. The operational challenge is to determine whether the underlying statutory purpose can be achieved more efficiently through better information, clearer documentation, improved coordination or appropriately designed human oversight. The result should be measured through complete-service performance rather than the isolated productivity of individual departments.
Managerial accountability consequently requires more than an inventory of deployed AI systems. Administrations should identify the intended operational benefit of each application, the process owner responsible for achieving it, the resources necessary for implementation and the evidence required before expanding deployment. Without these elements, responsibility can become fragmented: technical teams report that a model functions, employees report that it saves time, procurement teams confirm that licences were delivered and service managers remain unable to demonstrate an improvement in final outcomes.
Employees: the productivity dividend and the changing substance of public work
AI adoption affects public employees not only by accelerating specific tasks but also by changing the distribution of attention, professional judgement and cognitive effort across their working day. Where routine drafting, summarisation or information retrieval becomes less demanding, employees may obtain additional capacity for analytical work, interactions with citizens, complex cases or professional development. These possibilities are supported in part by qualitative findings from the United Kingdom’s public-sector trials, in which participants described shifting attention toward activities they regarded as more valuable. However, the available evidence does not establish that such reallocation occurs uniformly or that it automatically improves measurable service outcomes. DWP Copilot Evaluation — Department for Work and Pensions, January 2026.
The distinction between assistance and intensification is economically consequential. AI may reduce the time necessary to complete established work while management leaves the total workload unchanged, creating space for greater professional attention or improved working conditions. Alternatively, the administration may increase expected output, potentially expanding service capacity but also altering the intensity of work and the demands placed on employees. A third possibility is that employees assume additional verification responsibilities because AI-generated material must be examined before use. The net workload effect then depends on the balance between tasks eliminated, new tasks introduced and the resources needed to correct mistakes.
It would therefore be insufficient to evaluate the workforce impact of AI by measuring the number of operations completed per employee. An increase in completed tasks may coexist with deteriorating accuracy, increased cognitive pressure or reduced opportunities for professional development. Conversely, a reduction in routine workload may improve employees’ ability to handle complex cases even where overall throughput initially remains unchanged. A comprehensive evaluation should distinguish productivity, professional competency, service quality and employee experience rather than treating them as interchangeable concepts.
The distribution of these effects may vary across professional categories. Lawyers and administrative specialists must assess legal validity and procedural requirements; technical staff must manage information security, reliability and system integration; managers must coordinate processes and assign resources; frontline employees may require tools that improve their ability to explain decisions and assist citizens. These different roles create different training requirements. General familiarity with AI tools is not equivalent to the ability to supervise a consequential automated assessment, evaluate a supplier’s security controls or redesign a legally regulated administrative procedure.
The contractual dimension is equally important. In Italy, the public-employment bargaining structure preserves distinctions among sectors and professional arrangements. In the broader European context, employment conditions and mechanisms for workforce consultation differ among jurisdictions. AI implementation must therefore be reconciled with the relevant national rules rather than assuming that productivity improvements create an unrestricted managerial entitlement to alter employment terms. Public Employment Bargaining Framework 2025–2027 — ARAN, October 2025.
Citizens: beneficiaries of public services rather than residual recipients of efficiency gains
The public ultimately evaluates administrative performance through the services it receives and the decisions that affect its rights and obligations. This creates a fundamental difference between productivity in public administration and efficiency measured solely through internal organisational accounts. A commercial enterprise may use automation to increase production, reduce operating expenses or expand margins, while a public authority must also consider statutory duties, equal treatment, accessibility and procedural fairness. These obligations affect what constitutes a legitimate improvement and how the benefits should be distributed.
An administrative service may be economically more productive if citizens receive decisions sooner, if fewer applications require correction or if vulnerable users obtain more effective assistance. These outcomes may justify investment even where expenditure does not decline. Conversely, an administration that reduces staff hours devoted to an activity while increasing errors or limiting meaningful review may produce an apparent accounting benefit without a corresponding increase in public value.
This distinction is especially significant for AI systems involved in consequential decisions. The European regulatory framework and GDPR protections demonstrate that public-service efficiency must be considered alongside the legal effects of automated processing and the rights of affected individuals. Faster decision-making cannot be evaluated independently of whether the decision is lawfully produced, sufficiently accurate and open to appropriate challenge. AI Act — European Commission, 2026 and General Data Protection Regulation — European Parliament and Council, April 2016.
Citizen-facing measurement consequently requires more than administrative statistics on completed tasks. It should examine the accessibility of services, the distribution of waiting times, the handling of complex cases, the quality of explanations and the availability of effective review. It should also distinguish between improvements experienced by citizens and changes that primarily benefit internal administration. Both may be legitimate objectives, but they should be reported separately to make the allocation of resources transparent.
The political economy of the AI productivity dividend
The economic value of recovered working time depends on the alternative uses available to an administration. A public employee who completes an activity more quickly does not necessarily create an immediate budgetary saving because salaries, office infrastructure and other operating costs may remain unchanged. Financial benefits can arise when the administration reduces avoidable expenditure, decreases reliance on external services, limits overtime or uses existing capacity to accommodate additional demand that would otherwise require new resources. These are different forms of economic benefit, and their fiscal implications should not be combined without explicit assumptions.
The distinction becomes particularly important when governments seek to estimate the aggregate economic impact of public-sector AI. Multiplying an average daily time saving by the number of employees and their hourly compensation produces a theoretical estimate of the value of working time. It does not establish the amount of expenditure that a government can remove from its budget. To convert that estimate into realised fiscal savings, the administration must demonstrate how its expenditure changed, account for the costs of AI and identify any offsetting changes in service quality or capacity.
A different calculation is required when the objective is service expansion. If AI enables an organisation to process additional applications without increasing employment, the relevant outcome is the additional number of completed services, adjusted for complexity and quality. The value of this gain depends on the public purpose of the service and cannot necessarily be expressed as an immediate reduction in expenditure. Similarly, when recovered time is redirected toward more careful verification, the economic benefit may appear through lower error rates, fewer disputes or reduced rework, rather than through a higher volume of initial outputs.
These distinctions reveal why budgetary authorities and operational managers require different but connected information. Financial reporting must establish what has actually happened to expenditure, while service-performance reporting must establish what has changed for citizens and administrative users. Workforce reporting must explain how staff activities have changed. Governance reporting must confirm that the relevant legal and security requirements remain satisfied. A comprehensive productivity assessment should integrate these dimensions without collapsing them into a single monetary figure.
Allocation matrix: who controls each part of the productivity process?
The following matrix represents a comparative functional framework. Actual legal competence depends on the jurisdiction, institution, statutory mandate and applicable employment arrangements.
| Institutional actor | Relevant authority or role | Decision affecting recovered time | Accountability requirement | Principal governance risk |
|---|---|---|---|---|
| Legislature | Establishes legal mandates and public-finance authorisations | Determines statutory responsibilities and appropriations | Legislative scrutiny and public accountability | Expectations inconsistent with legal obligations |
| Government and central coordination bodies | Establish policy and cross-administration priorities within their powers | Sets strategic objectives and implementation priorities | Policy evaluation and public reporting | Adoption targets substituting for outcomes |
| Budgetary authority | Controls expenditure within the applicable fiscal framework | Determines funding and evaluates fiscal consequences | Financial records and audit | Treating theoretical labour value as cash savings |
| Administrative leadership | Governs authorised organisational processes | Reallocates capacity and redesigns workflows | Service-performance and management reporting | Local efficiency without institutional benefit |
| ICT, data and security functions | Exercise technical responsibilities under institutional arrangements | Determine safe and technically feasible deployment conditions | Security and operational assurance | Unmanaged data exposure or technological dependence |
| Public employees | Perform professional and administrative duties | Apply AI assistance and exercise required human judgement | Professional and procedural accountability | Nominal oversight without effective verification |
| Workforce and employee-representation bodies | Exercise rights and functions under applicable employment rules | Participate where employment arrangements require consultation or bargaining | Employment compliance | Disputes over workload and professional roles |
| Data-protection and other competent supervisory authorities | Exercise legally defined oversight powers | Enforce applicable requirements | Regulatory accountability | Weak safeguards or unlawful processing |
| Internal audit and external audit institutions | Examine performance and proper use of resources within their mandates | Evaluate evidence and controls; do not ordinarily manage daily work allocation | Audit findings and follow-up | Undetected errors in benefit reporting |
| Citizens and service users | Hold applicable legal rights and provide service feedback | Influence accountability through lawful participation, complaints and review mechanisms | Accessible information and remedies | Productivity gains that do not improve public outcomes |
Authority, workforce and public value
Who Decides How AI-Recovered Time Is Used?
Analytical relationship between public authority, administrative execution, employment obligations and final public-service outcomes.
Legislature and Government
Statutory mandates, appropriations, public priorities and policy accountability.
Administrative Leadership
Service objectives, workflow design, resource allocation and implementation.
Public Employees and Specialists
Human judgement, technical supervision, professional duties and service delivery.
Verified capacity released from administrative work. Its allocation is a governance decision, not an automatic consequence of software performance.
Capacity and quality
More completed cases, shorter delays, improved accessibility and stronger verification.
Costs and workforce development
Verified avoidable expenditure, reduced rework, training and institutional resilience.
Continuous Oversight and Accountability
Applicable employment rights, data protection, human responsibility, audit, procurement controls and citizens’ legal remedies operate across the entire allocation process.
Analytical institutional model; not an official legal hierarchy. Responsibilities vary by jurisdiction and administrative mandate. No numerical productivity gains are assumed.
The accountability problem: preventing improvements from disappearing inside administrative statistics
An important institutional risk arises when public authorities measure AI deployment through indicators that are only indirectly related to the stated purpose of administrative reform. Licence counts, usage frequencies, employee satisfaction and the number of automated activities can establish that a technology has been adopted, but they cannot prove that a public institution delivers better services. Even a measured reduction in task duration may be insufficient if the administration cannot document what happened to the resulting capacity.
This is not simply a problem of inadequate statistical reporting. It affects the distribution of responsibility. When productivity is measured only at the level of individual employees, managers may attribute improvements to personal performance without examining organisational constraints. When it is measured only through total expenditure, governments may overlook improvements in service quality or capacity that do not immediately reduce costs. When it is measured only through completed case volumes, institutions may neglect the complexity of individual cases and the quality of decisions. A credible accountability framework must make these different outcomes visible and assign responsibility for the information required to evaluate them.
The United Kingdom’s evidence demonstrates the importance of acknowledging what an evaluation cannot establish. The cross-government Copilot report documented average reported time savings but recognised that the subsequent use of recovered time could not be precisely identified through the experimental design. The DWP evaluation improved the estimation of task-level effects through comparison-group analysis but did not convert those findings into verified Treasury savings. These are substantive methodological distinctions, not deficiencies that can be resolved by presenting a larger headline estimate. Cross-Government Copilot Findings — Government Digital Service, June 2025 and DWP Copilot Evaluation — Department for Work and Pensions, January 2026.
Accountability also depends on preserving the distinction between technical performance and administrative responsibility. A model may generate an apparently accurate summary, recommendation or classification, but the relevant institution must determine whether the output is suitable for its legal and operational purpose. Where a system contributes to a consequential decision, the adequacy of human oversight and the applicable rights of affected individuals cannot be established through model accuracy alone. Appropriate supervision requires a combination of legal authority, professional competence, reliable records and effective procedures for correcting errors.
The institutions responsible for audit and supervision have different functions from those responsible for daily implementation. Auditors may examine the evidence supporting reported improvements, while data-protection authorities may assess compliance with personal-data rules. Public managers must retain responsibility for organising work and responding to operational findings. These functions should reinforce one another without transferring managerial responsibilities to auditors or allowing technical suppliers to determine the standards by which public institutions evaluate themselves.
From reporting minutes saved to documenting the use of public capacity
A more rigorous governance approach would distinguish five components of the productivity dividend. The first is the operational gain achieved through AI-assisted activities. The second is the proportion of that gain that remains available after verification, correction, training and other new tasks are considered. The third is the share of net recovered capacity that can actually be reassigned within the applicable institutional and employment framework. The fourth is the observable consequence for services, expenditure or workforce capabilities. The fifth is the durability of the outcome after the application moves beyond its initial experimental phase.
This framework has a practical advantage: it prevents governments from claiming a uniform financial benefit from heterogeneous forms of productivity improvement. Recovered capacity devoted to complex cases has a different economic meaning from avoided overtime expenditure. Additional completed applications are different from a reduction in errors. Improvements in employee satisfaction are different from lower service-delivery costs. These outcomes can all matter to an institution, but they should not be conflated merely because AI contributed to them.
For administrative reporting, the most consequential evidence would be a documented allocation of recovered time, the organisational decision authorising that allocation and the observed outcome. Such reporting would enable an institution to distinguish whether its AI investments expanded service capacity, reduced backlogs, strengthened quality control or altered the cost of delivering established services. It would also permit managers and oversight bodies to examine whether the reported gains remain credible after accounting for implementation costs and possible changes in demand or case complexity.
A government adopting this approach would not need to dictate the same allocation of recovered time across every public institution. Different services face different constraints and legitimate priorities. The objective would be to ensure that the decision is explicit, within the authority of the responsible institution and supported by evidence appropriate to the intended benefit. This creates a basis for more meaningful comparisons across departments without imposing a mechanically uniform productivity target.
Strategic decision matrix for 2027–2031
| Strategic choice | Authority required | Intended effect | Time-to-effect | Reversibility | Principal downside |
|---|---|---|---|---|---|
| Standardise task-level AI measurement | Administrative management and relevant evaluation functions | Establish credible local efficiency baselines | Near term, after suitable baseline collection | Generally high | Additional reporting without service improvements |
| Link AI programmes to complete-service indicators | Service leadership and other competent operational authorities | Demonstrate changes in final outcomes | Medium term, dependent on process and data integration | Moderate to high | Attribution difficulties where several reforms occur together |
| Establish documented capacity-allocation decisions | Management under relevant employment arrangements | Explain how recovered time is used | Near to medium term | Depends on staffing and procedural changes | Administrative rigidity or excessive monitoring |
| Integrate AI evaluation with financial reporting | Budgetary authorities and competent financial management functions | Distinguish operational value from realised savings | Medium term | Generally high for reporting methods | Misclassification of theoretical gains as fiscal benefits |
| Institutionalise professional and legal oversight | Relevant legal, operational, security and supervisory functions | Preserve decision quality and lawful deployment | Continuing requirement | Depends on applicable legal obligations | Disproportionate controls if poorly designed |
| Evaluate long-term organisational effects | Management, workforce functions and competent evaluators | Assess sustainable capacity, competencies and workload | Multi-year | Variable | Uncertain attribution and delayed recognition of benefits |
The table describes policy options rather than commitments already adopted by the jurisdictions examined. The implementation periods are qualitative planning categories, not forecasts of when governments will achieve specific numerical productivity gains.
The 2031 institutional test: evidence of value rather than technological adoption alone
The central issue over the next five years will be whether public institutions can establish an auditable connection between AI-enabled task acceleration and the outcomes that justify public expenditure. The evidence already available demonstrates that AI assistance can reduce the time required for selected forms of administrative work. It also shows that the strength of that evidence varies significantly according to evaluation design and the type of outcome measured. These findings establish the relevance of AI to public-sector productivity, but they do not resolve how governments should distribute the recovered capacity or demonstrate that any single allocation model is universally appropriate.
The most informative future evidence would document what happens after an administrative activity becomes faster. An institution that uses AI to reduce document preparation time should be able to show whether the resulting capacity was used to address a backlog, perform additional reviews, process more applications, improve accessibility or reduce expenditure. If none of these outcomes can be established, the administration may still have achieved a useful task-level improvement, but it should not claim an unverified service-wide or fiscal benefit.
The same evidence would help determine whether new technological dependencies are changing the distribution of authority inside public administration. Systems that organise information, prepare assessments and coordinate workflows can influence how officials spend their time and which matters receive attention. Governance arrangements must ensure that such technical processes remain subordinate to legitimate institutional objectives and that their operation can be evaluated and corrected. The question is not whether human officials should perform every routine administrative activity themselves, but whether the institution retains meaningful control over the functions, resources and decisions for which it remains legally responsible.
These considerations lead to a final distinction between productive capacity and public value. Productive capacity concerns what an organisation is able to accomplish with its resources. Public value concerns whether that capability is used to fulfil legitimate public responsibilities effectively, lawfully and equitably. AI may expand the former without automatically improving the latter. The relationship between them is determined by institutional decisions, professional competencies, legal obligations and the systems through which performance is measured.
Chapter 3 — Key judgments
The productivity dividend generated by artificial intelligence is governed through a distribution of public authority rather than owned exclusively by the technology supplier, the employee or the administrative manager. Employees contribute professional expertise and exercise relevant responsibilities; managers organise authorised work and allocate capacity; budgetary and political authorities determine priorities within their respective mandates; and supervisory institutions examine the legality and accountability of the resulting activities. Citizens experience the consequences through public-service outcomes and the exercise of their rights.
The evidence supports the importance of distinguishing employee-level efficiency from institutional performance and fiscal effects. It does not establish that recovered time is automatically converted into additional services, reduced costs or improved working conditions. Those outcomes depend on subsequent decisions and must be observed separately.
The assessment would materially change if comparable public-sector reporting established how recovered time was allocated, which authorities approved the allocation, how employee roles changed and whether completed services improved after controlling for implementation costs, demand and case complexity.
Open official record: institutional allocation decisions, audited AI-related cost and benefit assessments, workforce-impact evaluations, service-level outcome data and evidence of sustained performance after full operational deployment.
Pillar I — Consolidated Institutional Assessment
The three chapters establish a common analytical finding without implying that Italy and its European counterparts operate through identical administrative systems. Artificial intelligence can increase the productive capacity of public institutions, but the conversion of that capacity into public value depends on governance decisions made after the initial technological improvement.
Italy’s recently adopted guidance provides a developing framework for the organisation of AI deployment. France’s experiments demonstrate measurable improvements in selected tasks and provide evidence concerning adoption and institutional control over approved AI tools. Germany’s initiatives illustrate the importance of addressing complex administrative workflows and the operational conditions required for wider implementation. The United Kingdom offers more detailed quantitative evidence of employee-level time savings and shows why careful statistical evaluation does not eliminate the need to measure final outcomes. The European Union establishes important common legal requirements while leaving much of the organisation of public services and employment to the relevant national systems.
Together, these developments identify the principal institutional requirement for the 2027–2031 period: public authorities need evaluation frameworks capable of distinguishing technological performance, organisational capacity, final service outcomes, fiscal consequences and legal accountability. They also need governance arrangements that identify who may redirect recovered working time, which objectives justify that allocation and how its effects will be assessed.
The resulting conclusion is not that every minute saved must be converted into a larger volume of administrative output, nor that workforce reduction is the necessary expression of productivity. The legally and economically relevant objective is to ensure that recovered capacity is allocated through authorised institutional decisions and that its contribution to public purposes can be demonstrated.
The future productivity of public administration will depend not only on how quickly artificial intelligence performs work, but on whether governments can account for the use of the human capacity it releases.
Pillar II — International Models and the Economics of AI Deployment
Federal AI governance, state-directed administrative modernisation, technological sovereignty, cybersecurity, workforce transformation and the economic distribution of artificial-intelligence productivity gains.
Geographic and institutional coverage: United States, China and Russia, with cross-jurisdictional examination of cybersecurity, shadow AI, human capital and technological dependencies. Assessment horizon: 2027–2031.
Chapter 4. The United States: Federal AI Governance and Agency Performance
The American federal model: accelerated deployment under decentralised operational responsibility
The United States has developed a distinctive institutional approach to artificial intelligence in government, combining centrally issued executive-branch requirements with substantial responsibility for implementation inside individual federal agencies. Its significance extends beyond the rapid adoption of generative AI. The federal system provides a particularly important case study of how technological investment, regulatory authority, procurement rules, administrative capacity and institutional accountability interact when government organisations attempt to convert automation into measurable operational improvements. The central finding is that the United States has documented rapid growth in federal AI use cases and established governance mechanisms intended to accelerate deployment, but the official evidence does not demonstrate an equivalent increase in realised government-wide productivity. The most consequential weakness is the continuing distance between the reported expansion of AI applications and the capacity of agencies to document workforce readiness, operational outcomes and the benefits attributable to individual systems.
The American framework must be understood through the institutional relationship between the Office of Management and Budget, agency leadership, Chief Artificial Intelligence Officers, technology and procurement functions, Congress and the Government Accountability Office. These actors exercise different powers. The executive branch can establish government-wide administrative requirements for covered agencies, while operational implementation remains dependent on agency missions, appropriations, statutory responsibilities and technical capabilities. Congress determines important elements of the legislative and fiscal environment and exercises oversight. The Government Accountability Office (GAO) evaluates implementation and identifies deficiencies, but it does not assume responsibility for operating agency systems. This distribution of authority matters because federal AI governance is not a single centrally controlled technological programme. It consists of multiple institutional decisions made within a common policy environment.
A major policy development occurred on 3 April 2025, when the Office of Management and Budget issued Memorandum M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust. The memorandum replaced the previous M-24-10 framework and established revised expectations for federal AI adoption, agency leadership, inventories and risk management. The policy reflected the administration’s emphasis on reducing barriers to AI deployment while retaining safeguards applicable to privacy, civil rights, civil liberties and uses presenting substantial consequences for individuals. It also reinforced the role of Chief AI Officers and agency governance arrangements, requiring covered agencies to develop mechanisms for coordinating AI investments, risk assessment and deployment priorities. M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust — Office of Management and Budget, April 2025.
The economic significance of this policy is not confined to the removal of administrative obstacles. The memorandum identifies a category of high-impact AI in which system outputs serve as a principal basis for decisions or actions having significant consequences for rights or safety. This creates an important distinction between applications that can be introduced through comparatively streamlined processes and those requiring additional safeguards. The determination depends on the system’s actual function and impact, not merely on the presence of a human reviewer. A human-in-the-loop arrangement does not automatically remove an application from the high-impact category if the AI output remains a principal basis for a consequential decision. Consequently, the federal framework preserves an important boundary between accelerating government work and permitting technology to exercise influence over sensitive administrative outcomes.
The same date also saw the issuance of OMB Memorandum M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government, which revised the requirements governing federal procurement of AI capabilities. Its provisions address commercial competition, vendor sourcing, data portability, interoperability, intellectual-property considerations and cross-functional participation in acquisition decisions. The memorandum recognises that government organisations may encounter costly dependencies when acquiring AI systems without adequate provisions for exchanging data, changing suppliers or integrating future technologies. These concerns have direct implications for administrative productivity because the economic value of an AI application depends not only on its initial performance but also on whether an agency can maintain, modify and replace it without disproportionate expenditure or disruption. M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government — Office of Management and Budget, April 2025. The White House
Taken together, these memoranda establish a governance structure in which adoption, procurement and operational risk are interdependent. A federal agency may identify an application capable of reducing the time required to analyse documents, manage correspondence or retrieve information, but implementation also requires appropriate authorisation, a sustainable acquisition arrangement, technical integration and an evaluation of whether the system supports the agency’s mission. The formal introduction of a technological capability is therefore only one stage in a process that can succeed economically only when the institution acquires and maintains the organisational conditions necessary to use it.
Federal AI deployment: the difference between growth in reported use cases and realised performance
The most important official quantitative baseline is provided by the US Government Accountability Office’s report GAO-25-107653, published on 29 July 2025. The GAO examined AI inventories covering 11 selected federal agencies and found that the total number of reported AI use cases rose from 571 in 2023 to 1,110 in 2024, while reported generative-AI use cases increased from 32 to 282 over the same period. These figures represent a substantial expansion in documented activity, but the reporting universe is limited to the agencies selected for the examination and does not constitute a complete inventory of every federal organisation. Moreover, the figures describe reported use cases rather than independently measured productivity outcomes. Artificial Intelligence: Generative AI Use and Management at Federal Agencies, GAO-25-107653 — Government Accountability Office, July 2025. GAO Files
The development-stage distribution is particularly revealing. Of the 282 generative-AI use cases reported for 2024, 84 cases, representing approximately 30%, were in the initial phase, where a requirement and intended purpose had been identified. A further 75 cases, approximately 27%, were in acquisition or development. The remaining 123 cases were distributed among implementation and assessment, operation and maintenance, or retirement. It would therefore be inaccurate to describe all 282 reported use cases as fully operational systems delivering measurable benefits. The inventory included activities at different stages of maturity, and some applications were not yet capable of producing sustained operational outcomes. GAO Files
The functional distribution also provides valuable evidence about the priorities guiding early generative-AI adoption. According to the GAO’s analysis, approximately 61% of reported generative-AI use cases belonged to mission-enabling functions, including activities such as improving written communications, increasing access to information, summarising reports and monitoring programme status. Government services accounted for 15%, while health and medicine represented 9%. Together, these three categories accounted for 84% of the reported generative-AI use cases in the selected agency inventory. These figures are significant because mission-support activities often allow organisations to introduce AI without immediately changing the substantive legal authority governing public decisions. They also show why adoption figures can expand rapidly before comparable service-level productivity evidence becomes available. GAO Files
The composition of the inventory suggests that generative AI was being adopted extensively for activities whose immediate benefits are relatively easy for employees to perceive but whose institutional consequences are harder to measure. Improvements in the drafting of an internal report, the retrieval of technical material or the preparation of routine correspondence may release professional capacity, yet they do not necessarily affect the completion time of a benefits determination, a regulatory inspection, a tax examination or a healthcare-related service. The economic question is therefore whether federal agencies possess the operational metrics and management arrangements needed to identify how those gains influence their respective missions.
The GAO also identified definitional issues affecting comparisons between reporting periods. In particular, changes in guidance concerning the reporting of research and development activities influenced the inventories submitted by some agencies. Such differences demonstrate why year-on-year increases must be interpreted as growth in reported and classified use cases, rather than as a precise measure of the total expansion of federal AI capability or expenditure. The increase from 32 to 282 is an important indicator of reported adoption, but its economic implications cannot be inferred from the numerical change alone.
US Federal AI Deployment — Verified GAO Indicators
| Indicator | 2023 | 2024 | Interpretation |
|---|---|---|---|
| Total reported AI use cases | 571 | 1,110 | Reported inventory across 11 selected agencies |
| Reported generative-AI use cases | 32 | 282 | Expansion of generative-AI reporting and adoption |
| Generative-AI use cases in initial phase | Not established in this comparison | 84 | Intended use identified; not equivalent to operational deployment |
| Generative-AI use cases in acquisition or development | Not established in this comparison | 75 | Systems under procurement or development |
| Generative-AI cases in other lifecycle stages | Not established in this comparison | 123 | Combined implementation, operational and retired categories |
| Mission-enabling share | Not established in this comparison | 61% | Activities supporting agency missions |
| Government services share | Not established in this comparison | 15% | Government-service applications |
| Health and medicine share | Not established in this comparison | 9% | Healthcare-related applications |
Source: Government Accountability Office, GAO-25-107653, July 2025. Percentages are reported category shares, not shares of expenditure or realised benefits.
The distribution reveals a fundamental issue in the economics of deployment. A growing inventory represents an expanding portfolio of technological investments and experiments, not a completed portfolio of productivity gains. The appropriate questions for federal oversight concern which applications enter sustained operation, how much they cost over their full lifecycle, what changes occur in relevant service indicators and whether improvements persist after accounting for maintenance, verification and workforce development.
The Internal Revenue Service: a concrete test of the relationship between technological ambition and institutional capacity
A more recent GAO investigation provides unusually direct evidence of the organisational constraints that can affect government AI adoption. On 24 March 2026, the GAO published report GAO-26-107522, examining artificial-intelligence governance and implementation at the Internal Revenue Service. The audit identified 126 active AI use cases in the IRS inventory as of June 2025, but it also found weaknesses in workforce planning, information quality, strategic management and the completeness of application records. The GAO concluded that staffing reductions had contributed to shortages of personnel with the technical competencies needed to develop and support AI systems, while the agency lacked an adequate workforce plan for identifying and addressing the relevant skills gaps. Artificial Intelligence: IRS Actions Needed to Address Skills Gaps, Information Quality, and Strategic Management, GAO-26-107522 — Government Accountability Office, March 2026. U.S. GAO
This finding is especially important because it identifies a mechanism through which the pursuit of administrative efficiency can undermine the capacity required to obtain future technological benefits. If an institution reduces the number of experienced personnel before establishing which professional and technical capabilities remain necessary, it may encounter difficulties maintaining existing systems, developing new applications, validating model outputs or managing suppliers. The resulting constraint is not necessarily a shortage of AI software. It may be a shortage of employees capable of integrating that software into the institution’s work.
The GAO recommended that the IRS identify its AI-related skills requirements, establish a workforce plan, improve the quality assurance of inventory records and strengthen the relationship between AI applications and the agency’s strategic objectives. The agency agreed with the eight recommendations made in the report. Agreement with recommendations is a documented institutional response, but it should not be treated as proof that the weaknesses had already been corrected or that the proposed measures had produced operational improvements. U.S. GAO
The IRS case demonstrates that the economic returns from AI can depend on capacities that are not directly visible in measures of task completion. Technical specialists, data engineers, system architects, legal experts and experienced operational staff may contribute indirectly to productivity by maintaining information quality, identifying suitable applications and supervising consequential decisions. Reductions in these functions can create bottlenecks that offset efficiencies elsewhere. An institution may therefore face a situation in which it possesses more AI applications but insufficient human capacity to operate them reliably.
The information-quality dimension is equally consequential. An incomplete AI inventory prevents managers from knowing which systems are being used, which applications support strategic objectives and where technical or financial dependencies may be accumulating. Without reliable inventories, expenditure and performance reporting may omit systems that consume resources or create risk. The problem is not merely administrative recordkeeping: it affects the capacity to evaluate whether public investment is producing its intended benefits.
The economics of procurement: competition, interoperability and the life-cycle cost of government AI
AI procurement introduces a financial dimension that is frequently absent from claims about time saved. A government institution does not purchase only the output of an algorithm. It acquires access to an operational arrangement that can involve software licences, computing infrastructure, integration services, information security, staff training, vendor support, model updates, auditability and technical continuity. The economic value of an AI-enabled service depends on the balance between the benefits it generates and the complete costs necessary to deliver and sustain those benefits.
OMB Memorandum M-25-22 addresses several of these issues directly. Its emphasis on competition, data portability and interoperability reflects the risk that agencies may become dependent on a supplier whose services are difficult or expensive to replace. Such dependency can change the economics of an initially successful implementation. A low initial subscription cost may be offset by high future switching costs, limited access to operational records, difficulties integrating with other government systems or contractual restrictions on how data and technical components may be reused. The White House
From an economic perspective, the relevant procurement decision must therefore consider the entire period during which the system is expected to remain operational. Initial deployment expenditure is only part of the calculation. Recurring charges, changes in computing demand, the cost of technical modifications, the availability of competing suppliers and the resources required to discontinue the arrangement all affect the final return. Agencies that treat procurement as an isolated purchase rather than as a component of long-term institutional capability risk overstating the expected net benefit of automation.
The fiscal consequences of AI adoption are also affected by the relationship between staff expenditure and operational capacity. If a federal agency saves several minutes per employee each day while retaining its existing workforce, the resulting improvement may permit additional work or better service quality. It does not automatically create an equivalent reduction in the federal budget. Realised cash savings require actual avoidable expenditure, while improved output requires observable changes in mission performance. These outcomes should be assessed separately, particularly when senior decision-makers are evaluating whether to scale an application beyond its initial deployment.
The American governance problem: adoption maturity versus institutional readiness
The combination of rapid inventory growth, revised federal governance requirements and the GAO’s IRS findings demonstrates that the principal challenge is not simply identifying more opportunities for automation. It is establishing whether agencies possess the organisational conditions required to translate those opportunities into sustained value. Leadership arrangements, reliable inventories, workforce plans, procurement controls and risk-management procedures are not substitutes for operational performance, but they affect whether performance improvements can be achieved and verified.
This creates an important distinction between AI adoption maturity and administrative productivity maturity. The former concerns whether an agency can identify, procure, deploy and manage AI applications. The latter concerns whether those applications produce reliable improvements in statutory functions, public-service quality, institutional capacity or fiscal performance. A government can make substantial progress in the first dimension while remaining unable to demonstrate the second.
For the United States, the principal future evidence requirements are therefore specific. The federal record needs to distinguish operational applications from those still under development, identify the costs and responsibilities attached to each system, document the workforce capacity required for maintenance and verification, and connect deployments with changes in relevant agency outcomes. Such evidence would allow federal authorities to evaluate the economic consequences of AI investment without relying on the number of reported applications as a proxy for productivity.
Verified US federal generative-AI deployment
Generative AI Adoption and Lifecycle Evidence
Reported use cases across 11 selected federal agencies. Figures are counts of reported applications, not productivity or budgetary savings.
Generative AI Use Cases
Common scale: 0–300 cases
The increase reflects the growth in reported generative-AI use cases, including applications not yet fully operational.
2024 Lifecycle Distribution
All 282 reported generative-AI cases.
The final category combines implementation and assessment, operation and maintenance, and retired cases. It must not be labelled entirely operational.
Source: US Government Accountability Office, GAO-25-107653, 29 July 2025. The 11-agency sample is not the complete US federal government. Reporting definitions can affect year-on-year comparisons.
Chapter 4 — Key judgments and consequential evidence gaps
The United States provides strong evidence of growing AI adoption across federal agencies, accompanied by an evolving framework for governance and procurement. The 2025 OMB memoranda establish requirements intended to facilitate responsible deployment, while GAO reporting demonstrates that significant portions of the inventory remain at development or implementation stages. The IRS audit adds an important operational finding: deficiencies in workforce planning and application records can constrain the effectiveness of AI even when an institution has identified numerous opportunities for use.
The assessment would materially change with consistent agency-level reporting of fully operational applications, full lifecycle expenditure, staff competencies, service outcomes and documented changes in operational capacity. Such information would make it possible to distinguish diffusion of AI technology from realised improvements in federal administrative performance.
Open official record: agency-level full-cost evaluations, outcome-based assessments of operational AI use cases, implementation records addressing GAO workforce recommendations and reliable evidence of how recovered administrative capacity is reassigned.
Chapter 5. China and Russia: State Coordination, Administrative Modernisation and Technological Sovereignty
AI deployment as an instrument of national administrative architecture
China and Russia approach artificial intelligence within broader programmes of state-directed digital transformation, technological infrastructure development and administrative modernisation. Although their political systems, economic structures, industrial capabilities and international technological dependencies differ, both have adopted official strategies that connect the development of digital technologies with government functions and national technological priorities. Their relevance to the international productivity debate lies particularly in the relationship between central coordination, shared infrastructure, institutional autonomy and control over information systems. The principal analytical finding is that centralised direction can influence how AI resources are distributed across government, but neither announced targets nor shared technological infrastructure establishes that the resulting applications deliver a measurable productivity dividend. The capacity to mandate or coordinate deployment is analytically separate from the capacity to demonstrate improvements in completed administrative services.
This distinction becomes important when comparing these countries with the United States and European states. The American federal system places substantial operational responsibility within individual agencies, subject to executive-branch guidance and legislative oversight. European governments operate through national administrative structures influenced by common EU rules, where applicable. China’s policy framework explicitly emphasises coordinated infrastructure, model reuse and the prevention of duplicated government AI deployments. Russia’s strategic framework links digital administration to common state systems, electronic services, information exchange and technological independence. These institutional arrangements may influence the cost, speed and consistency of implementation, but their operational results depend on data quality, professional competencies, system reliability, local administrative capabilities and the effectiveness of monitoring.
The comparison also requires a distinction between technological sovereignty as an objective and technological sovereignty as an achieved condition. Government strategies can prescribe domestic technological development, shared computing infrastructure and reduced dependence on external suppliers. Whether those policies produce reliable, affordable and sustainable capabilities is a separate empirical question. In the absence of audited operational records, announcements concerning sovereign AI systems should not be converted into claims of demonstrated independence, lower costs or higher administrative productivity.
China: the AI+ strategy and the extension of artificial intelligence into public governance
China’s State Council developed a broad policy framework for integrating artificial intelligence into economic and social functions through its AI+ Action initiative, identified as State Council Document Guo Fa [2025] No. 11. The policy establishes objectives for combining AI with science and technology, industry, consumption, public welfare, governance and international cooperation. It sets milestones extending to 2027, 2030 and 2035, including targets concerning the diffusion of intelligent terminals and AI agents. In particular, the document sets ambitions for adoption rates exceeding 70% by 2027 and 90% by 2030 for specified categories of next-generation intelligent terminals and intelligent agents. These are broad national technology-adoption objectives, not measured productivity improvements in government administration. Opinions on Deepening the Implementation of the AI+ Action — State Council of the People’s Republic of China, Document Guo Fa [2025] No. 11. xxgk.jcgov.gov.cn
The strategic significance of AI+ lies in its attempt to treat artificial intelligence as an enabling infrastructure rather than as a narrow collection of productivity applications. The policy connects technological development with changes in production, public services and administrative processes. However, the presence of broad targets can create measurement challenges. The diffusion of intelligent agents or terminals across society does not reveal how many administrative procedures have been shortened, what additional costs have been incurred or whether public employees have been able to concentrate on more demanding responsibilities. Broad technological penetration is therefore a measure of deployment or adoption, not a sufficient measure of administrative performance.
A more specific institutional framework was published on 10 October 2025, when the Cyberspace Administration of China and the National Development and Reform Commission issued the Guidelines for the Deployment and Application of Large AI Models in Government Affairs. The guidelines provide a detailed account of government use cases, deployment models, information-management requirements and operational safeguards. Their importance lies in translating the AI+ policy direction into a structure intended to guide government departments at different administrative levels. Guidelines for the Deployment and Application of Large AI Models in Government Affairs — Cyberspace Administration of China and National Development and Reform Commission, October 2025. cac.gov.cn
The document identifies several categories of applications. In public-service delivery, it describes intelligent question-answering systems, assistance with administrative applications and proactive matching between public policies and eligible individuals or enterprises. In government office work, it addresses drafting support, information retrieval and the classification and allocation of incoming tasks. It also discusses public governance, emergency management, policy evaluation and support for administrative assessments. These functions differ significantly in their technical and legal implications. An application providing information about administrative requirements is not equivalent to one assisting enforcement or determining how public resources should be allocated. The official guidelines recognise this distinction by referring to the selection of appropriate deployment approaches according to the complexity and sensitivity of the application.
The Chinese document provides an especially detailed basis for analysing the institutional allocation of technical capacity. For general applications such as document drafting and question answering, it discusses the use of mature, appropriately registered model products and services. For specialised applications, including certain regulatory or risk-related functions, it recognises the possibility of developing domain-specific capabilities using expert knowledge and specialised data. It also emphasises the requirements associated with state secrets, work-related confidential information and sensitive material. This distinction demonstrates that the policy does not prescribe a single model for every administrative function; instead, it seeks to coordinate deployment according to operational requirements and information sensitivity.
Shared infrastructure and the economics of coordinated deployment
One of the most distinctive features of China’s 2025 government AI guidelines is the explicit emphasis on centralised infrastructure, reuse and the prevention of isolated model deployments. The guidelines encourage the use of integrated computing resources and the reuse of government AI models across administrative units. They identify the possibility of central government departments and provincial authorities providing services to subordinate institutions, while stating that county-level administrations and lower levels should, in principle, reuse higher-level resources rather than independently developing and deploying separate government large models. The guidelines also promote provincial AI service platforms integrated with existing government cloud and digital-service infrastructure. cac.gov.cn
This approach addresses a potential economic inefficiency associated with fragmented technology investment. If every small administrative office independently acquires computing infrastructure, develops a specialised model and maintains separate security controls, governments may incur repeated fixed costs and create systems that are difficult to integrate. Shared infrastructure can potentially reduce duplication, concentrate technical expertise and facilitate the deployment of standard capabilities across numerous departments. It may also create opportunities for more consistent security management and coordinated updates.
Nevertheless, centralisation produces its own economic and operational dependencies. A shared infrastructure arrangement may reduce duplication while increasing reliance on a central provider whose performance affects many downstream institutions. If local administrations require specialised functions that the common platform does not support, the apparent economies of scale may be offset by adaptation costs, integration problems or reduced operational flexibility. The optimal arrangement depends on the diversity of administrative tasks, the sensitivity of information, connectivity, reliability requirements and the cost of maintaining local alternatives.
The Chinese guidelines recognise this tension through their support for both coordinated shared resources and specialised applications where justified by operational needs. The distinction matters because administrative productivity is not necessarily maximised by imposing the same technical solution on every function. A common document-retrieval or drafting capability may benefit from shared infrastructure, while an application supporting a highly specialised regulatory decision may require additional datasets, expert supervision and tailored validation.
The policy also addresses what it describes as digital formalism, warning against pursuing technological novelty without practical benefits, duplicating systems, imposing ineffective tools or repeatedly collecting the same information. This is a particularly significant aspect of the official record. The Chinese guidance explicitly identifies a risk that administrative digitalisation could create additional procedural burdens rather than reduce them. Its treatment of that problem shows that the state-directed approach recognises the possibility that technology deployment can increase, rather than decrease, bureaucratic work when organisational design is inadequate. cac.gov.cn
From the standpoint of productivity, this creates an important test for shared AI infrastructure. The correct performance indicator is not simply the number of departments connected to a common platform. It is whether those departments can use the platform to reduce unnecessary work, improve service outcomes, maintain appropriate controls and avoid transferring new burdens to other parts of the administration.
Operational safeguards, knowledge management and model reliability in China
China’s government AI guidelines also establish a structured approach to operational management. They emphasise the classification and management of government data, the creation of reliable knowledge bases, the maintenance of information provenance and the continuous updating of resources used by AI systems. These requirements have direct economic implications because the usefulness of an AI assistant depends on the quality of the information it can retrieve and interpret. An application that generates rapid answers from outdated regulations, incomplete records or inaccurate administrative guidance can increase the volume of work required to correct its mistakes.
The guidelines discuss safeguards against fabricated or unreliable outputs, including content review, real-time controls and cross-checking mechanisms appropriate to particular applications. They also identify risks associated with prompt-injection attacks, excessive consumption of computing resources and the disclosure of confidential information. Government departments are instructed to establish security responsibilities, manage access permissions, maintain operational logs and develop incident-response arrangements. These requirements describe a formal model of operational control, but the published guidance does not establish that every department has implemented them with equal effectiveness. cac.gov.cn
An important distinction concerns the relationship between public-facing services and internal administrative support. A government chatbot providing information directly to citizens creates a different exposure from an internal assistant whose output is reviewed by a professional employee. In the first case, inaccurate information may immediately influence public behaviour or administrative applications. In the second, errors may be detected before they affect an external decision, although that protection depends on the quality of review. The guidelines’ emphasis on content controls and the supporting role of AI reflects the need to distinguish these operational settings.
The formal Chinese framework therefore provides a comparatively detailed description of the intended governance architecture for government AI. Its strongest evidence concerns policy design, resource coordination, identified applications and prescribed safeguards. The principal unresolved question is implementation: how consistently departments apply the guidance, whether shared infrastructure produces lower full-life-cycle expenditure, and whether measurable improvements occur in the speed, quality and accessibility of completed public services.
China — Government AI Deployment Architecture
| Governance component | Official policy approach | Intended operational function | Economic mechanism | Main unresolved question |
|---|---|---|---|---|
| Central strategic direction | AI+ Action with milestones for 2027, 2030 and 2035 | Integrate AI across economic and public functions | Coordinated investment and diffusion | Which milestones translate into measured administrative outcomes? |
| National government AI guidance | October 2025 CAC–NDRC guidelines | Define applications, deployment methods and safeguards | Standardised implementation expectations | How consistently is guidance applied? |
| Shared computing resources | Reuse of higher-level government infrastructure | Reduce unnecessary independent systems | Potential economies of scale | Are savings greater than central operating and integration costs? |
| Provincial AI service platforms | Coordinated provision of models and resources | Support subordinate institutions | Shared technical capacity | Can specialised local needs be met effectively? |
| Government knowledge bases | Data quality, provenance and continuous updating | Improve reliability of AI outputs | Less search time and fewer errors | How are quality and maintenance costs measured? |
| Security and confidentiality | Data classification, access control and operational monitoring | Protect government information | Reduced exposure to operational incidents | What is the verified level of implementation? |
| Model-output supervision | Review, validation and controls against unreliable outputs | Maintain accuracy and administrative confidence | Lower correction and rework costs | Are controls effective in real deployment? |
| Performance evaluation | Monitoring of effectiveness and system operation | Identify benefits and deficiencies | Continuous improvement | Are end-to-end service metrics published? |
Source: Government Large-Model Deployment Guidelines — CAC and NDRC, October 2025. The economic mechanisms are analytical interpretations of the official governance arrangements, not independently verified realised savings.
Russia: digital public administration and the strategic importance of common state systems
Russia’s institutional framework is best examined through the government’s strategic direction for the digital transformation of public administration, approved by Government Order No. 637-r of 16 March 2024. The document established objectives extending to 2030 and assigned implementation responsibilities to the Ministry of Digital Development together with relevant federal executive authorities and state institutions. Its scope extends beyond generative AI and includes digital public services, electronic document management, government information systems, administrative interaction and infrastructure modernisation. This broader scope is essential: the strategy should not be interpreted as an AI-specific productivity programme simply because certain forms of automation and advanced digital technology can support its implementation. Strategic Direction for the Digital Transformation of Public Administration, Government Order No. 637-r — Government of the Russian Federation, March 2024. static.government.ru
The strategy identifies targets concerning the digitisation of government services and the expansion of electronic administrative processes. One particularly significant objective is the provision of 100 government services online with results available at the time of application by 2030. This target differs fundamentally from an internal measure of employee productivity. It concerns the experience of the service recipient and the completion of an administrative transaction. Its achievement would depend on the ability of the relevant systems to retrieve reliable information, apply appropriate procedural requirements and provide valid results without unnecessary delays. The official target does not establish that those 100 services already operate in this manner. static.government.ru
The document also includes a target of 100% electronic document management for local-government bodies and state and municipal institutions by 2030, alongside objectives concerning interoperability and the development of a common information environment for administrative interaction. It addresses modernisation of government workstations and digital systems, including measures intended to simplify routine work and increase technological independence. These objectives are important because they concern the infrastructure through which AI applications could become useful at scale. A generative-AI assistant integrated into an administration with fragmented records and incompatible information systems may struggle to produce reliable results, while a more coherent digital environment can reduce some of the technical barriers to automation. static.government.ru
The significance of Russia’s strategy is consequently not confined to the potential acceleration of individual administrative tasks. Its central economic proposition is that common digital systems, interoperable records and increasingly automated processes can change how government institutions exchange information and deliver services. This approach places a substantial premium on infrastructure coordination and the reduction of duplicated administrative activity. It also creates dependencies on the institutions responsible for operating the shared systems.
A government department that receives data through an integrated platform may be able to avoid repeatedly requesting the same information from applicants. An electronic service that can verify required records automatically may reduce the need for manual processing. However, these potential gains must be distinguished from measured outcomes. The existence of a national target or infrastructure programme does not establish the completion rate, cost, reliability or accessibility of the resulting services.
Technological sovereignty and the economic consequences of restricted supplier choice
The Russian strategy also addresses technological independence, including the substitution of certain technologies and the reduction of risks associated with external restrictions. Such objectives are particularly relevant when assessing the costs and resilience of government information systems. Reliance on foreign software, hardware or cloud services can create exposure to procurement disruption, licensing restrictions, maintenance limitations and difficulties obtaining updates. Domestic alternatives may reduce some of these dependencies, but their economic value depends on performance, security, cost, interoperability and the ability to support operational requirements.
The analytical distinction between security of supply and economic efficiency is crucial. A domestically controlled technology may offer greater institutional control over certain operations while imposing higher acquisition, development or maintenance costs. Conversely, a less expensive external solution may create dependencies that become costly under changing legal or commercial conditions. Neither outcome can be determined solely from the nationality of the supplier. The relevant assessment requires a complete examination of the available alternatives, the likelihood and consequences of disruption, and the costs of maintaining adequate service continuity.
Russia’s digital-administration strategy establishes technological independence as an objective, but the reviewed record does not provide a comprehensive audited comparison of the costs and performance of all domestic and foreign alternatives used by public authorities. It would therefore be inappropriate to interpret the strategic emphasis on domestic systems as proof that Russia has achieved complete technological autonomy or that its government AI deployments generate a measurable financial advantage.
The same caution applies to the relationship between electronic administration and AI. Digital document management, service portals, data exchange and automated workflows can all improve administrative performance without necessarily using modern generative models. Attribution requires identifying which technology produced a particular change and whether that change occurred independently of simultaneous organisational reforms. A reduction in service completion time following the introduction of a digital platform cannot automatically be credited to AI if other components of the transformation contributed to the result.
Russia’s administrative targets: status, definitions and accountability
| Strategic indicator | Official target or direction | Target horizon | Correct interpretation | Evidence required to establish achievement |
|---|---|---|---|---|
| Instant-result online services | 100 government services | By 2030 | Target for immediate-result digital services | Verified list of operational services and observed results |
| Electronic document management | 100% coverage of specified local, state and municipal bodies | By 2030 | Coverage objective within the defined institutional scope | Implementation and operational records |
| Interagency information exchange | Development of a unified information environment | Strategic period to 2030 | Infrastructure and interoperability objective | Technical integration and performance evidence |
| Government workplace modernisation | Modernised digital working environment | Strategic period | Technological capability objective | Deployment and usability records |
| Routine administrative automation | Simplification and automation of specified processes | Strategic period | Intended operational improvement | Before-and-after process measurements |
| Technological independence | Reduced dependence on selected external technologies | Strategic period | Policy objective | Verified supplier, infrastructure and resilience data |
The distinction between strategic objectives and verified operational outcomes is particularly important for international comparisons. The Russian figures cannot be placed alongside measured British employee time savings or French experimental task savings as though they described the same economic phenomenon. They provide evidence of planned digital-government capabilities and the direction of institutional investment.
China and Russia compared: coordination does not eliminate the need for verification
China’s and Russia’s frameworks both place importance on coordinated digital infrastructure, but the available official documents emphasise different implementation functions. China’s October 2025 guidelines provide detailed prescriptions for deploying large AI models in government, including shared resources, model reuse, sensitive information and output control. Russia’s Government Order No. 637-r establishes broader digital-administration objectives, including instant-result services, electronic document management and technological independence. The first provides direct guidance on government large-model deployment; the second describes a digital-government transformation programme within which various automation technologies may be used.
This distinction affects the analytical unit of comparison. China can be assessed through the governance of large-model deployments and the administrative arrangements intended to support them. Russia can be assessed through digital-service targets and the infrastructure necessary for increasingly automated administration. Neither official record, by itself, establishes a national causal estimate of working time recovered by AI or the allocation of that capacity among government functions.
The principal common feature is the potential to reduce duplicated infrastructure and improve coordination among public institutions. The principal common limitation is that centrally established targets do not substitute for independent operational measurement. Central authorities may have the capacity to establish uniform standards or require shared systems, but the benefits still depend on accurate data, effective local implementation, reliable professional oversight and services that actually meet the needs of citizens and enterprises.
Comparative state-coordination and infrastructure architecture
AI and Digital Administration: Two Coordination Models
Qualitative institutional structures derived from official Chinese government AI guidance and Russia’s digital-government strategy. No performance ranking is implied.
Strategic integration of AI into public governance and economic development.
Common infrastructure, shared platforms and model reuse across administrative levels.
Public services, document processing, task allocation and decision support.
Operational results, security, cost, data quality and measurable service outcomes.
Government-wide direction for service digitisation and information infrastructure.
Electronic records, information exchange and government infrastructure.
Electronic document management, simplified procedures and instant-result service targets.
Verified target achievement, service completion times and full operating costs.
Common Analytical Constraint
Central direction and common platforms can influence deployment costs and technical coordination. Their existence does not demonstrate achieved productivity, reliable service quality or fiscal savings.
Sources: CAC–NDRC Government AI Guidelines, October 2025; Russian Government Order No. 637-r, March 2024. Analytical scheme, not an official organisation chart.
Chapter 5 — Key judgments and consequential evidence gaps
China’s official government AI guidance establishes an extensive framework for coordinated large-model deployment, shared computing resources, information quality, operational safeguards and the prevention of duplicative technological investment. Russia’s digital-administration strategy establishes measurable targets for electronic government services and common administrative systems through 2030. Both frameworks connect technological development with the organisation of state functions, but they differ in scope and the specificity of their AI provisions.
Neither framework establishes a comparable nationwide estimate of AI-generated working time savings, the allocation of the resulting capacity or the net fiscal benefit of public-sector AI deployment. Future assessment requires implementation records, operational cost data, service completion metrics and reliable evidence of whether shared infrastructure improves performance across administrative levels.
Open official record: verified deployment and cost reports, service-level performance measurements, independent evidence of target achievement, and records establishing the operational effectiveness of centrally coordinated AI and digital-government systems.
Chapter 6. Shadow AI, Human Capital, Cybersecurity and the Institutional Value of Time
The invisible adoption problem: when employee innovation develops faster than institutional control
The international expansion of generative artificial intelligence has created a governance problem that is only partially captured by formal deployment statistics. Public employees can access powerful AI capabilities through commercial subscriptions, personal accounts, browser interfaces and software applications without waiting for their institutions to establish approved deployment programmes. Such use can create immediate operational benefits, especially in drafting, translation, research, summarisation and document preparation. However, when the institution does not know which systems are being used, what information they receive or how their outputs are verified, the resulting activity becomes difficult to assess from the perspectives of security, legality, accountability and productivity. The principal finding is that unmanaged AI adoption can create genuine but institutionally invisible efficiency gains, while simultaneously introducing risks that may offset or reverse those gains. The appropriate governance response must therefore address not only the technology itself but also the organisational incentives that encourage employees to use unapproved tools.
The term shadow AI describes the use of AI tools or capabilities outside established organisational approval, procurement or governance arrangements. It overlaps with the broader concept of shadow IT but creates additional technical issues because generative models may receive sensitive text, interpret instructions embedded in untrusted material and produce plausible but inaccurate outputs. The risk is not limited to the disclosure of information. An unapproved system can influence administrative work even when the institution has not assessed its accuracy, retained appropriate records or established procedures for reviewing consequential outputs. This creates uncertainty about which activities depend on AI, whether decisions remain reproducible and whether measured employee performance reflects changes in professional work, new software assistance or both.
The available official record does not establish a harmonised global prevalence rate for shadow AI in public administration. Surveys of particular employee populations cannot automatically be generalised across governments, and the absence of recorded unauthorised use does not demonstrate that such use is absent. Institutions also differ in how they define AI systems, distinguish embedded software capabilities from standalone generative services and record the use of personal accounts. Consequently, claims that a specific proportion of all government employees use shadow AI would require a clearly defined and representative evidence base that is not established by the official materials examined here.
The underlying governance mechanism can nevertheless be assessed through competent cybersecurity guidance. The Australian Signals Directorate’s Australian Cyber Security Centre, working with international cybersecurity partners including US, British, German, Canadian, Japanese and other agencies, issued guidance on secure engagement with artificial intelligence. The document identifies considerations for organisations using internally hosted and third-party AI systems, including the handling of sensitive information, exposure to external services and the need to evaluate the risks of implementation. The multinational character of the guidance is relevant because government organisations face similar technical attack surfaces even when their administrative laws and institutional structures differ. Engaging with Artificial Intelligence — Australian Signals Directorate and International Cybersecurity Partners, January 2024. cyber.gov.au
The principal institutional dilemma is that employees may discover useful applications before their organisations can authorise them. A blanket prohibition on all unapproved tools may reduce certain immediate exposures, but it can also leave practical demand unresolved and encourage informal workarounds. Unrestricted use produces the opposite problem: the institution may benefit from higher individual productivity while losing visibility over data processing, supplier dependence and the provenance of administrative outputs. The more sustainable approach is to identify legitimate operational requirements, provide suitable authorised capabilities and maintain controls proportionate to the information and decisions involved.
This distinction is particularly important in public administration because users may process personal information, confidential business records, internal legal assessments, public procurement materials or documents relevant to regulatory decisions. Their responsibility for protecting such information does not disappear when a convenient AI interface becomes available. An authorised tool must be evaluated in relation to the actual data and functions involved, while employees require practical guidance on what may be entered, what must be verified and how the resulting material may be used.
Shadow AI as an information-governance and productivity-accounting problem
The economic implications of shadow AI extend beyond the immediate possibility of a cybersecurity incident. When employees independently use AI to accelerate work, the organisation may fail to identify the source of an improvement, reproduce it elsewhere or determine whether the gain is sustainable. A unit that produces reports more quickly may appear to have improved its internal procedures when the actual change results from employees using external applications outside approved systems. If the institution subsequently changes staff assignments or performance expectations without recognising that dependency, it may create unrealistic operating assumptions.
A related issue concerns information asymmetry between employees and management. Employees may possess direct knowledge of the tasks for which AI is useful, while managers control procurement, information access and formal implementation. If the governance process is too slow to incorporate operational experience, useful applications may remain outside institutional visibility. Conversely, employees may not know the technical or legal risks of providing documents to external systems. Effective governance requires mechanisms through which operational innovation can be evaluated without allowing individual convenience to override institutional safeguards.
The economic cost of an unauthorised deployment is also difficult to assess when external services are purchased through personal accounts or embedded in software subscriptions. Costs may be paid privately, reimbursed informally or concealed within broader software expenditure. Such arrangements can prevent the institution from understanding the resources required to reproduce the workflow, negotiating appropriate contractual terms or ensuring service continuity if an employee leaves. Where AI becomes an essential component of a process without documented institutional ownership, a personal productivity practice may evolve into an unrecognised operational dependency.
The correct response should therefore distinguish between the discovery of a useful application and the authority to deploy it on institutional information. The first can originate from an employee’s practical knowledge; the second requires assessment under the applicable administrative, information-security, procurement and data-governance arrangements. This distinction creates an avenue for institutional learning without treating unauthorised use as automatically suitable for continued operation.
Cybersecurity: the principal mechanisms through which AI can destroy its apparent productivity dividend
AI-related cybersecurity risks should be analysed according to the technical mechanisms through which they can affect government information systems and administrative decisions. The National Institute of Standards and Technology’s Generative Artificial Intelligence Profile, NIST AI 600-1, published on 26 July 2024, provides an official framework for examining risks associated with generative AI, including information integrity, data privacy, security, system reliability and the management of generated outputs. The document is a companion to NIST’s AI Risk Management Framework and is intended to help organisations incorporate relevant safeguards into the design, deployment and evaluation of generative-AI applications. It is a voluntary framework rather than a universally binding legal instrument. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 — National Institute of Standards and Technology, July 2024. NIST
A particularly important attack mechanism is prompt injection, in which instructions embedded within material processed by an AI system attempt to influence its behaviour outside the legitimate task. A government assistant may retrieve a document, website or email that contains text designed to manipulate the model’s subsequent output. The threat is especially consequential for systems that can interact with external applications, retrieve sensitive information or initiate actions in other software environments. The resulting risk depends on the permissions available to the application, the separation between trusted instructions and untrusted data, and the effectiveness of independent controls.
The distinction between a prompt-injection attempt and a successful compromise is important. The presence of malicious instructions in a document does not establish that a system has executed them or disclosed information. Equally, the absence of observed incidents does not demonstrate that the underlying attack surface is secure. A credible evaluation must determine whether the system can identify untrusted material, restrict its permissions and prevent generated text from directly authorising sensitive operations. These are technical control questions that cannot be resolved through general employee awareness training alone.
Another major risk concerns the confidentiality of information submitted to AI services. The institution must know which data are transferred to external systems, under what contractual and technical arrangements they are processed, and whether their use complies with applicable legal requirements. The problem is not limited to model training. Information may be retained in application logs, processed by additional service components, transmitted between providers or made available to personnel or systems under contractual arrangements that differ from the institution’s expectations. The relevant assessment must therefore examine the actual service architecture rather than relying on broad claims that a tool is private or secure.
Generated misinformation and professional overreliance create a different class of exposure. An AI assistant can produce fluent but inaccurate interpretations of legislation, administrative procedures or technical records. If employees fail to verify such outputs, the initial time saving may be offset by incorrect decisions, additional correspondence, legal challenges or the need to repeat work. These effects can occur long after the original task has been completed, making them difficult to capture through immediate employee productivity measurements.
Excessive permissions and uncontrolled automation become particularly significant as organisations move toward agentic systems. A tool that merely produces text for human review has a different operational risk profile from one authorised to retrieve confidential files, modify records, send communications or initiate workflow actions. The latter can create larger productivity opportunities because it reduces the need for manual transfers between systems. It also expands the consequences of configuration errors, unreliable output and compromised instructions. Appropriate controls must distinguish advisory functions from actions that change institutional records or affect external parties.
The Chinese government AI guidelines explicitly recognise several of these mechanisms, including prompt injection, attacks that consume excessive resources, management of sensitive data, operational logging and the need for incident-response arrangements. Their presence in an official government deployment framework demonstrates that these risks are not exclusive to Western procurement environments; they arise from the technical characteristics of AI applications and the information systems with which they interact. Government AI Deployment Guidelines — Cyberspace Administration of China and National Development and Reform Commission, October 2025. cac.gov.cn
Cybersecurity and Shadow AI Risk Matrix
| Risk mechanism | Operational exposure | Possible administrative consequence | Primary control objective | Performance indicator |
|---|---|---|---|---|
| Unauthorised external AI use | Institutional data entered into unapproved services | Loss of information-control visibility | Identify use and provide approved alternatives | Authorised adoption and investigated incidents |
| Prompt injection | Untrusted documents or retrieved content influence model behaviour | Incorrect output or unauthorised downstream actions | Separate untrusted content from privileged instructions | Adversarial-test outcomes |
| Sensitive-data disclosure | Information processed outside approved arrangements | Confidentiality breach and legal exposure | Data classification, access restrictions and supplier controls | Confirmed incidents and compliance findings |
| Inaccurate generated content | Fabricated or unsupported information enters official work | Errors, rework and potentially unlawful outcomes | Verification and evidence checking | Error and correction rates |
| Excessive system permissions | AI applications can modify records or initiate actions | Operational disruption or improper transactions | Least privilege and independent authorisation | Permission-review findings |
| Uncontrolled supplier dependencies | Critical workflows rely on external platforms | Service interruptions and rising switching costs | Portability and continuity planning | Recovery and migration-test results |
| Model or knowledge-base deterioration | Outputs become less reliable as information changes | Increasing correction effort | Continuous evaluation and data maintenance | Quality measurements over time |
| Incomplete AI inventories | Undocumented applications and information flows | Weak management and audit visibility | Comprehensive registration and periodic reconciliation | Inventory completeness and exceptions |
The table identifies control objectives rather than asserting a uniform legal requirement for every administration. Controls must be selected according to the actual system, sensitivity of information, jurisdiction and potential effects on affected individuals.
The economics of cybersecurity: why prevention and validation belong in productivity calculations
Cybersecurity is often treated as an additional cost of adopting artificial intelligence, but this accounting description is incomplete. Security measures influence the reliability and continuity of the administrative capability being acquired. An AI application that generates documents more quickly while introducing unacceptable exposure to confidential records does not provide the same economic value as a system that produces equivalent outputs within appropriately controlled arrangements. Similarly, an application that requires frequent correction, incident investigation or emergency suspension may impose costs not captured in the initial estimate of time saved.
A comprehensive productivity calculation should therefore consider both routine operating costs and the potential consequences of incidents. Routine costs include access management, logging, security reviews, model testing, employee training and the maintenance of approved systems. Incident-related costs may include investigation, recovery, service disruption, legal response and the correction of affected outputs. The existence of these cost categories does not justify assigning arbitrary monetary values to potential incidents. Instead, institutions should document observed expenditure and use defensible risk assessments when considering uncertain future losses.
The relationship between security investment and operational efficiency is not necessarily negative. A well-designed approved AI environment can reduce the need for fragmented procurement, support consistent access arrangements and provide a reusable basis for multiple administrative applications. Conversely, complicated approval systems that are poorly aligned with practical work may increase administrative overhead and encourage avoidance. The economic objective is to build controls that are effective, proportionate and integrated with operational processes rather than appended as repeated manual procedures.
This principle becomes particularly important when government organisations seek to expand AI capabilities across departments. A centrally managed service can spread fixed security and infrastructure costs over multiple applications, but it can also concentrate operational dependencies. The appropriate evaluation must therefore examine both economies of scale and the consequences of failure at shared components. A single central platform may reduce duplication while creating a larger potential interruption if essential infrastructure becomes unavailable.
Human capital: why workforce capability determines the realised value of technological investment
Human capital is one of the principal constraints on the productivity of public-sector AI because deploying technology does not automatically create the professional skills required to use, supervise and maintain it. An AI assistant can accelerate technical operations, but employees must still recognise whether the output is reliable, relevant and appropriate to the administrative function. Managers must identify suitable uses and understand their effects on workflows. Technical staff must maintain systems and integrate them with existing data resources. Legal and security professionals must evaluate compliance and information risks. The productivity dividend therefore depends on complementary investment in competencies that differ substantially among professional roles.
The GAO’s March 2026 IRS audit provides direct institutional evidence of this relationship. The investigation found that staffing reductions had contributed to shortages of employees with the expertise required to develop and support AI tools. It also identified the absence of a suitable workforce plan for assessing and addressing AI-related skills needs. These findings establish that human-capital constraints were not merely theoretical concerns within the examined agency. They affected an institution that had already identified numerous AI applications and was considering additional deployment. Artificial Intelligence: IRS Actions Needed to Address Skills Gaps, Information Quality, and Strategic Management — GAO, March 2026. U.S. Government Accountability Office (GAO)
The implications extend beyond the number of employees possessing specialised technical qualifications. Effective AI adoption requires a distribution of competencies across different organisational functions. A public manager needs sufficient understanding to evaluate whether an application addresses a genuine workflow constraint and whether its benefits are appropriately measured. A legal professional needs to recognise where generated assessments may be unreliable or inconsistent with applicable law. A procurement specialist needs to understand the consequences of contractual terms governing information access, intellectual property and supplier replacement. An information-security specialist must assess the permissions and data flows associated with AI services. These functions require different forms of expertise and cannot all be replaced by a general course on writing effective prompts.
Training should therefore be evaluated according to demonstrated changes in professional capability rather than attendance alone. Completion of a training module can establish that employees received instruction, but it does not necessarily demonstrate that they can identify an unreliable AI output, respond appropriately to a suspected prompt-injection attempt or distinguish a low-risk drafting application from a system requiring more substantial oversight. More informative assessments would use practical exercises linked to the relevant administrative responsibilities, combined with observed performance after training.
The timing of workforce investment is also economically consequential. Introducing AI capabilities before employees understand their limitations can create rework and security exposure. Delaying deployment until every employee has advanced technical expertise would be equally impractical and could prevent institutions from obtaining useful benefits from lower-risk applications. A more proportionate approach would establish baseline competencies for authorised use and provide specialised training to personnel responsible for consequential decisions, technical management and oversight.
Workforce transformation: substitution, complementarity and emerging organisational dependencies
The effect of artificial intelligence on public employment cannot be reduced to a single relationship between automation and workforce reduction. AI may substitute for particular activities while complementing the professional capabilities required to complete an administrative function. For example, automated information retrieval may reduce the manual time needed to assemble relevant documentation, while making it possible for a legal professional to devote greater attention to interpretation and case-specific reasoning. Similarly, automated document classification may reduce routine handling but increase the importance of personnel capable of investigating unusual cases or correcting classification errors.
The resulting distribution of work depends on which tasks are accelerated, which remain necessary and whether the institution changes its operating procedures. An application that reduces routine workload may allow an organisation to accommodate growing demand without increasing staffing. It may instead improve quality or reduce pressure on employees. In some circumstances, it may support lower future expenditure, but that outcome requires a demonstrated change in resource requirements. These possibilities cannot be assumed from technical performance alone.
AI can also create new professional functions. Model evaluation, data curation, system integration, incident response, supplier management and process redesign require continuing attention. The growth of these functions means that some apparent labour savings may be offset by new workloads. This does not necessarily make the investment economically undesirable: the new activities may support better services and more reliable institutions. It does mean that gross task savings must be distinguished from net changes in organisational effort.
A further issue concerns the concentration of expertise. If only a small group of technical specialists understands the systems on which an administration increasingly depends, operational continuity may become vulnerable to staff turnover or changes in suppliers. The relevant workforce objective is therefore not simply to maximise the number of employees using AI, but to ensure that institutions retain sufficient distributed competence to operate, supervise and modify their systems.
Human Capital and AI Governance — Functional Competency Matrix
| Professional function | Essential competency | Operational responsibility | Evaluation method | Risk of inadequate capability |
|---|---|---|---|---|
| Administrative leadership | Workflow analysis, evaluation and institutional risk | Determine appropriate deployment and resource allocation | Service-level performance review | Adoption without operational benefit |
| Legal professionals | AI limitations, evidence validation and applicable law | Assess legal suitability of AI-assisted work | Case-based evaluation and quality review | Unsupported or unlawful reasoning |
| Public-service employees | Authorised use, verification and data handling | Perform tasks using approved systems | Practical exercises and work-quality assessment | Incorrect outputs or sensitive-data exposure |
| IT and systems engineering | Integration, reliability and model operations | Maintain secure and functional systems | Testing and operational monitoring | Service interruption or technical dependence |
| Cybersecurity specialists | Threat modelling, access control and incident response | Protect systems and information | Security assessments and response exercises | Compromise or uncontrolled permissions |
| Data professionals | Provenance, data quality and maintenance | Ensure reliable information inputs | Dataset audits and quality measurements | Inaccurate or outdated outputs |
| Procurement professionals | Supplier assessment, portability and lifecycle economics | Acquire sustainable capabilities | Contract review and total-cost evaluation | Vendor lock-in and uncontrolled expenditure |
| Internal audit and evaluation functions | Evidence quality, attribution and benefit measurement | Examine reported results | Reproducible evaluations | Misleading productivity or fiscal claims |
A particular advantage of this differentiated approach is that it avoids treating training as a homogeneous input. A frontline employee, a legal specialist and an AI infrastructure engineer do not require identical competencies merely because they operate within the same institution. Their training should be tailored to the decisions and responsibilities associated with their functions, with the institution maintaining sufficient understanding across organisational boundaries to coordinate their work.
Shadow AI, security controls and institutional learning
Converting Shadow AI into Accountable Capability
Proposed process for evaluating useful employee-led AI practices before institutional deployment.
Potential reduction in routine work, better information retrieval or improved drafting.
Identify the actual task
Record the purpose, input information, expected output and observed benefit. Do not require confidential data for initial screening.
Evaluate data, permissions and accuracy
Examine confidentiality, legal requirements, supplier arrangements, model limitations and potential consequential effects.
Test within approved conditions
Measure quality, task duration, correction costs and practical suitability using authorised systems.
Institutional Adoption
Use approved tools, establish human responsibility, record deployment and monitor actual outcomes.
Restrict or Redesign
Prevent sensitive use, document the reason and assess whether a safer alternative exists.
Analytical model. Actual approval authority depends on the administration and jurisdiction. Reference standards: NIST AI 600-1 and International Guidance on Engaging with AI.
The institutional value of time: incorporating security, competencies and resilience into economic evaluation
The international evidence establishes that the economic performance of government AI systems cannot be interpreted independently of the conditions necessary for their reliable operation. Cybersecurity, professional competencies, data quality and technological continuity are not external considerations added after the productivity calculation. They influence whether the underlying productivity gain exists, how long it lasts and whether it can be converted into a sustainable improvement in administrative services.
This creates a distinction between gross efficiency and net institutional value. Gross efficiency concerns the reduction in resources required to carry out a particular activity under specified conditions. Net institutional value also considers the work necessary to maintain the system, verify its output, train employees, manage information security and respond to errors. It must further account for the contribution of the application to final service outcomes. A tool that reduces drafting time substantially but requires repeated correction may generate a smaller net benefit than its initial performance suggests. An application that creates a modest time saving while substantially improving the consistency of administrative information may produce value through a different mechanism.
A rigorous economic evaluation should therefore begin with a clearly defined operational baseline. The institution must know how much time, expenditure and professional effort the relevant activity required before deployment. It must then establish which components change after implementation and whether the difference can reasonably be attributed to AI rather than to unrelated staffing, procedural or technological reforms. It should also record the resources necessary for ongoing operation, including software expenditure, infrastructure, technical maintenance, oversight and employee training.
A second layer concerns the distribution of benefits and costs across departments. An AI-enabled unit may produce documents more quickly while transferring additional verification work to another department. If only the first unit is evaluated, the resulting benefit can be overstated. Similarly, centralised infrastructure may reduce expenditure within individual offices while increasing costs at the shared service provider. The correct accounting boundary must encompass the units whose resources materially change because of the deployment.
The third layer concerns institutional resilience. Public services may need to continue operating during supplier interruptions, system failures, cybersecurity incidents or changes in the availability of qualified personnel. A system that offers substantial efficiency gains under normal conditions but introduces an unacceptable risk of service interruption may have a less favourable economic profile than a more conservative alternative. Resilience should therefore be evaluated through recovery arrangements, data portability, continuity testing and the ability to maintain essential services when AI capabilities are unavailable.
The fourth layer concerns the sustainability of workforce competencies. If employees increasingly rely on AI to produce assessments, summaries or technical material, institutions should monitor whether they retain sufficient independent expertise to detect errors and manage exceptional cases. A reduction in routine manual activity may be beneficial, but it could also reduce opportunities to develop certain professional skills if training and supervision are not adapted. The significance of this possibility varies across occupations and should be evaluated through observed performance rather than assumed as an inevitable consequence of automation.
International governance comparison: operational control of AI-generated capacity
| Governance dimension | United States | China | Russia | Cross-jurisdictional implication |
|---|---|---|---|---|
| Principal policy mechanism examined | OMB M-25-21 and M-25-22 | AI+ policy and 2025 government LLM guidelines | Government Order No. 637-r | Different institutional instruments govern deployment |
| Deployment coordination | Central federal requirements with agency-level implementation | Explicit coordination and reuse of model and computing resources | Common digital-government systems and strategic direction | Coordination models influence integration and duplication |
| Procurement and supplier control | Explicit attention to competition, portability and interoperability | Selection of appropriate model services and shared resources | Technological independence and common infrastructure objectives | Supplier choices affect costs and resilience |
| Risk management | High-impact AI category and agency responsibilities | Lifecycle controls, content safeguards and confidentiality requirements | Digital-system protection and infrastructure modernisation | Actual implementation requires separate verification |
| Workforce capability | GAO identifies concrete agency skills and planning gaps | Official guidance requires training and improved staff competencies | Strategy includes modernisation of government working systems | Skills and operational readiness are distinct from software availability |
| Official quantitative evidence | AI use-case inventories and agency audits | Technology adoption targets and deployment guidance | Digital-service targets to 2030 | Indicators cannot be directly combined |
| Principal productivity gap | Incomplete connection between adoption and measured mission outcomes | Limited comparable service-level AI productivity evidence | Limited AI-specific causal productivity evidence | End-to-end performance remains the key comparative requirement |
The comparison shows why national AI governance should not be reduced to a distinction between centralised and decentralised systems. The practical effectiveness of either arrangement depends on the quality of the institutions operating within it. A centrally coordinated system can reduce duplication but may concentrate dependencies. Agency autonomy can encourage specialised applications but may produce inconsistent standards or repeated procurement. The appropriate economic assessment concerns the relationship between the chosen governance arrangement and the actual performance, cost and reliability of the services it supports.
Strategic scenarios for 2027–2031: three non-exclusive development pathways
Three analytically distinct developments warrant monitoring over the five-year horizon. They are not mutually exclusive, and no numerical probabilities are assigned because the available official evidence does not provide a sufficiently comparable historical base rate for reliable international forecasting.
The first pathway is controlled institutional integration. Under this development, administrations increasingly provide authorised AI environments, establish clear responsibilities for deployment and improve the connection between technical systems and existing workflows. Employee-led experimentation becomes an input to formal evaluation rather than remaining permanently outside institutional oversight. The mechanism supporting this pathway is the combination of approved tools, organisational learning and investment in technical competencies. Evidence favouring it would include improvements in application inventories, fewer recurring control deficiencies, reliable system-performance reports and documented reductions in complete-service processing times.
The second pathway is fragmented technological expansion. Under this development, institutions deploy growing numbers of AI applications without adequate integration, common data standards or clear ownership of operational outcomes. Employees obtain local efficiency gains, but organisations incur duplicated procurement, inconsistent information controls and additional management burdens. This pathway is compatible with rapid growth in reported use cases and would not necessarily be visible through adoption statistics alone. Indicators would include persistent inventory inaccuracies, repeated procurement of overlapping capabilities, rising integration expenditure and limited improvement in final service performance despite extensive deployment.
The third pathway is capability concentration and resilience pressure. Under this development, governments increasingly depend on a relatively limited number of technological platforms, suppliers or internal centres of expertise. Shared infrastructure may offer economies of scale, but operational continuity becomes more sensitive to supplier arrangements, cyber incidents, platform availability and shortages of specialised personnel. Evidence supporting this pathway would include increasing concentration of critical workloads, difficulties migrating between systems, shortages of employees with necessary technical competencies and service disruptions associated with shared infrastructure dependencies.
These pathways demonstrate that the future value of public-sector AI is not determined by the sophistication of individual models alone. Institutional design influences whether benefits can be reproduced, whether costs remain manageable and whether administrations retain sufficient control over their responsibilities.
Decision framework: what governments should measure before expanding deployment
A useful expansion decision requires evidence concerning five distinct questions: whether the application performs its intended function; whether its outputs are sufficiently reliable for the proposed use; whether the institution can operate and supervise it sustainably; whether the resulting changes improve relevant administrative outcomes; and whether the total benefits justify the costs and risks associated with implementation.
The decision should also be proportionate to the possible consequences of failure. A drafting assistant used for non-sensitive internal correspondence may be suitable for a relatively streamlined evaluation, provided appropriate information controls are maintained. An application influencing eligibility for public benefits, regulatory enforcement or another consequential determination may require much more extensive validation and oversight. The existence of an AI system that performs well on a general benchmark does not establish its suitability for a specialised administrative function.
The decision to expand deployment should therefore be conditional rather than automatic. Successful task-level tests can justify further examination, but operational scaling should depend on evidence that the institution can reproduce the benefit, maintain the necessary safeguards and manage the effects on downstream work. Where these conditions are absent, a limited pilot or redesign may be more appropriate than immediate expansion.
Chapter 6 — Key judgments
The proliferation of accessible AI tools has made the boundary between individual experimentation and formally authorised institutional adoption increasingly consequential. Shadow AI can create useful operational improvements, but it also introduces uncertainties concerning data handling, output reliability, procurement, oversight and the reproducibility of working methods. The principal governance requirement is to make legitimate innovation visible and assessable without assuming that every employee-selected tool is suitable for processing institutional information.
Cybersecurity and human capital are central determinants of the net economic value of AI deployment. The NIST framework and multinational cybersecurity guidance provide official foundations for assessing technical risks, while the GAO’s 2026 IRS audit demonstrates that workforce shortages and inadequate application records can materially constrain an agency’s implementation capability. These findings establish that expanding access to AI systems without corresponding organisational and professional capacity may fail to produce sustainable performance improvements.
The assessment would materially change with representative public-sector evidence concerning unauthorised AI use, measured rates of security and quality incidents, the effectiveness of approved alternatives, workforce capability development and the full organisational costs of deployed systems.
Open official record: representative and clearly defined shadow-AI assessments, verified incident and correction costs, workforce-readiness evaluations, authorised-tool adoption records, system-resilience testing and sustained service-performance measures.
Pillar II — Consolidated International Net Assessment
The comparative evidence from the United States, China and Russia demonstrates that governments can adopt substantially different institutional arrangements for artificial-intelligence deployment while facing related economic and operational constraints. The United States has documented rapid growth in federal AI use cases and revised executive-branch requirements concerning adoption, procurement and risk management. China has developed detailed guidance for coordinating large-model deployment across government departments and administrative levels, emphasising common infrastructure, resource reuse, data governance and safeguards against ineffective digitalisation. Russia has established digital-government objectives extending to 2030, including electronic document management, interoperability and the provision of selected public services with immediate online results.
These arrangements should not be interpreted as equivalent evidence of achieved productivity. The American record contains comparatively detailed use-case inventories and agency audit findings, but many reported applications were not fully operational and important workforce and information-quality weaknesses remained. China’s official record establishes policy requirements and an extensive government AI deployment architecture, but does not provide a comparable nationwide estimate of how much public-sector working time has been recovered and how that capacity has been used. Russia’s strategy provides measurable digital-service objectives, but the reviewed record does not permit those objectives to be attributed specifically to generative AI or converted into realised fiscal savings.
The cross-jurisdictional evidence also reveals a problem that cannot be resolved solely through procurement or technological regulation. Public institutions increasingly depend on the ability of employees to understand model limitations, identify unreliable outputs, protect sensitive information and determine when AI-assisted work requires additional professional judgement. Cybersecurity, human capital and organisational resilience therefore influence the amount of usable capacity that remains after the costs of verification, maintenance and operational control are considered.
The central strategic conclusion is that AI-generated productivity should be evaluated as a net institutional capability rather than a gross technological saving. The relevant outcomes are not limited to the number of models acquired, employees equipped with assistants or applications entered into national inventories. They include the reliability of completed services, the quality of administrative decisions, the ability to maintain essential operations, the allocation of recovered professional capacity and the full expenditure required to sustain those improvements.
Over the 2027–2031 horizon, the strongest evidence of successful government AI deployment will come from administrations capable of demonstrating a consistent relationship between authorised technology, available professional competencies, manageable security risks and measurable improvements in public functions. This is the basis on which future claims of higher government productivity, more efficient expenditure and improved administrative performance can be evaluated.
OPEN-SOURCE ANALYTICAL ASSESSMENT | 20 NOVEMBER 2026
Pillar III — Measurement, Strategic Choices and the 2031 Outlook
The economics of artificial-intelligence productivity in government: outcome measurement, fiscal accountability, workforce allocation, institutional resilience, strategic scenarios and the decisions that will determine whether automated efficiency becomes public value.
Institutional scope: Italy, France, Germany, the United Kingdom and the European Union, with comparative implications for the United States, China and Russia. Strategic horizon: 2027–2031.
Chapter 7. Measuring the Conversion of Time Savings into Public Value
The measurement problem: the distinction between productive work, administrative throughput and public value
The most consequential unresolved question in public-sector artificial intelligence is not whether machines can accelerate tasks that previously required substantial human effort. It is whether public institutions can demonstrate that those accelerations produce improvements in the services and responsibilities for which they are funded. The central analytical judgment is that task-level time savings constitute an intermediate operational benefit, while public-sector productivity must ultimately be assessed through complete-service performance, the resources consumed to achieve it and the quality of the resulting administrative outcomes. A government cannot establish that its administration has become more productive merely by documenting that employees draft correspondence more quickly, retrieve information in fewer minutes or use AI assistants frequently. The relevant economic evidence must establish what happens to the capacity released by those improvements and whether its subsequent allocation contributes to legitimate institutional objectives.
This distinction becomes particularly important when assessing the cumulative effects of AI across large public organisations. Administrative services consist of multiple activities with different resource requirements, legal obligations and operational constraints. Some involve routine information processing, while others require professional interpretation, consultations, verification or formal authorisation. The acceleration of a preliminary task may increase the volume of material reaching a subsequent stage without changing the capacity of that stage to process it. Consequently, local gains in operational efficiency can coexist with unchanged waiting times or even increasing backlogs. The economic outcome depends on whether the accelerated activity was a binding constraint and whether management can redirect resources toward the parts of the process that determine final completion.
The international evidence provides an increasingly useful basis for distinguishing these phenomena. The United Kingdom’s Department for Work and Pensions reported an adjusted estimate of approximately 19 minutes saved per working day in its Microsoft 365 Copilot evaluation, but the study’s central measurements concerned task efficiency and employee experience, not an equivalent reduction in all service-delivery times. The evaluation recognised limitations arising from non-random licence allocation, self-reported measurements and the absence of a pre-trial baseline. These qualifications do not invalidate its findings; they define what the evidence can legitimately establish. An Evaluation of DWP’s Microsoft 365 Copilot Trial — Department for Work and Pensions, January 2026.
The OECD’s 2025 examination of 200 identified government AI use cases adds a broader institutional perspective. It found that 57% of the cases supported the automation, streamlining or tailoring of services and processes, while other applications addressed decision support, forecasting, accountability and anomaly detection. Importantly, these categories describe the objectives associated with selected use cases rather than measured national productivity improvements. The OECD also reported that only 15% of governments in its 2023 evidence had an AI investment framework, illustrating the gap between technological interest and formal arrangements for evaluating investments. Governing with Artificial Intelligence: The State of Play and Way Forward in Core Government Functions — OECD, June 2025.
A meaningful system of measurement must therefore distinguish the objectives of an AI programme from its actual operational consequences. Declared objectives may justify experimentation and the allocation of resources, but verified outcomes are required before institutions can claim realised productivity. This distinction is particularly relevant to public finance, where estimates of potential economic benefit can influence procurement, staffing and future expenditure even when the underlying technology has not been tested at operational scale.
The production function of public administration: why conventional productivity measures are insufficient
In conventional economic analysis, productivity concerns the relationship between outputs and the inputs required to produce them. Applying this principle to public administration is more complicated because public services frequently have outputs that cannot be evaluated through market prices, and the social value of their delivery depends on legality, accessibility, timeliness and substantive quality. Two administrative decisions may require similar working time while having very different levels of complexity and public significance. Likewise, processing a greater number of applications does not necessarily constitute better performance if the additional output is associated with a higher rate of errors or inadequate examination of individual circumstances.
This creates a measurement challenge in which neither labour hours nor completed transactions can be interpreted independently. Labour productivity measured as cases per staff-hour may increase when employees use AI to complete routine work, but the resulting indicator can become misleading if the complexity of the case mix changes. An administration dealing with an increasing proportion of complex applications may appear less productive even when its workforce has improved its technical performance. Conversely, an organisation may report a large increase in cases completed after introducing AI because it has concentrated on straightforward applications while more difficult cases remain unresolved.
The appropriate unit of measurement must therefore correspond to a clearly defined service or administrative function, with sufficient information about the characteristics of the cases being processed. Case-mix adjustment is particularly important for services involving legal interpretation, social benefits, regulatory investigations or other activities in which complexity varies substantially. In these settings, counting every completed case as an equivalent output can produce distorted comparisons among offices and over time.
The quality of administrative decisions must also be incorporated. A faster process that generates more corrections, appeals or reversals can impose additional costs on citizens and institutions. Such costs may appear after the initial transaction and in departments other than the one where AI was deployed. An assessment restricted to the time saved by the original employee would therefore omit part of the administrative burden generated by the system.
A broader productivity framework should distinguish at least four economic relationships: the quantity of valid services delivered per unit of input; the time required to complete those services; the quality and accessibility of the outcomes; and the full cost of maintaining the organisational and technological arrangements through which they are produced. These dimensions can improve independently and should not be collapsed into an arbitrary composite score.
The time-accounting framework: separating active processing, waiting and rework
A fundamental requirement for evaluating AI in government is to distinguish between active processing time and elapsed service time. Active processing time measures the period during which personnel or systems perform work on a case. Elapsed service time includes the entire interval between the defined start and completion of a procedure, including time spent waiting for another department, a legal approval, additional information or an external response. Artificial intelligence may reduce active processing time without materially changing elapsed time when waiting periods account for most of the procedure.
The distinction can be illustrated through an explicitly hypothetical example. Suppose an administrative process requires 45 minutes of initial document preparation, 30 minutes of substantive review, 15 minutes of final processing and 390 minutes of waiting between stages. Its total elapsed duration is 480 minutes. If AI reduces the initial preparation task by two-thirds, that activity falls from 45 to 15 minutes, producing a gross saving of 30 minutes. The complete process now takes 450 minutes, assuming all other stages and waiting periods remain unchanged. A 66.7% improvement in the selected task produces only a 6.25% reduction in total elapsed service time. The example is a calculation based on stated assumptions, not a measured result from any government.
The economic significance changes when the same technology is introduced at a binding constraint. If initial document preparation is the activity limiting the number of files that an office can process, reducing its duration may increase throughput substantially, provided that subsequent units have sufficient spare capacity. If another stage remains the bottleneck, throughput may change little even though the initial task becomes faster. The result depends on the capacity of the overall process, not merely on the proportional reduction in one activity.
A further complication concerns rework. AI-generated outputs may require verification, correction or repetition. If the initial activity becomes faster while the probability of an error increases, the net benefit depends on the frequency and cost of those errors. The appropriate record should therefore distinguish first-pass processing from subsequent corrections and identify the departments responsible for them. This is particularly important where an error discovered late in a procedure requires several earlier stages to be repeated.
For performance management, the administration should record the creation time of a case, the start and completion of each significant stage, transfers between units, periods awaiting additional information, rework events and the final administrative outcome. Without these data, it may be possible to measure individual employee time savings but impossible to determine whether they improved the complete service.
A quantitative model for evaluating the conversion of recovered time
A practical analytical framework can use a small number of transparent relationships, while recognising that their inputs must be measured rather than assumed.
Gross task time saved is the difference between the time required to perform a defined task before and after AI assistance, holding the relevant task requirements and quality standards sufficiently constant.
Net operational capacity recovered is gross time saved less additional work attributable to the new system, including validation, correction, monitoring and other necessary activities. This is a capacity measure rather than a financial saving.
Service-time improvement is the change in elapsed time required to complete the defined administrative service. It must be measured separately because a reduction in active work does not guarantee an equivalent reduction in waiting or approval time.
Net fiscal benefit is the expenditure actually avoided or other monetised benefit legitimately recognised under the relevant appraisal framework, less attributable implementation and operating costs. The treatment of non-cashable capacity gains must remain explicit.
Public value incorporates the consequences for the legally defined service, including quality, accessibility, fairness and the ability to meet demand. Where benefits cannot be credibly monetised, they should be reported through appropriate non-monetary indicators rather than forced into financial estimates.
The United Kingdom’s Digital and Data Benefits Framework, published on 7 April 2026, provides an official example of the importance of such distinctions. The framework identifies productivity, user experience, channel shift, data sharing and technological capability as different potential benefit streams, while explicitly warning that analysts must prevent double counting. It recommends sensitivity analysis concerning adoption rates, productivity improvements and cost reductions and requires the framework to be used alongside HM Treasury appraisal guidance rather than as a standalone business case. Digital and Data Benefits Framework — Government Digital Service and Department for Science, Innovation and Technology, April 2026.
This is a critical methodological development because a single AI deployment can affect several indicators simultaneously. If a government values employee time saved and then separately counts additional services produced through the same recovered hours, it may double count the benefit. If reduced administrative delays are monetised through benefits to service users, those same benefits must not be counted again under a different programme heading without a defensible reconciliation. A properly constructed business case should identify which benefits are distinct, which are alternative valuations of the same effect and which depend on assumptions about future organisational changes.
Administrative productivity scorecard: required indicators, definitions and attribution rules
| Measurement domain | Indicator | Required unit and observation | Analytical purpose | Main attribution risk |
|---|---|---|---|---|
| Task efficiency | Median active task duration | Minutes per comparable task | Establish direct acceleration | Differences in complexity |
| Recovered capacity | Net staff time released | Staff-hours after verification and rework | Measure usable operational capacity | Unrecorded new workloads |
| Service duration | Median end-to-end completion time | Hours or calendar days per completed case | Establish effects on final service | Changes in demand and queues |
| Service reliability | 90th-percentile completion time | Hours or days | Monitor delayed cases | Complex cases concentrated in upper tail |
| Throughput | Completed valid cases | Cases per period and staff-hour | Assess service capacity | Variation in case mix |
| Pending demand | Backlog size and ageing | Open cases and days pending | Identify unresolved congestion | Changes in incoming demand |
| Quality | Corrections and reversals | Cases and rates | Detect errors and rework | Changes in review intensity |
| Accessibility | Successful completion by user category | Rates within defined groups | Assess distributional outcomes | Changes in user composition |
| Fiscal performance | Fully loaded cost per completed service | Currency and price basis | Connect inputs to outputs | Omitted system and transition costs |
| AI operating costs | Licences, infrastructure and support | Annual expenditure | Establish continuing cost | Incomplete cost allocation |
| Workforce allocation | Recovered time assigned to authorised functions | Hours by activity | Determine where efficiency gains go | Informal or undocumented changes |
| Sustainability | Retained improvement after deployment | Comparable periods | Test durability | Concurrent organisational reforms |
The scorecard should be interpreted as a reporting structure rather than as a recommendation to maximise every indicator simultaneously. Some measures can legitimately move in different directions. More intensive human review may increase active processing time while reducing errors and appeals. Greater accessibility may increase the number of complex applications successfully submitted, thereby raising demand and affecting the average time required to complete a case. A government that evaluates only speed or volume may misinterpret these changes as declining performance when they reflect improvements in the reach or quality of public services.
The appropriate measurement framework should therefore be designed around the statutory purpose and operating characteristics of the relevant institution. For a routine administrative certificate, timeliness and cost may be prominent. For a complex regulatory investigation, accuracy, legal sustainability and the quality of evidence may carry greater importance. For a citizen-facing support service, accessibility and successful resolution may be central. Standardisation should concern definitions and reporting discipline, not the imposition of identical performance objectives on fundamentally different public functions.
Establishing credible causality: trials, comparison groups and administrative data
The strongest evaluations of AI deployment do more than compare performance before and after introducing a system. They examine whether observed changes can reasonably be attributed to AI rather than to other developments occurring during the same period. This is particularly important in government, where staffing levels, service demand, legislation, budget allocations and digital systems may change simultaneously.
Randomised trials can be useful when employees, offices or cases can be assigned to treatment and comparison conditions in a legally and operationally appropriate manner. Such designs may reduce selection bias, but they require sufficient sample sizes, clearly defined outcomes and safeguards against differences in treatment intensity or the exchange of practices between groups. Random assignment may be inappropriate for certain high-consequence services or where withholding an established improvement would create unacceptable operational concerns.
Where randomisation is unavailable, quasi-experimental approaches can provide useful evidence if their assumptions are credible. Difference-in-differences analysis can compare changes in treated and untreated groups, but it requires a defensible basis for assuming comparable trends in the absence of intervention. Interrupted time-series designs can examine changes around deployment dates, but they must account for seasonality, pre-existing trends and other contemporaneous reforms. Regression-based comparisons can adjust for observable differences among employees or offices, but they cannot eliminate all confounding from unobserved characteristics.
The DWP evaluation illustrates the value and limits of an adjusted comparison-group design. Its analysis controlled for demographic, occupational and AI-related interest variables while explicitly acknowledging non-random licence allocation and the absence of a pre-trial baseline. The fully adjusted estimate of 19 minutes daily is more informative when interpreted alongside those methodological qualifications than when presented as an unconditional productivity coefficient applicable to all civil servants. DWP Copilot Evaluation, Methodology and Chapter 4 — Department for Work and Pensions, January 2026.
For service-level evaluations, administrative transaction records are especially valuable because they can capture actual completion dates, queue durations, corrections and outcomes rather than relying exclusively on employee perceptions. However, such records must be sufficiently complete, consistent and legally usable. Data quality becomes a prerequisite for credible evaluation, particularly where offices record processing stages differently or where legacy systems do not provide reliable timestamps.
The public-sector AI benefit conversion model
Why a Faster Task Does Not Guarantee a Faster Service
Illustrative sequential administrative process. All values are hypothetical minutes, with no change in waiting or review requirements.
Chapter 7 — Key judgments
The conversion of AI time savings into public value requires at least three distinct forms of evidence: reliable measurements of task-level effects, documented changes in complete administrative processes and outcome indicators that establish whether citizens, institutions or public budgets benefit. These measurements must account for quality, complexity, waiting time, rework and the costs of maintaining the technology. No single estimate of minutes saved can substitute for this framework.
The OECD’s international evidence and the United Kingdom’s official evaluation guidance demonstrate the importance of assessing multiple benefit streams without double counting. Administrative transaction records and appropriately designed comparisons provide a stronger basis for institutional productivity claims than technology inventories or employee satisfaction alone.
The assessment would materially change if governments published reproducible evaluations that linked AI deployment to end-to-end service outcomes while separately identifying the resources required for implementation, the treatment of recovered capacity and the consequences for decision quality.
Open official record: administrative transaction datasets, before-and-after service evaluations, full-cost AI implementation records, documented workforce reallocation and independently interpretable assessments of service quality.
Chapter 8. Fiscal Implications, Workforce Reallocation and Institutional Risks
The public-finance problem: why recovered working time is not equivalent to expenditure that governments can remove from their budgets
The fiscal implications of AI adoption require a more demanding analysis than the measurement of employee time savings. When an AI system reduces the time required for a public employee to complete a task, the institution obtains additional potential working capacity, but the corresponding salary expenditure ordinarily remains in the budget unless a separate decision changes staffing, overtime, contracting or other costs. The benefit may be economically significant without generating immediate cash savings. The central finding is that governments must distinguish three separate results: operational capacity released, expenditure genuinely avoided and public value created through better or additional services. Treating these outcomes as interchangeable risks overstating the budgetary return on AI investments and can create unrealistic expectations concerning workforce reductions or fiscal consolidation.
The distinction is particularly important because public expenditure frequently finances resources that cannot be reduced continuously in proportion to minor changes in task duration. Salaries, premises, infrastructure, statutory obligations and many support services contain fixed or relatively inflexible components. If an AI assistant saves an employee 20 minutes per day, the institution does not automatically reduce its payroll by the monetary value of those minutes. It may instead use the recovered capacity to process additional applications, improve professional supervision, reduce queues or meet demand that would otherwise require more staff. Those outcomes can represent genuine value for money, but they require measurement frameworks different from the accounting of expenditure reductions.
The United Kingdom’s Digital and Data Benefits Framework, issued on 7 April 2026, provides particularly relevant official evidence because it differentiates the value of staff productivity from other benefit categories and explicitly warns against double counting. Its discussion of public-sector AI identifies an earlier Government Digital Service estimate of £23.7 billion in potential benefits across healthcare, education and policing. The underlying components are £13.6 billion for healthcare under an assumption of automating 25% of tasks, £7.5 billion for education under an assumption of automating approximately one-third of tasks and £2.6 billion for policing under an assumption of automating 29% of tasks. These figures are prospective estimates associated with specified assumptions; they are not audited reductions in public expenditure or evidence that the corresponding improvements had already occurred. Digital and Data Benefits Framework, Section 2: AI — UK Government, April 2026.
The decomposition of this estimate illustrates the scale of potential economic effects while simultaneously demonstrating the risk of misinterpretation. Healthcare, education and policing have different operating models, professional responsibilities and service objectives. Time recovered in clinical administration, classroom preparation or police information processing cannot necessarily be redistributed or monetised in the same manner. In healthcare, administrative assistance may free professional attention for patient care, but the resulting benefit depends on clinical capacity, scheduling and the organisation of services. In education, reduced preparation workloads may contribute to teaching, assessment or professional development rather than lower staffing expenditure. In policing, efficiency gains may increase investigative or frontline capacity while leaving overall expenditure unchanged. The financial value of these outcomes is therefore not established by multiplying a task-automation assumption by aggregate labour costs.
The fiscal question becomes more complex when AI is introduced alongside wider digital transformation. Investments in structured data, cloud services, modern case-management systems and cybersecurity may be necessary for effective AI deployment, but those assets can also generate benefits independently of AI. Attribution requires distinguishing the contribution of the model from the contribution of supporting infrastructure and organisational reform. Without this distinction, an institution may assign too much of the benefit to AI or count the same digital improvement under several different investment programmes.
The October 2026 Green Book and the discipline of public investment appraisal
An important development for the economic assessment of government AI occurred with the publication of the 2026 edition of HM Treasury’s Green Book, issued on 28 October 2026. The Green Book establishes the UK government’s framework for appraising policy, projects and programmes through the consideration of objectives, monetisable and non-monetisable benefits, public-sector financial effects, risk, uncertainty and distributional consequences. It defines value for money through a balanced judgement concerning the use of public resources to achieve objectives, rather than through a narrow test of immediate cash savings. The Green Book 2026 — HM Treasury and Government Finance Function, October 2026.
This framework is directly relevant to AI because its expected benefits often involve uncertainty concerning adoption, technical performance, employee behaviour and the ability of institutions to change their workflows. The Green Book requires consideration of a business-as-usual counterfactual, the rationale for intervention, feasible alternatives and the costs and benefits associated with those alternatives. It also requires appraisal to address optimism bias, the tendency for projected benefits to be overstated while costs and implementation periods are underestimated. In AI programmes, optimism bias may arise when early demonstrations are assumed to represent ordinary operating conditions, when adoption rates are treated as equivalent to effective usage or when the costs of integrating systems with existing administrative infrastructure are underestimated.
The 2026 framework also provides a basis for distinguishing social cost-benefit analysis from social cost-effectiveness analysis. Where important benefits can be valued credibly in monetary terms, a cost-benefit approach can assist comparison among options. Where the benefits are materially important but cannot be monetised reliably, cost-effectiveness analysis and explicit treatment of non-monetary outcomes may be more suitable. This distinction is especially significant for AI systems used in services involving fundamental rights, administrative justice or the quality of public decisions. The value of a lawful and accurate decision cannot always be reduced to the number of minutes required to produce it.
The Green Book also distinguishes the public-sector financial effect of an intervention from its wider social value. A project can improve outcomes for households or businesses without producing a corresponding reduction in government expenditure, and a programme that reduces the cost of an internal activity may have limited social value if it does not improve the relevant public function. This distinction provides a conceptual basis for evaluating AI investments without converting all benefits into a theoretical wage-based estimate.
For European administrations outside the United Kingdom, the Green Book is not a governing legal instrument. Its analytical distinctions are nevertheless useful for comparative assessment, provided national public-finance rules and appraisal requirements remain authoritative. Italy’s budgetary system, Germany’s federal arrangements and France’s public-finance procedures must each be evaluated within their own institutional frameworks. The principle that realised expenditure, potential capacity and wider public value should be distinguished does not require identical fiscal legislation across jurisdictions.
The three fiscal accounts of AI deployment
A credible governmental AI business case should maintain three connected but separate accounts.
The first account concerns operational capacity. It records the staff-hours potentially released from activities, adjusted for verification, correction and the additional work required to operate the AI system. This account answers whether the organisation can perform its duties with less active effort or whether existing resources can support more activity. It does not, by itself, establish savings in the government’s cash budget.
The second account concerns public expenditure. It records the incremental financial costs of deploying and maintaining AI, alongside expenditures that can genuinely be avoided. Costs can include licences, cloud or local computing, technical integration, cybersecurity, specialist staff, training, evaluation, procurement and eventual decommissioning. Avoided expenditure may include reduced overtime, external contracting, duplicated software or planned expenditure that would otherwise have been necessary. Claims concerning avoided future expenditure require a credible counterfactual and should remain separate from reductions actually recorded in accounts.
The third account concerns public-service and social outcomes. It examines changes in waiting time, throughput, accuracy, accessibility, quality and the consequences for citizens or enterprises. These changes may create economic value even when staff expenditure remains unchanged. They should be monetised only where the methodology and evidence permit, and any non-monetised effects should remain visible in the decision.
The three-account structure reduces the risk of double counting. For example, the same hour released by an employee cannot simultaneously be valued in full as a cash saving and as capacity used to deliver an additional service unless the financial and operational relationships justify distinct benefits. The government must establish whether the hour has actually reduced an avoidable cost, has been reassigned to another activity or remains available without an observed additional output.
Fiscal classification of government AI benefits
| Benefit or cost category | Appropriate measurement | Public-budget consequence | Main evidentiary requirement | Common accounting error |
|---|---|---|---|---|
| Gross task time saved | Minutes or hours | None automatically | Comparable task measurements | Counting time as cash |
| Net capacity recovered | Staff-hours after additional work | May enable redeployment | Validation and rework data | Ignoring additional tasks |
| Overtime expenditure avoided | Recorded overtime payments | Potential cash saving | Payroll and workload records | Assuming planned reductions were realised |
| External contracts reduced | Actual contract expenditure | Potential cash saving | Contract changes and payments | Ignoring transition costs |
| Additional valid cases completed | Quality-adjusted service output | May remain budget-neutral | Transaction records | Monetising the same labour twice |
| Shorter waiting times | Elapsed time for applicants | Social benefit; possible indirect fiscal effect | Service-time and user data | Assuming all waiting time is monetisable |
| Reduced errors and appeals | Corrections, appeals and related costs | Potential savings across several units | Case records and cost attribution | Counting gross reductions without causality |
| AI licences and computing | Contractual and metered expenditure | Direct cost | Financial and supplier records | Excluding recurring charges |
| Integration and data preparation | Staff and external implementation costs | Direct or allocated cost | Project and personnel accounts | Treating existing resources as free |
| Cybersecurity and oversight | Operating and assurance expenditure | Continuing cost | Security and staffing records | Excluding necessary safeguards |
| Workforce training | Training and opportunity costs | Implementation cost; possible future benefit | Training and performance records | Claiming competence from attendance alone |
| Supplier exit and decommissioning | Migration, termination and continuity costs | Future or contingent cost | Contract and technical assessment | Ignoring switching dependencies |
The table represents an analytical accounting framework. Costs and benefits should be incorporated only where evidence supports their existence and amount, with appropriate treatment of real prices, the appraisal period, discounting and distributional effects under the applicable national framework.
Hypothetical fiscal example: how an apparently large productivity dividend can disappear
Consider a deliberately simplified administrative organisation employing 1,000 workers who perform tasks potentially supported by AI. Assume that 60% of the employees use an authorised assistant regularly, that the average gross time saving among those users is 20 minutes per working day and that the organisation operates on an assumed schedule of 220 working days annually. Under those assumptions, gross annual recovered working time equals 44,000 staff-hours. This number describes potential capacity and has not been derived from an actual government deployment.
If 30% of those hours are subsequently required for verification, correction, training and additional operational work associated with the system, net recovered capacity falls to 30,800 staff-hours. The calculation does not establish whether the organisation can use those hours productively. If the relevant employees cannot be reassigned to tasks where demand exists, or if downstream approval capacity remains constrained, some of that capacity may not translate into additional completed services.
Suppose that the administration subsequently documents 5,000 hours of genuinely avoidable overtime, valued at an assumed marginal cash cost of €35 per hour. The resulting gross cash saving would be €175,000. If the AI programme incurs annual incremental operating and support expenditure of €240,000, the difference between those two financial categories would be negative €65,000, before considering any other attributable costs or benefits. The hypothetical programme could nevertheless deliver worthwhile public value if its remaining recovered capacity produced faster services, fewer errors or additional outputs whose benefits justified the expenditure.
The example illustrates why the same programme can simultaneously display a large gross productivity gain, a substantial amount of usable recovered capacity and no immediate net cash saving. It also demonstrates why claims that AI will automatically finance itself through employee time savings require careful scrutiny.
The three accounts of AI productivity
Recovered Time Is Not Treasury Cash
Worked example with explicitly assumed inputs. No values represent measured results from a government.
| Account | Assumption | Calculated result | Interpretation |
|---|---|---|---|
| Staff population | 1,000 employees | 600 regular users | 60% assumed uptake |
| Gross time | 20 minutes per day; 220 days | 44,000 hours | Potential labour capacity |
| Additional work | 30% of gross hours | 13,200 hours | Validation and support |
| Net time | 44,000 less 13,200 | 30,800 hours | Not a monetary saving |
| Overtime avoided | 5,000 hours × €35 | €175,000 | Assumed cash benefit |
| Annual AI operation | Assumed expenditure | €240,000 | Incremental cash cost |
| Financial difference | €175,000 − €240,000 | −€65,000 | Excludes other benefits and costs |
ILLUSTRATION ONLY. All inputs are assumptions. The negative cash difference does not demonstrate that the project lacks social value. Other benefits would require separate evidence and valuation.
Workforce reallocation: the economic consequences of choosing capacity expansion, professional development or fiscal adjustment
The allocation of recovered working time is one of the most consequential decisions facing government institutions because different uses of the same capacity can produce different forms of value. An administration facing substantial demand and long waiting times may allocate recovered hours to additional case processing. Another may prioritise professional verification, accessibility improvements or more complex services. A third may reduce avoidable overtime or external contracting. The appropriate decision depends on the legal mandate and operational circumstances of the service rather than on an assumption that all AI savings should produce the same outcome.
In Italy, the issue intersects with the organisational responsibilities established under public-employment law and with the continuing implementation of the national AI framework. The adoption guidelines issued by AgID on 9 November 2026 emphasise organisational impact, maturity assessment, data governance and professional competencies. Their publication was also formally announced in the Gazzetta Ufficiale on 16 November 2026. These developments establish an institutional basis for examining AI adoption within administrative management, but they do not determine a universal allocation of recovered staff time. Adoption of the Guidelines for AI in Public Administration — Gazzetta Ufficiale, November 2026 and Guidelines for the Adoption of AI in Public Administration — AgID, November 2026.
The decision to redeploy capacity must also consider the difference between employees’ technical productivity and the requirements of their professional role. A public employee may be able to prepare routine documentation more quickly, but the resulting capacity may not be immediately transferable to another function requiring different expertise. Training can improve this mobility, but it involves costs and time. Similarly, the ability to reduce staffing expenditure depends on the institution’s actual employment arrangements and future resource requirements. An administrative director cannot infer that an employee’s reduced drafting time permits the elimination of a position when the person also performs other essential functions.
The experience of the US Internal Revenue Service illustrates a related risk. The GAO’s March 2026 audit identified 126 AI use cases in the agency’s June 2025 inventory, including 77 in development, while reporting deficiencies in workforce planning and losses of employees possessing AI-related expertise. Officials in one relevant research and analytics group reported the loss of 63 employees who had been working full- or part-time on AI. These findings show how reductions in technical capacity can threaten the execution of future AI programmes even when those programmes are intended to improve efficiency. Artificial Intelligence: IRS Actions Needed to Address Skills Gaps, Information Quality, and Strategic Management — GAO, March 2026.
The implication is that a credible workforce strategy must account for two distinct effects: the substitution of AI for particular activities and the additional human capabilities required to maintain reliable AI-supported services. Reducing routine work may create opportunities for more demanding professional tasks, but expanding deployment may simultaneously increase demand for data specialists, security professionals, system engineers, evaluators and managers capable of redesigning processes. The net employment effect depends on the balance between these forces and cannot be inferred from the share of tasks considered technically automatable.
The institutional risks of reallocating capacity prematurely
A significant strategic risk arises when organisations treat expected productivity improvements as certain before completing the necessary operational evaluation. If an administration reduces staffing or other resources on the basis of projected AI savings, it may discover that the remaining workforce cannot perform required validation, maintain systems or address exceptional cases. The apparent fiscal benefit can then be offset by delays, expensive external support, rising error rates or the need to restore capacity.
This is particularly important for services where demand is volatile or the consequences of failure are significant. An AI application may perform well on routine cases but encounter difficulties with unusual documentation, complex legal questions or incomplete records. If workforce capacity has already been reduced, the administration may lack sufficient expertise to resolve those exceptions. The risk is therefore not simply technical model failure but an institutional mismatch between the capacity assumed in a business case and the capacity required under real operating conditions.
The opposite risk also exists. An administration may accumulate AI subscriptions and technical infrastructure without changing the organisation of work, resulting in recurring expenditure with limited measurable benefit. In this case, the institution has preserved its workforce but failed to convert technological capability into improved outcomes. Avoiding both problems requires staged investment and evaluation, with explicit conditions for expansion, modification or termination.
Fiscal risk also arises from supplier dependence. Changes in licence pricing, computing demand, data-processing requirements or integration costs can alter the economics of an AI application after its initial deployment. Contracts should therefore be evaluated over a realistic lifecycle, including the costs of switching providers and maintaining service continuity. The relevant question is not whether the initial procurement appears inexpensive, but whether the institution can sustain the required capability at an acceptable cost under plausible future conditions.
The 2026 Green Book provides a useful appraisal principle: public projects should be evaluated through their full social costs and benefits, public-sector financial impacts, distributional effects and risks, including uncertainty concerning the assumptions supporting the proposal. The framework explicitly requires sensitivity analysis and the identification of assumptions whose deterioration could change the value-for-money conclusion. The Green Book 2026 — HM Treasury, October 2026.
Institutional investment options: authority, expected effects and downside exposure
| Decision option | Required authority | Intended effect | Implementation burden | Reversibility | Principal risk |
|---|---|---|---|---|---|
| Expand general-purpose assistants | Procurement and authorised administrative management | Reduce selected routine workloads | Moderate; recurring expenditure | Relatively high if contracts permit exit | Low effective use and limited service outcomes |
| Invest in workflow integration | Service ownership, ICT and relevant administrative leadership | Reduce handoffs and operational delays | Substantial technical and organisational work | Moderate | Integration cost and disruption |
| Establish specialist AI evaluation teams | Competent management and workforce authorities | Improve verification, quality and measurement | Continuing staffing and training | Moderate | Additional bureaucracy without useful decisions |
| Redirect capacity toward backlogs | Service management within applicable employment arrangements | Increase valid case completion | Depends on workforce flexibility | Generally moderate to high | New bottlenecks |
| Reduce avoidable external expenditure | Budget and contract authorities | Create measurable cash savings | Contractual and operational adjustments | Variable | Loss of specialist support |
| Consolidate technology platforms | Government or agency technology governance | Reduce duplication and coordinate security | Potentially substantial | Moderate to low after deep integration | Concentrated dependencies |
| Maintain controlled pilots for uncertain applications | Authorised programme leadership | Gather stronger evidence before scaling | Limited to moderate | High | Delayed benefits and pilot proliferation |
| Retire underperforming applications | Competent service and procurement authorities | Stop recurring costs or reduce risk | Exit planning and migration | Depends on data and system portability | Disruption if operational dependencies are undocumented |
These courses of action are analytical options rather than measures already adopted by the governments examined. Their suitability must be evaluated against the relevant institution’s legal duties, financial resources, workforce characteristics and service objectives.
Chapter 8 — Key judgments
The economic value of AI-generated time savings cannot be equated with immediate fiscal savings. Public institutions may obtain significant operational capacity without reducing expenditure, and such capacity can create social value through faster, more accessible or higher-quality services. Realised cash benefits require evidence of genuinely avoided costs and must be assessed after accounting for the financial resources required to deploy and operate AI systems.
The UK government’s April 2026 benefits framework and October 2026 Green Book provide official methodological foundations for distinguishing potential benefits, realised financial impacts, uncertainty and the risk of double counting. The GAO’s examination of the IRS demonstrates why workforce capabilities must remain part of any fiscal assessment: savings obtained by reducing specialised personnel may compromise the institution’s ability to support AI deployment.
The assessment would materially change with audited records connecting AI investment to realised expenditure changes, quality-adjusted service output, workforce redeployment and costs incurred throughout the operating lifecycle.
Open official record: independently interpretable AI business cases, audited outturn expenditure, documented avoidance of contracts or overtime, workforce allocation records, total-cost-of-ownership assessments and evidence of the social benefits generated by additional administrative capacity.
Chapter 9. Strategic Scenarios, Early-Warning Indicators and the Final Net Assessment
The 2031 question: whether governments will govern AI as a technology programme or as a transformation of public institutions
The decisive strategic issue for the period 2027–2031 is whether governments will succeed in moving beyond the deployment of individual AI applications toward a system of administrative governance in which technological investment, workforce capacity, operational performance and public accountability are evaluated together. The distinction is not primarily about the sophistication of future AI models. More capable systems may accelerate a wider range of administrative activities and support increasingly complex workflows, but their institutional value will still depend on the conditions under which they operate and the decisions made about the capacity they release. The principal judgment is that the most consequential uncertainty concerns the ability of governments to redesign administrative processes and measure complete-service outcomes, rather than their ability to acquire additional AI tools.
The evidence already available establishes that governments can obtain measurable task-level benefits from AI assistance, but it also identifies weaknesses in the structures needed to reproduce those benefits at scale. The United Kingdom has produced detailed evaluations of employee time savings and official guidance on the appraisal of digital benefits. Italy has formalised adoption guidance that recognises organisational maturity and continuous improvement. Germany’s administrative initiatives demonstrate the potential relevance of AI to complex operational procedures. France has examined employee adoption and task performance through a large government trial. The United States has established federal governance and procurement requirements while its audit institutions have identified weaknesses in agency workforce planning and AI inventories. China has issued detailed government deployment guidance, while Russia has adopted digital-administration objectives extending to 2030. These developments create a substantial international policy agenda, but they do not establish a common rate at which AI adoption produces productivity growth.
The central uncertainty can be divided into three questions. The first is whether improvements in technical capability will continue to reduce the cost and effort required to perform relevant administrative activities. The second is whether public institutions will possess sufficient managerial, professional and technical capacity to incorporate those improvements into complete services. The third is whether the resulting benefits will be documented, distributed and controlled through arrangements consistent with the applicable legal and public-finance frameworks. The answers are interdependent, but they need not develop at the same pace. Technical capabilities can improve rapidly while administrative organisations remain constrained by fragmented data, limited staffing flexibility or inadequate performance measurement.
The next five years should consequently be assessed through institutional scenarios rather than unsupported numerical forecasts of productivity growth. The available official record does not provide a sufficiently consistent international historical series to support credible probabilities for these outcomes. Assigning precise likelihoods would create an appearance of predictive certainty without an adequate empirical basis.
Scenario A — Integrated administrative transformation
Under the first scenario, governments connect AI investment with systematic redesign of administrative processes, reliable performance measurement and a clearly defined distribution of responsibilities. Individual task improvements are evaluated within the complete procedures they support, while managers receive sufficient information to determine whether recovered capacity can reduce waiting times, expand service provision, improve quality or produce avoidable expenditure. Training programmes are linked to operational roles, and technical procurement is designed around interoperability, security and the ability to maintain or replace systems.
The mechanism supporting this development is organisational complementarity. AI reduces the effort required for selected activities, while changes in workflow, information management and personnel allocation permit institutions to use the resulting capacity. In such a setting, the benefit of AI does not depend exclusively on the magnitude of the initial task saving. A relatively modest improvement at a binding bottleneck may generate substantial service-level value, while a larger improvement in a non-constrained activity may have limited effects on final completion times. Institutions that can identify and act on these differences will be better equipped to select appropriate applications and avoid unnecessary expenditure.
Evidence consistent with this scenario would include sustained reductions in end-to-end completion times, improvements in quality-adjusted throughput, reduced backlog ageing and detailed records showing how recovered staff capacity was reassigned. Financial reporting would distinguish genuine expenditure reductions from non-cashable benefits, while post-implementation evaluations would examine whether the gains persisted after changes in technology, demand and workforce composition.
An important institutional condition is the availability of reliable administrative data. Governments cannot govern productivity effectively if they do not know how long procedures take, where files accumulate or how often work must be repeated. The quality of transaction records and the ability to integrate information across departments may therefore be as important to the realisation of AI benefits as model performance itself.
The principal risk within this scenario is excessive complexity in the evaluation and governance system. A government could respond to the need for accountability by establishing procedures so burdensome that they absorb a material proportion of recovered capacity. Effective implementation would therefore require proportionate controls, reusable evaluation methods and clear responsibility for decisions, rather than multiplying reporting obligations for every low-risk activity.
Scenario B — Technological expansion without institutional productivity
The second scenario is characterised by continued growth in AI deployments while administrative performance changes only modestly or inconsistently. Government institutions purchase additional licences, introduce new assistants and expand their inventories of AI use cases, but the organisation of work remains substantially unchanged. Employees report faster completion of selected activities, while unresolved queues, fragmented information systems and downstream approval requirements limit the effects on complete services.
This scenario is plausible because technical adoption and organisational reform are governed by different decision processes. Procurement can often proceed through identifiable budgets and contracts, while workflow redesign may require cooperation among departments, changes to established procedures and agreement concerning the redistribution of responsibilities. Technology projects may therefore be easier to initiate than institutional reforms, particularly when the benefits of the latter are distributed across several organisations.
The resulting risk is that AI spending becomes a recurring component of administrative expenditure without a corresponding increase in public value. Some employees may benefit from reduced routine work, but those gains remain local and difficult to reproduce. Departments may maintain overlapping tools, while the absence of common evaluation methods prevents managers from determining which applications are worth expanding. In this environment, reported adoption can continue increasing even when service-level indicators remain largely unchanged.
The principal signposts would include rising AI expenditure, increasing numbers of licences or applications, limited improvement in completed-service performance and persistent difficulties reconciling the reported benefits with administrative transaction records. A further indicator would be the absence of documented decisions concerning the use of recovered staff time. If an institution can demonstrate that employees work faster but cannot explain how the saved capacity contributes to its mission, the productivity dividend remains institutionally unresolved.
This scenario is particularly relevant to the interpretation of government AI inventories. The GAO’s July 2025 examination of selected federal agencies established rapid growth in reported use cases, but the inventory included applications at different stages of development and did not provide a corresponding measure of realised productivity. This demonstrates why deployment counts must remain separate from performance outcomes. Artificial Intelligence: Generative AI Use and Management at Federal Agencies — Government Accountability Office, July 2025.
Scenario C — Automation dependence and institutional resilience pressure
The third scenario arises when governments obtain meaningful automation benefits but become increasingly dependent on technical systems, suppliers or specialised personnel whose failure or unavailability could disrupt essential administrative functions. The benefits may initially appear substantial because AI applications become embedded in document management, case processing, information retrieval and operational coordination. However, the concentration of capabilities in a limited number of platforms or teams can create new dependencies that are not fully reflected in initial productivity measurements.
The economic mechanism is a change in the structure of operating risk. When an institution performs a task manually, its continuity depends on the availability and competence of employees and supporting administrative systems. When that task becomes heavily dependent on AI, continuity may also require reliable model access, computing capacity, specialised data resources, integration services and suitable fallback arrangements. The loss of a critical component may affect several downstream functions simultaneously.
A shared AI platform can reduce duplicated infrastructure, but the consequences of interruption may become more concentrated. Similarly, an institution may reduce its dependence on routine manual work while increasing its dependence on a smaller group of specialists capable of maintaining and supervising automated systems. These trade-offs cannot be evaluated through average task savings alone.
The official evidence identifies relevant mechanisms without establishing that this scenario will inevitably occur. The GAO’s IRS audit demonstrated that staffing reductions and inadequate workforce planning can weaken the ability of an agency to support AI applications. OMB’s federal procurement guidance recognises the importance of interoperability, competition and data portability. Chinese government AI guidance emphasises shared infrastructure while prescribing safeguards and lifecycle management. These records show that resilience and control are material considerations across different governance systems. GAO-26-107522 — Government Accountability Office, March 2026, M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government — Office of Management and Budget, April 2025, and Government Large-Model Deployment Guidelines — Cyberspace Administration of China, October 2025.
Indicators consistent with this scenario would include increasing concentration of critical workloads, the absence of tested fallback procedures, difficulties migrating applications between suppliers and shortages of personnel able to maintain or independently evaluate the systems. The appropriate institutional response would depend on the significance of the affected services and the feasibility of reducing dependencies through procurement, technical architecture and workforce development.
Scenario comparison: diagnostic evidence and implications
| Scenario | Mechanism | Evidence that would support it | Evidence that would weaken it | Principal consequence |
|---|---|---|---|---|
| Integrated administrative transformation | AI adoption combined with process redesign and accountable capacity allocation | Sustained service-time reductions, higher valid throughput, reliable net-benefit reporting | Improvements confined to individual tasks | Greater demonstrated public value |
| Technological expansion without institutional productivity | Tools and licences expand faster than organisational reform | Growing expenditure with unchanged service outcomes | Documented end-to-end gains and avoided expenditure | Weak conversion of investment into outcomes |
| Automation dependence and resilience pressure | Critical administrative functions increasingly depend on concentrated technology and expertise | Supplier concentration, migration barriers, outages, specialist shortages | Tested continuity arrangements and diversified operational capability | Greater exposure to disruptions and switching costs |
These scenarios can overlap. An administration may achieve substantial productivity gains in some services while retaining fragmented systems elsewhere. It may also become more productive and more dependent on a critical technology supplier at the same time. The appropriate evaluation therefore concerns the conditions under which each mechanism operates, not the assignment of an entire country to a single category.
The early-warning architecture: observing institutional changes before they become systemic problems
A useful strategic monitoring system should identify observable developments capable of confirming or weakening the central assessment. These indicators must be connected to decisions that institutions can actually make. Monitoring is not valuable merely because information can be collected; it becomes valuable when a change in the evidence can justify expanding a deployment, redesigning a procedure, correcting a security weakness or revising a financial assumption.
The first category concerns service performance. Administrations should monitor complete-case duration, throughput, backlog size and ageing, and the frequency of corrections or reversals. These indicators reveal whether local automation is affecting the institution’s final outputs. The second category concerns financial sustainability, including recurring licence charges, infrastructure expenditure, support costs and verified avoided expenditure. The third concerns organisational capacity, including the availability of skilled personnel, training effectiveness, system maintenance and the distribution of recovered working time. The fourth concerns risk and resilience, including operational incidents, unauthorised AI use, supplier concentration and the ability to maintain essential services during disruptions.
A further category concerns changes in the regulatory environment. Within the European Union, the consolidated AI Act establishes future application dates for specified high-risk requirements, including 2 December 2027 for the relevant Annex III categories and 2 August 2028 for specified Annex I product-related systems, following the 2026 amendments. These dates provide concrete planning milestones for affected providers and deployers. Institutions must still determine whether particular systems fall within the relevant categories and which obligations apply. AI Omnibus Enters into Force — European Commission, July 2026 and Consolidated Regulation (EU) 2024/1689 — EUR-Lex, July 2026.
For Italy, the publication of AgID’s AI adoption guidelines in November 2026 establishes a near-term implementation reference point. The agency separately identified its guidance on AI development and procurement as undergoing adoption procedures at the time of its published overview. Their eventual formal adoption and subsequent implementation should therefore be monitored as separate institutional developments rather than assumed to have occurred already. Artificial Intelligence: Guidelines and Implementation Status — AgID, November 2026.
Strategic watch indicators and decision triggers
| Indicator | Record or observation required | Signal of progress | Warning condition | Relevant decision |
|---|---|---|---|---|
| Median service duration | Administrative transaction records | Sustained reduction adjusted for case mix | Task savings without service improvement | Redesign downstream workflow |
| Upper-tail completion time | 90th- or 95th-percentile service records | Fewer severely delayed cases | Growing long-duration backlog | Review complex-case capacity |
| Net recovered hours | Time studies and activity records | Verified capacity available for reassignment | Gross savings offset by rework | Modify application or review procedures |
| Quality-adjusted throughput | Completed cases and error records | More valid outputs with acceptable quality | Higher volume accompanied by more corrections | Strengthen validation or narrow deployment |
| Cost per completed service | Financial and operational records | Sustainable unit-cost improvement | Rising costs without corresponding outcomes | Review procurement and operating model |
| Workforce capability | Skills assessments and staffing records | Required competencies maintained | Critical specialist shortages | Revise training and recruitment arrangements |
| Supplier concentration | Contracts and technical dependency inventories | Manageable portability and continuity | Essential functions depend on a single irreplaceable component | Test exit and contingency options |
| AI inventory completeness | Reconciled deployment and procurement records | Accurate records of applications and ownership | Material omissions or unknown owners | Strengthen governance and asset control |
| Security and reliability | Incident logs and adversarial testing | Effective controls and manageable incidents | Repeated failures or unacceptable exposure | Suspend, redesign or restrict affected use |
| Benefit attribution | Evaluation and counterfactual evidence | Credible causal contribution to outcomes | Benefits based solely on assumptions | Delay claims or further scaling |
| Workforce allocation | Management decisions and activity records | Documented legitimate use of recovered capacity | No identifiable disposition of time savings | Establish allocation and reporting responsibilities |
| Regulatory readiness | Applicable legal obligations and implementation records | Compliance arrangements aligned with operative dates | Unresolved duties close to application milestones | Prioritise remediation |
The warning conditions are qualitative because universally applicable numerical thresholds cannot be established from the available evidence. An acceptable error rate, service delay or supplier-concentration level will depend on the legal and operational significance of the relevant activity. Individual administrations should establish quantified thresholds only where baseline records, service obligations and risk assessments provide a defensible basis for doing so.
The 2027–2031 decision sequence
The five-year horizon creates a useful planning structure, but the sequence should not be interpreted as a forecast that governments will complete particular reforms by specific dates. It is a proposed progression of institutional decisions, organised around the dependencies required to transform AI adoption into measurable outcomes.
During 2027, administrations undertaking substantial AI investment should establish the operational records necessary to evaluate their existing deployments, including application ownership, baseline service measurements, full-cost information and relevant workforce competencies. Within the European Union, the December 2027 milestone for specified Annex III high-risk systems adds a concrete regulatory consideration. The principal strategic decision is whether institutions possess adequate evidence and controls to justify wider deployment.
During 2028, attention should extend toward the integration of AI into complete administrative workflows and the evaluation of results beyond initial trials. The August 2028 European milestone for specified product-related high-risk systems is relevant to affected organisations, while the broader economic challenge concerns whether applications have produced repeatable improvements without excessive additional costs or new dependencies. Expansion decisions should be linked to observed outcomes rather than the availability of increasingly capable models alone.
During 2029, governments should be able to examine whether earlier investments have produced durable effects across several operating cycles. By this stage, the most useful evidence would concern changes in cost per completed service, backlog reduction, professional capabilities, system resilience and the distribution of benefits among different user groups. Where improvements have not appeared, institutions should assess whether the obstacle lies in the technology, its integration, the organisation of work or the assumptions underlying the original business case.
During 2030, several national digital-government strategies reach significant milestones, including Russia’s objectives concerning electronic document management and instant-result public services and China’s broader AI+ development ambitions. Those targets should be evaluated through actual implementation evidence, not presumed achievement. International comparison should distinguish government AI effects from improvements attributable to wider digital infrastructure or procedural reform.
By 2031, the strategic assessment should concentrate on whether public administrations possess sustainable and accountable arrangements for continuously evaluating, modifying and replacing AI-enabled systems. The meaningful measure of maturity will be the institution’s ability to demonstrate reliable public-service outcomes, not the accumulated number of technologies introduced since the beginning of the decade.
The 2027–2031 institutional decision roadmap
Five Years to Accountable AI Productivity
Proposed decision sequence for public institutions, distinct from enacted regulatory milestones.
Application inventories, service baselines, cost records and responsible ownership.
Test complete processes, quality controls and repeatability before scaling.
Measure durable service, cost, workforce and resilience effects.
Distinguish achieved national milestones from commitments and digitalisation effects.
Operate continuous evaluation, supplier review and capacity allocation.
Verified European Regulatory Milestones
- 2 December 2027: Application of specified EU AI Act requirements for Annex III high-risk systems.
- 2 August 2028: Application of specified requirements for high-risk AI systems associated with Annex I products.
Applicability depends on the legal classification and circumstances of the system. Other AI Act provisions follow different commencement dates.
Sources: European Commission, AI Omnibus, July 2026; Consolidated EU AI Act, July 2026. The five-year roadmap is analytical, not an official timetable or a forecast of achievements.
Strategic choices by jurisdiction: coordinating different administrative systems without imposing a false international benchmark
The final international comparison must recognise that different institutional systems possess different authorities over procurement, workforce organisation, administrative procedures and public expenditure. A common productivity framework can improve the interpretation of evidence, but it cannot eliminate these differences or justify identical operational decisions. The relevant strategic question for each jurisdiction is therefore what additional institutional evidence and authority are required to transform its existing AI programmes into measurable improvements.
For Italy, the most immediate issue concerns implementation. The adoption of AgID’s guidelines in November 2026 provides a recent formal reference for organisational planning, but the productivity consequences depend on decisions made within individual administrations. A measurable service-performance framework should connect AI projects with identified procedures, their responsible managers, relevant staffing arrangements and the data required to evaluate outcomes. The existence of national guidance is a foundation for these activities, not evidence that they have already produced savings. AgID Guidelines for the Adoption of AI in Public Administration — November 2026.
For France, the principal opportunity concerns moving from evaluated employee activities to reproducible service outcomes. The government’s extensive assistant experiment provides evidence of adoption, perceived usefulness and task-level gains. The next analytical requirement is to document whether these gains can be reproduced across different administrative functions and whether institutions can identify how capacity is reassigned. This would distinguish successful general-purpose assistance from the additional organisational changes needed to improve complete procedures. L’Assistant IA: Evaluation of the Public-Employee Experiment — DITP, June 2026.
For Germany, the important question is whether local AI-assisted workflow improvements can be reproduced across administrations with different technical and organisational conditions. The federal government’s documented initiatives concerning permitting and other administrative processes provide useful opportunities for evaluation, but findings from particular pilots should not be treated as nationally representative. German public authorities must also evaluate whether technological-sovereignty objectives improve operational resilience at an acceptable lifecycle cost. Federal Government AI Policy and Administrative Initiatives — Bundesregierung, November 2026.
For the United Kingdom, the challenge is to connect an increasingly sophisticated body of evaluation and appraisal guidance with evidence of actual public-service and fiscal outcomes. Its employee time-saving studies provide relevant operational inputs, while the 2026 Digital and Data Benefits Framework and Green Book establish methods for assessing economic value, uncertainty and the distinction between financial and social effects. The decisive future evidence will concern the implementation of those methods across different public institutions. Digital and Data Benefits Framework — UK Government, April 2026 and The Green Book 2026 — HM Treasury, October 2026.
For the European Union, the question concerns the relationship between regulatory harmonisation and the measurement of administrative performance. The AI Act establishes obligations relevant to deployment, while member states retain different administrative structures and workforce arrangements. A common approach to performance definitions could improve cross-country learning without requiring identical procurement systems or public-employment policies. Such coordination would be most valuable where it enables governments to distinguish comparable service outcomes from superficially similar adoption statistics. Regulatory Framework for Artificial Intelligence — European Commission, 2026.
For the United States, the principal requirement is to connect federal AI inventories and governance arrangements with the operational and financial results of individual agencies. The GAO’s work demonstrates that substantial growth in AI use cases can coexist with weaknesses in workforce planning, inventory completeness and strategic management. Evidence concerning service performance and lifecycle costs would enable a more informative assessment of the benefits generated by federal AI investments. GAO-25-107653 — Government Accountability Office, July 2025 and GAO-26-107522 — Government Accountability Office, March 2026.
For China, the central empirical question is whether coordinated deployment, shared computing resources and common government AI platforms improve services at a lower total institutional cost. The official guidance provides a detailed architecture for implementation, but comparable evidence concerning realised working-time savings, operational reliability and administrative outcomes remains necessary to evaluate its economic effects. Government AI Large-Model Deployment Guidelines — CAC and NDRC, October 2025.
For Russia, future assessment should distinguish achievements in digital-government transformation from benefits attributable specifically to artificial intelligence. The government’s strategic objectives concerning electronic services, common information systems and administrative modernisation establish relevant implementation targets through 2030, but their achievement and economic consequences require operational verification. Strategic Direction for the Digital Transformation of Public Administration, Government Order No. 637-r — Russian Government, March 2024.
Final decision matrix — Institutional priorities and evidence required
| Jurisdiction | Principal evidence available | Decision-relevant gap | Competent institutional function | Consequential next assessment |
|---|---|---|---|---|
| Italy | National legislation and AgID adoption guidance | Service-level results and allocation of recovered capacity | Public-administration leadership and relevant oversight bodies | Auditable implementation and outcome evaluations |
| France | Employee trial, adoption indicators and task measurements | Replication across complete services | Ministerial operational and evaluation functions | End-to-end service and quality measurement |
| Germany | Federal initiatives and local workflow pilots | Scalability, costs and transferability | Competent federal, state and municipal authorities | Multi-site operational and lifecycle evaluation |
| United Kingdom | Employee evaluations and formal appraisal frameworks | Conversion into verified service and fiscal outcomes | Departments, finance and evaluation functions | Realised benefits and post-implementation appraisal |
| European Union | Common AI legal framework | Comparable administrative productivity definitions | EU and national competent institutions within their mandates | Harmonised evidence and regulatory implementation |
| United States | Agency inventories, OMB policy and GAO audits | Mission outcomes, workforce readiness and full costs | Federal agencies, OMB and oversight institutions | Agency-level performance and investment evaluation |
| China | AI+ strategic direction and government deployment guidance | Realised results of shared infrastructure and model reuse | Competent central and subnational authorities | Verified deployment, reliability and service outcomes |
| Russia | Digital-government targets and infrastructure strategy | AI-specific attribution and verified target achievement | Competent federal and subnational authorities | Implementation evidence and service-performance audit |
This comparison is deliberately non-ranking. The documents examined differ in purpose, scope, coverage and measurement method, and none supports a harmonised assessment of national AI-generated productivity.
Final net assessment — The institutional economics of AI-generated time
The international evidence supports a central conclusion concerning the relationship between artificial intelligence and public administration: the economic value of time saved by AI is determined by the institutional decisions that follow the technical improvement. A faster administrative task can release professional capacity, but the resulting public benefit depends on whether that capacity is available for another legitimate purpose, whether the organisation possesses the authority and competencies required to use it, and whether its allocation produces an observable improvement in final outcomes.
This conclusion does not diminish the significance of technological progress. Public administrations conduct substantial volumes of activities involving document preparation, information retrieval, record classification, correspondence and analytical support. AI can reduce the effort required for many such tasks, and official evaluations have begun to demonstrate measurable results under specified conditions. The United Kingdom’s adjusted assessment of approximately 19 minutes saved per working day among evaluated Copilot users is an important example of task-level evidence. France’s measured experimental gains demonstrate similar opportunities in particular forms of administrative work. These findings establish that AI assistance can have material effects on how public employees perform their duties.
They do not, however, establish the amount of additional public value generated across entire institutions. That requires a different category of evidence. Governments must demonstrate whether applications are completed sooner, whether more valid services are delivered, whether errors and appeals decline, whether citizens experience improved accessibility and whether the resources required to produce those outcomes have changed. The absence of these measurements creates a risk that government organisations will interpret technological diffusion as productivity growth without establishing the economic mechanism connecting the two.
The fiscal implications reinforce this distinction. Recovered labour capacity is not automatically a reduction in public expenditure. Public employees may remain in their positions while carrying out additional or more demanding work. That can produce legitimate economic value, but the benefit must be evaluated through the services delivered rather than recorded as a cash saving. Alternatively, automation may reduce expenditure on overtime, external support or duplicated systems, but such benefits require evidence of actual or credibly avoided costs, net of the resources necessary to operate the technology. Governments that conflate these categories risk overstating the financial returns from AI and making resource decisions on an insufficient evidentiary basis.
The workforce consequences are similarly dependent on institutional choices. Artificial intelligence can reduce routine workloads and create opportunities for professional development, more careful verification and additional public-service capacity. It can also intensify work, create new demands for technical supervision or increase dependency on specialised personnel. These outcomes are not predetermined by the technology. They reflect the interaction between managerial decisions, applicable employment arrangements, professional competencies and the structure of the services in which AI is introduced.
Cybersecurity and technological sovereignty add further dimensions to the assessment. Government institutions increasingly depend on software, computing infrastructure, models, data services and external suppliers whose availability and reliability influence administrative continuity. Shared infrastructure can reduce duplication, while excessive supplier concentration can create operational vulnerabilities. Data governance and verification may impose costs, but they also protect the integrity of administrative functions. The appropriate economic evaluation must therefore consider the complete operating system rather than the apparent productivity of the model in isolation.
The international comparison demonstrates that these issues arise under different models of state organisation. The United States combines federal policy requirements with agency-level implementation and congressional oversight. European administrations operate within national institutional structures influenced by a common AI regulatory framework. China’s government guidance places substantial emphasis on coordinated resources and shared deployment. Russia’s digital-administration strategy establishes objectives concerning common infrastructure and online services. These approaches create different conditions for the allocation and control of technological resources, but none removes the need for credible measurement of operational outcomes.
The strongest strategic opportunity for 2027–2031 is consequently the development of institutions capable of measuring, allocating and accounting for the capacity recovered through AI. This requires reliable service data, competent public management, appropriate professional training, full-cost investment appraisal and meaningful oversight. It also requires the willingness to modify or discontinue deployments whose expected benefits are not realised. An administration that continues purchasing technology regardless of observed outcomes cannot demonstrate responsible management of the resulting productivity dividend.
The distinction between technical efficiency and public value ultimately determines the meaning of administrative modernisation. A government can deploy increasingly sophisticated AI systems without improving the services experienced by its population. Conversely, a well-designed application addressing a specific operational constraint may produce substantial value even if its technical complexity is modest. The relevant criterion is whether the technology enables the institution to fulfil its responsibilities more effectively, reliably and economically.
The authority to decide how recovered time is used remains a matter of public governance. Technology suppliers influence what can be automated. Employees possess essential knowledge of operational requirements and exercise professional judgement. Public managers organise authorised activities and allocate available resources. Political and budgetary institutions determine priorities within their respective legal mandates. Supervisory and audit functions evaluate compliance and performance, while citizens experience the consequences through the services and decisions that affect them.
These responsibilities cannot be replaced by a single productivity metric. Their coordination requires an institutional framework in which the effects of AI are visible, decisions concerning its use are accountable and the resulting benefits can be evaluated against legitimate public objectives.
By 2031, the defining question will not be how many public employees have access to artificial intelligence, how many applications have been deployed or how many minutes individual tasks can save. It will be whether governments can demonstrate that the capacity released by AI has been transformed into measurably better public services, sustainable institutional capabilities and a more accountable use of public resources.
That is the central economic and governance test of artificial intelligence in the modern state.




















