Governments are using AI—but how many can prove it improves public services?
The OECD reports government AI use in 35 of 36 participating member countries, yet only 10 say they measure financial or non-financial impact. The comparison maps national practice; it does not audit individual systems or prove public benefit.
The short version
- What happened
- Thirty-five of 36 reporting OECD countries use AI in at least one government activity, but a country-level yes does not reveal the number, scale or quality of deployed systems.
- Who it affects
- Residents may receive faster services, but need clear notice, explanations, correction routes and accountable human review when AI contributes to a consequential decision.
- What remains unproven
- The data are primarily country self-reports and do not independently audit each AI system or verify claimed benefits.
Keep in mind
The data are primarily country self-reports and do not independently audit each AI system or verify claimed benefits.
The denominator is countries, not projects, decisions or people affected; one pilot and many large deployments can produce the same binary response.
Most underlying observations cover 2023–2024, so the report does not describe every deployment or policy change made during 2025 or 2026.
Missing countries and uneven reporting periods constrain comparisons, while the AI chapter is not a representative survey of residents or public servants.
Source analysis · 2 linked sources. Primary analysed source date: 1 October 2026.
Read the full report · 7 min The evidence, methods, implications and limitations
What changed in the OECD's final 2026 outlook
The OECD's final Digital Government Outlook 2026 shows that public-sector AI is no longer rare, but the machinery for proving that it works is still thin. Thirty-five of 36 reporting OECD countries said they used AI in at least one of four areas: internal administration, public-service delivery, policy design, or oversight and accountability. Use was most common in internal processes, reported by 31 countries, and public services, reported by 27. Only 13 reported use in policy design and 12 in oversight or accountability.
This article updates the portal's earlier short record in place. The earlier entry pointed to a pre-release page and carried an incorrect June source date. The controlling OECD report is dated 1 October 2026, runs to 200 pages and has DOI 10.1787/605281b2-es. The correction matters because publication date is part of the evidence trail: it should tell readers when the source became available, not when an anticipated page first appeared.[1][2]
The denominator is countries, not projects or citizens
The report combines the OECD Digital Government Index and open-data evidence across 36 OECD members and eight accession candidates. Its AI chapter mainly uses responses to the 2025 Digital Government Survey 3.0. Most AI percentages use 36 OECD countries and six accession candidates. The underlying reference period generally runs from 1 January 2023 to 31 December 2024, despite the survey and report labels referring to 2025 and 2026. Germany and the United States are missing from several AI measures; Indonesia and Thailand use an earlier reporting period in parts of the annex.
Each country is one observation in the headline ratios. A reported 'use' means the central government identified at least one qualifying application; it does not measure the number of systems, residents affected, transaction volume, accuracy or benefit. Country self-reporting is appropriate for mapping policy and administrative practice, but it can reflect different definitions and documentation standards. A nation with one limited pilot can sit in the same yes-or-no category as one running several mature services.[1][2]
Adoption is running ahead of safeguards and transparency
All 36 OECD countries reported at least one form of safeguard, yet operational controls were much less common. Fourteen countries required risk assessment before deployment, 12 reported internal review committees and 11 conducted post-deployment audits. Only 11 had a formal algorithmic-transparency standard and six maintained an open algorithm register. Twenty-five of the 36 had neither of those two transparency mechanisms, despite widespread use of AI in administration and services.
Procurement readiness also lagged investment. Thirty-two countries reported some funding for government AI, but only 21 offered central support for buying AI goods and services. That gap can leave individual departments negotiating data rights, audit access, liability, security and supplier lock-in without consistent expertise. For taxpayers, a funded pilot is not yet evidence of value; contracts need measurable outcomes and an exit route when systems fail or costs rise.[2]
Only ten governments reported measuring impact
The sharpest evidence gap concerns outcomes. Only 10 of 36 OECD countries—28%—reported any financial or non-financial impact measurement for government AI use cases. Yet 18 countries said adoption decisions were informed by evidence about possible efficiency gains or cost savings. The report therefore reveals a mismatch between the confidence used to justify adoption and the smaller number of governments that say they have measured results prospectively or retrospectively.
Training is broader than evaluation. Thirty-two countries reported programmes to support AI skills in government, including 28 offering practical-use training and 22 covering ethical use. Only 13 offered training specifically focused on AI in public services, and 13 on policy work. General literacy helps staff recognise risks, but people operating welfare, immigration, tax, health or justice systems need domain-specific instruction, escalation routes and authority to stop unreliable automation.[2]
What this means for people using public services
AI can make a public service faster by retrieving records, drafting correspondence or directing a person to the right process. It can also make an error harder to challenge when the model, supplier and department each sit behind different layers of accountability. Only 15 of 36 countries involved service users during implementation, according to the OECD's headline summary, and just eight reported a citizen feedback or complaints mechanism. Those are process measures, not proof that the mechanisms are accessible or effective.
The practical test is whether a resident can understand that AI affected a decision, obtain the relevant explanation, correct bad data and reach a responsible human before harm becomes irreversible. Governments should publish system inventories, intended purposes, affected populations, evaluation results and incident routes. They should preserve non-digital access where automated or online-only delivery would exclude people because of disability, language, connectivity, documentation or digital confidence.[2]
Limits and what would change the assessment
This is an official cross-country benchmarking report, not an independent audit of every system. The OECD verifies and standardises survey responses, but governments remain the primary reporters. Missing countries and different reporting periods limit comparisons. Binary indicators also cannot reveal whether an assessment was rigorous, whether an algorithm register is complete, or whether a complaints process changed an outcome. The report should not be read as a league table of safe or effective national AI.
Confidence would increase if countries published service-level denominators: number of decisions, error and appeal rates, processing time, cost per case, accessibility failures and outcomes for groups exposed to discrimination. Independent audits and user research should test whether safeguards work in practice. The assessment would weaken if governments continue expanding AI inventories while impact measurement, open registers and redress remain rare. For now, the OECD supports a precise conclusion: adoption is widespread, but evidence of public value and operational accountability has not kept pace.[1][2]