AI in government and public services 2026: from pilots to accountable use

Executive summary

Public bodies are moving from experiments to routine use of AI, under closer scrutiny than most private organisations. Use is rising quickly where it is counted, while strategy, measurement and governance lag behind. This report brings together official inventories, audit findings and international evidence, and sets out the rules that now apply to high-impact uses.

Use is growing fast where it is counted. US federal agencies reported 3,611 AI use cases in their 2025 inventory, up 105% from 1,757 in 2024, and 445 were classed as high-impact [1]. Health and Human Services, NASA and Veterans Affairs reported the most active use cases; uses by the intelligence community and the Department of Defense are not included in the inventory [2].
Strategy and governance trail adoption. Of 87 UK government bodies surveyed by the National Audit Office, 37% had deployed AI, typically with one or two use cases each, and only 21% had an AI strategy [3]. Fewer than half of the bodies using AI said use cases were always or usually identified at organisational level before deployment [3].
Pilots are still the norm. The OECD’s analysis of 200 use cases across 11 government functions found AI most common in public services, justice and civic participation, and less common in policy evaluation, tax administration and civil service reform [4]. Many initiatives remain pilots, held back by skills gaps, difficulty accessing and sharing data, risk aversion and weak measurement of results [4].
Rules for high-impact uses are now explicit. US federal policy (OMB memorandum M-25-21, April 2025) requires minimum practices for high-impact AI, including pre-deployment testing, an impact assessment, ongoing monitoring, human oversight and a route to remedy or appeal, and requires agencies to stop using systems that do not comply [5]. In the EU, AI used to decide access to public services and benefits is high-risk under the AI Act, public bodies deploying it must carry out a fundamental rights impact assessment, and high-risk systems already used by public authorities must comply by 2 August 2030 [6].
Public trust depends on data quality and transparency. The OECD warns that systems built on inadequate or skewed data can produce adverse outcomes for some groups, and highlights gaps in transparency and privacy [4].

Key findings

01US federal agencies reported 3,611 AI use cases in 2025, up 105% on 2024, including 445 classed as high-impact

02Only 37% of UK government bodies surveyed by the National Audit Office had deployed AI, and 21% had an AI strategy

03The OECD found many government AI initiatives remain pilots, held back by skills, data access and weak measurement of results

AI use cases reported by US federal agencies

2024

1,757

2025

3,611

Source: US Office of Management and Budget, 2026

UK government bodies and AI

Deployed AI

37%

Have an AI strategy

21%

Source: National Audit Office, 2024

What leaders should do

Keep a public inventory of AI use, with a named owner and a risk classification for each system.
Agree how each use case will be measured before it is piloted, and stop pilots that cannot show results.
Build impact assessments, human review and appeal routes into high-impact services before they go live.

What this report covers

Where government uses AI
What audit bodies have found
Barriers: skills, data and measurement
Rules for high-impact public-sector AI
Transparency and public trust
What to do in 2027

Methodology

We reviewed public sources published between 2024 and October 2026, favouring material from 2025 and 2026. These include official inventories (US Office of Management and Budget), audit reports (UK National Audit Office), international analysis (OECD), legislation (EU AI Act) and summaries of US federal policy. Each figure is cited to a dated source. Where we relied on a secondary report of a primary document, the source list says so. Surveys are self-reported and should be read as indicators, not audited results. We did not survey organisations directly.