US federal agencies reported 3,611 AI use cases last year. That's more than double the year before (and most of them are still pilots).

So AI in public services is very real. It just looks a lot less like robot civil servants and a lot more like call centres, summaries and paperwork.

Here's what governments in the US, UK and EU are actually doing with it right now.

How the US federal government uses AI

The 2025 federal inventory came out in April. Nextgov pulled out the numbers:

→ 3,611 use cases across 56 agencies, up 105% from 1,757 → HHS reported the most, with 447 → 445 cases were flagged "high-impact", and the VA alone had 215 of them → Microsoft Copilot showed up in 102 use cases

The examples are pretty unglamorous. "Summarization of Clinical Data". A VoiceBot for call centre modernisation.

Now the less exciting part. Brookings went through the same data and found:

  1. nearly 60% of use cases with deployment info were still in pilot or pre-deployment
  2. about 63% involved contractors
  3. more than 85% of deployed high-impact cases were missing some required risk-mitigation information

That third one matters. High-impact means things like health and benefits, where a wrong answer lands on a real person.

How the UK is using AI in public services

On 14 May the UK launched GOV.UK Chat inside the GOV.UK app:

→ you ask questions in plain English (childcare, pensions, buying a first home) → answers come from tens of thousands of official guidance pages, out of 80,000+ on the site → some government call centres handle around 100,000 calls a day, and the government says research suggests up to half of callers' questions could be answered by the tool → it's opt-in, you need a GOV.UK One Login, and helplines stay open

Early trials showed the most demand for tax, driving and benefits questions. No accuracy figures were published though. The release just calls the answers "reliable".

And here's the twist most coverage skipped. Earlier in May, Global Government Forum reported the UK had closed four of its own AI pilots (Parlex, Redbox, Caddy and Medguard). Only Extract (digitising old planning documents) and Consult (analysing consultation responses) survived, picked for having the highest public impact.

The reason? Money and staff got moved to replacing old IT systems, which the department called "our most urgent task".

bad read: "the UK is going all in on AI" better read: "the UK is cutting the AI that doesn't pay off and fixing the plumbing first"

How the EU is regulating AI in government

The EU is mostly moving on the rules side.

On 7 May, negotiators agreed a deal to push back the AI Act's high-risk obligations for use-based systems from 2 August 2026 to 2 December 2027, according to Covington's analysis.

That list (Annex III) includes AI used by public authorities to judge who is eligible for public benefits and services. So the strictest rules for exactly this kind of government AI just got another 16 months.

What uni won't teach you about AI in government

Politics, policy and public admin degrees are strong on theory, institutions and law. Where they fall short:

  1. procurement is invisible. 63% of US use cases involved contractors. Who writes the contract shapes the tool more than any strategy paper
  2. implementation is treated as boring. The UK just showed old IT can matter more than shiny AI
  3. evaluation gets skipped. A tool with no published accuracy figures is a policy question, not a tech detail

If you've read our piece on AI-hallucinated citations in court, same lesson: someone has to check the output, every time.

The critical-thinking move: 5 questions for any government AI tool

Use these in seminars, dissertations or a civil service interview:

  1. Is it live or a pilot? Nearly 60% of US cases weren't fully deployed. Announcements aren't outcomes
  2. Who's on the receiving end? Internal summaries are low risk. Benefits decisions are not
  3. What's the error rate, and who published it? If nobody did, ask why
  4. Who built it, and who can audit it? Contractor tools can be harder to inspect
  5. What happens when it's wrong? A helpline, an appeal, a human. GOV.UK Chat keeps helplines open. Good

We made a similar argument on AI dynamic pricing in retail: the pushback always lands where people feel treated unfairly by a system they can't question.

What to do this month

  1. Open the US AI use case inventory and pick one agency. Find one public-facing case and one internal one
  2. Try GOV.UK Chat on a question you already know the answer to. Note what it gets right and what it skips
  3. Read the Brookings piece and write a 200-word summary. It's honestly better interview prep than most textbooks
  4. Follow the EU AI Act timeline from our public sector and policy hub, because those Annex III dates will shape graduate jobs in compliance

so TLDR: governments are using AI for call centres, summaries and paperwork, mostly still in pilots, whilst the rules for high-stakes uses keep getting pushed back. Learn to ask what's live, who checks it and who it affects.

If a government chatbot gave you the wrong answer about your student loan, who do you think should be responsible?

Sources

Get the next issue by email

theVANTAGEd is written by students to bring critical thinking to AI across real industries. One issue every two weeks.