AI Skills for Government: Practical AI for the Public Sector

Public sector AI use runs into three constraints most private sector guidance ignores.

Decisions have to be explainable — to the person affected, to an appeal, sometimes to a court or a committee. Records have to be defensible years later, when nobody remembers the context. And the people on the receiving end usually can't go elsewhere, which means an error isn't a lost customer, it's someone who's stuck with it.

None of that makes AI unusable. It makes the boundary between assistance and decision the thing you have to get right, and get right in writing.

The line that matters: drafting versus deciding

The workable rule is that AI can help produce material that a person then owns, and cannot produce an outcome that affects someone's entitlements, liberty, or record.

Summarising a case file so an officer reads it faster: fine. Drafting standard correspondence: fine. Producing a first-pass analysis a policy officer then interrogates: fine.

Determining eligibility, scoring an applicant, flagging someone for attention, or generating the reasoning that goes on the record as the basis for a decision: not without a human who has independently reached the same conclusion and can say why in their own words.

The test is not whether a human clicked approve. It's whether the human could explain the decision if the tool vanished.

Skill one: interrogate the output, don't review it

Reviewing means reading and looking for something obviously wrong. Interrogating means checking specific things: where did this figure come from, what's the source for this claim, what does this conclusion assume, what would change it.

Well-formatted output invites reviewing. Public sector consequences require interrogating. The difference is a habit, and it's the one worth training.

Skill two: place the checkpoints where reversal is impossible

Any process where AI touches a citizen-facing outcome needs an explicit map of where a human must intervene.

Put the checkpoints where an error can't be undone: before anything is communicated to a member of the public, before anything enters a permanent record, before anything triggers an entitlement change or an enforcement step.

Write this down. An unwritten checkpoint isn't a control — it's a habit that survives exactly as long as the person who has it.

Skill three: keep the record intact

Anything that might be reviewed later needs to show what was considered and why. If AI produced part of the reasoning, the record should reflect what it produced and what the officer did with it.

This is more than diligence. A file that can't distinguish officer judgement from generated text is a file that can't be defended, and the people who'll need to defend it won't be the ones who created it.

Skill four: watch for consistent unfairness

A human making inconsistent judgements produces scattered errors. An automated process applying a flawed rule produces the same error every time, at scale, in the same direction — and consistency is easily mistaken for fairness.

That pattern is harder to notice than random error precisely because it looks orderly. Anything applied at volume needs someone periodically asking not "is this consistent" but "is this consistently wrong about a particular group".

Aoife — the summary that lost the caveat

Aoife is a senior caseworker at a housing authority. Her team used AI to summarise lengthy application files so officers could triage faster.

One summary described an applicant's circumstances accurately but dropped a caveat from a medical letter — a consultant's note that a condition was expected to deteriorate. The summary said the condition was managed. The letter said it was managed for now.

The application was correctly assessed in the end, because the officer opened the source document. But it was close, and the summary was not wrong so much as flattened.

Her team's rule now: summaries are a navigation aid, not evidence. Any factor that materially affects a decision gets read in the original before it's relied on.

Ben — the checkpoint that wasn't written down

Ben runs a licensing team in a local authority. They'd built a workflow where AI drafted decision letters and an officer approved each one before sending. It worked well for eight months.

Then the officer who'd designed it moved teams. Her replacement, reasonably, treated approval as a formatting check — nobody had written down that the approval step existed to independently verify the reasoning, because to her it had been obvious.

Two letters went out with conditions that didn't match the application. Both were caught on appeal.

The control had never been a control. It was one person's understanding of why the step existed, and it left when she did.

The one thing

In the private sector, an AI error costs a customer. In the public sector it costs someone who often has no alternative and limited means to challenge it.

That doesn't mean using less AI. It means the boundary between assistance and decision has to be explicit, written down, and understood by whoever holds the job next — not just by the person who set it up.