Yes, for drafting and preparation, and no for decisions. A government agency can use an agent to summarise case files, and draft correspondence while officers train. Any decision about a citizen, must wait until the officers reviewing the agent's work are trained, because public bodies owe a higher duty of fairness and explanation.
Why do public bodies need more care?
Public bodies face transparency and fairness duties that private firms often do not. A citizen refused a benefit or a licence is usually entitled to reasons and to challenge the decision. Under the EU AI Act, some public uses, such as assessing eligibility for public assistance, are classed as high-risk, which brings human oversight duties. Decisions based solely on automated processing with significant effects are also restricted under GDPR.
An officer who approves an agent's draft decision without understanding it is not exercising oversight. If that decision is later challenged, the agency will need to show who decided and how.
What can the agent do from the start?
- Case summaries from the file and the citizen's correspondence, each fact linked to its source (depth 1 and 2).
- Draft letters acknowledging applications, requesting information or explaining process (depth 2, sent by an officer).
- Guidance lookups for staff, citing the current version of the rules (depth 1).
- Completeness checks on applications, listing missing documents (depth 2).
- Freedom of information and correspondence triage, proposing a route for a person to confirm (depth 2).
What waits for trained officers?
- Eligibility and entitlement decisions.
- Refusals, sanctions, penalties and enforcement steps.
- Anything involving a vulnerable person or a safeguarding concern.
- Sending correspondence without a person (depth 3), even for routine letters, until officers can sample effectively.
What does an early deployment look like?
Take a national agency processing grant applications from small businesses. It starts an agent checking completeness and drafting requests for missing documents. Officers review every draft.
In week three, an officer notices that the agent treats a self-declared turnover figure as verified, and drafts an approval-ready summary on that basis. The officer flags it: turnover must be supported by filed accounts. The correction goes into the agent's brief, and the agency's own limits now hold any summary where a key figure is declared rather than signed. During the same period, all grant officers complete foundational training, and the senior officers who sign off decisions work towards operator certification. The agency only considers letting the agent send routine acknowledgements without review once those officers are certified. A fuller programme is described in what an AI agent training programme for government agencies should include, and the planning steps in how government agencies should train staff before deploying AI agents.
Which controls fit a public body?
- The agency's own limits cap agents at depth 2 until a named senior official approves more.
- Every verdict filed in an append-only record, so any decision can be reconstructed for an appeal or an audit.
- Reasons written by the deciding officer, not copied from the agent.
- Provenance on every fact: declared information never counts as verified.
- A published explanation of where agents are used, consistent with the agency's transparency duties.
Agencies tend to repeat certain errors, such as training IT and not caseworkers; see what government agencies get wrong about AI agent training.
Where the Academy fits
Every officer can take Pryme Intelligence Academy's free Foundations track, about four hours across six modules. Officers who review agent drafts take the Operator track, with module tests set in the public sector and their role. The free readiness assessment, for teams of up to 25, shows where understanding is weakest and which agent to start with. The topic guide on training teams for AI agents covers the wider sequence.
Questions people ask
Can a citizen be told an agent was involved?
They should be. Telling citizens where agents prepare work, and that a named officer decides, supports trust and the agency's transparency duties.
Does training need to be recorded?
Yes. If a decision is challenged, the agency should be able to show that the officer who approved it was trained and current.
Can agencies use agents on sensitive cases at all?
For preparation, often yes, with tight access. Decisions on sensitive cases should stay with experienced, trained officers.