Government agencies tend to buy training with the system, train the digital team that ran the procurement, and leave caseworkers to learn on live cases. That misses what public service demands of anyone overseeing an agent: reasons that stand up to challenge and fairness checked case by case. The caseworker who approves the agent's draft carries both.
Why isn't the supplier's training enough?
Procurement usually buys a training package with the platform: how to log in, where the queue is, what each button does. It is delivered once, near go-live, to whoever is available.
That covers the system, not the decision. A caseworker approving an agent's draft decision letter needs to know what the agent was allowed to do, which bound held the case and why, and where each fact came from. A household income figure held in the agency's own records is sealed. The same figure stated by the applicant on a form is declared. A letter from an employer uploaded with the claim is external until checked. A decision that rests on the wrong one is open to challenge, and no screen tour teaches the difference.
Who actually needs training first?
Agencies often train the digital or transformation team first and most deeply, because they led the project. They will configure and monitor the agent, but they rarely decide cases.
The people who need training first are the caseworkers, assessors and team leaders who will meet held actions every day: the draft eligibility decision, the overpayment letter, the grant payment waiting for release at depth 4. Then the senior officials who own each service and answer for its outcomes, because they decide what depth to seek. The digital team should be trained too, but as builders and stewards, not as a substitute for the front line.
Why do written reasons matter more in public service?
Public bodies face additional transparency and fairness duties. A citizen may ask why a decision was made, an ombudsman or tribunal may review it, and the reasons on file must hold up.
Every verdict in a governed system is filed in an append-only record, so the reasoning trail exists. The weak link is the person's part of it. A caseworker who approves a draft refusal with no comment, or who overturns a draft approval with "not eligible", has left a record that cannot explain itself. Train reviewers to write reasons a colleague, a manager and the citizen could all follow: what was checked, which fact decided it, and which rule applied. Practise on real past cases, and mark the reasons as well as the outcomes.
How does fairness change the reviewer's job?
An agent applies the same bounds to everyone, which is part of its value. It can still produce unfair outcomes if a source is skewed, a rule bites harder on some groups, or a pattern of held cases falls mostly on people who struggle to provide documents.
Caseworkers are often the first to see this. Training should teach them to notice it and report it to the service owner, not just to fix the individual case. In the EU, the AI Act treats some public uses of AI as high-risk, and Article 27 requires certain deployers, public bodies among them, to carry out a fundamental rights impact assessment before using such systems. Article 4 adds a duty to ensure a sufficient level of AI literacy among the staff who operate them. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What happens when trained staff move on?
Public sector programmes run for years and staff rotate between departments. A cohort trained at launch can be spread across the organisation within a couple of years, while the agent carries on with new people approving its work.
Agencies need training that survives rotation: assessment built into induction for any role that approves held actions, a register of who is certified and until when, and recertification on a cycle. Certificates for agents lapse after 90 days and a run that clears no depth drops an agent to depth 1 the same day; people's certificates should expire too, on a longer cycle. An agency can adopt the practice of granting approval rights only to staff whose certificate is current, depth by depth.
What should an agency do instead?
- Treat the supplier's system training as the first hour, not the programme.
- Train caseworkers and team leaders first, then service owners, then the digital team as builders.
- Assess reasons as well as outcomes on real past cases.
- Teach fairness as part of review, with a clear route to report patterns.
- Build assessment into induction and keep a register with expiry dates.
The broader order is in the guide to training teams for AI agents. For an agency's version, see how government agencies should train staff before deploying AI agents, the AI agent training plan template for government agencies, and how much AI agent training costs for government agencies. Insurers face a similar problem with claim decisions customers can contest; what insurers get wrong about AI agent training covers their version.
Where the Academy fits
Pryme Intelligence Academy's free readiness assessment lets a casework team of up to 25 take the Foundations test set in its own sector, with a report showing scores by concept, the gaps, and the first agent to start with. Caseworkers then take the Operator track. Certificates are issued through VerifyDoc.ai, and a public register lists certified people by name, credential and expiry.
Questions people ask
Do senior officials need hands-on training?
They need enough to own the service: what depth each agent works at, who approves its held actions, and what the record shows. They do not need to clear the queue. A short, assessed session on the vocabulary is usually right.
Should citizens be told an agent helped with their case?
Tell people when they are interacting with an AI system, and be clear in decision letters about how their case was handled and who decided it. A named person approving the decision is a strong reassurance.
How do we train staff who handle sensitive cases?
Use anonymised real cases from the same service, with extra care around personal and special-category data. Assess them on the hardest cases they will meet, not just routine ones.