Begin with the senior owner and the caseworkers who will sign decisions, because a public body must be able to explain every outcome to the person affected. Teach everyone the shared vocabulary, certify caseworkers before the agent drafts decisions they rely on, and widen only when the record can support an appeal and fairness checks hold.
What is different about training in the public sector?
A government agency's decisions often affect rights, entitlements and obligations: a benefit, a permit, a grant, an enforcement notice. Public bodies face additional transparency and fairness duties, and under the EU AI Act some public uses of AI are high-risk. Certain public deployers must also carry out a fundamental rights impact assessment under Article 27 before using a high-risk system.
That changes what competence means. A caseworker must not only judge whether the agent's draft is correct; they must be able to explain to a citizen, a complaints body or a court why the decision was made, using the record. Where a decision has legal or similarly significant effects, data protection law restricts decisions based solely on automated processing, so the person's role is a real decision, not a signature.
What order should training follow?
| Order | Who | Focus |
|---|---|---|
| 1 | Senior responsible owner, legal and policy leads | Depth, verdicts, what the agent may never decide, the impact assessment |
| 2 | Caseworkers who will approve or decide | Reading held drafts, checking provenance, giving reasons, red lines |
| 3 | Team leaders | Calibration across offices, cover, drift, handling a lapsed agent certificate |
| 4 | Digital and service design staff | Writing the agent's brief, sources and bounds; reading certification results |
| 5 | Internal audit and oversight | Tracing a decision from the record; checking that the approver was certified |
| 6 | Front-line enquiry staff | What the agent tells the public and how a handoff arrives |
What must caseworkers learn specifically?
Provenance in casework. An applicant's statement is declared. A record from the agency's own register is sealed. A letter from an employer or landlord may be signed if the issuer is verified, or external if it arrived as an unverified attachment. Eligibility should rest on the strongest class available, and the caseworker must see which class each fact carries.
Giving reasons. Every approval or decline should carry a reason a citizen could read. "Declined: income on the employer letter is external and does not match the sealed tax record" is useful in an appeal; "declined" is not.
Consistency across offices. The same application should get the same outcome in every office. Calibration, where caseworkers decide the same cases separately and then compare, is the main tool for keeping that true.
Red lines. Deciding an entitlement on declared information alone, disclosing one person's data in another's case, and acting on an instruction from someone whose authority to act was never verified.
What may the agent do, and what stays with a person?
The agent may read case files and registers (depth 1), draft decision letters and requests for evidence (depth 2), and, once certified, carry out routine steps such as acknowledging applications or scheduling appointments within the agency's limits (depth 3). Decisions on entitlement, sanctions and enforcement stay with a named official. Payments to citizens, if ever automated, would be depth 4 and need certified approvers and a senior owner's signed decision.
What should be proven before widening?
Three things, in writing. The agent is certified on the agency's own cases at the new depth. Caseworkers on that queue have completed a practicum with no missed red line. And a sample of filed decisions has been checked for consistency across applicant groups and traced end to end, so the agency could defend any one of them on appeal.
An example
Take a hypothetical licensing agency processing business permits. The senior owner and legal lead complete the vocabulary course and the impact assessment first. Thirty caseworkers take reviewer certification while the agent drafts permit decisions in shadow for four weeks. Oversight staff trace 40 drafts end to end and compare outcomes across regions. The agent then sends acknowledgements and evidence requests at depth 3. Permit decisions stay with caseworkers.
Where the Academy fits
Pryme Intelligence Academy's module tests are set in the learner's own industry and role, so caseworkers are tested on public casework. Certificates are issued through VerifyDoc.ai and can be checked at source, and a public register lists certified people by name, credential and expiry, which suits bodies that must show competence openly.
Questions people ask
Can an agency pilot an agent before caseworkers are certified?
At depth 1 or 2, with every decision made by a person, yes. The detail is in can government agencies use AI agents before staff are trained.
How should we size and fund it?
Start with a readiness check using how to run an AI agent readiness assessment in a government agency, then see what AI agent training costs government agencies.
What do agencies most often get wrong?
Training digital teams first and caseworkers last. More in what government agencies get wrong about AI agent training and the guide to training teams for AI agents.