How can government agencies prove staff are trained to oversee AI agents?

Public evidence faces more scrutiny, so build it to be shown. For every decision an agent helped prepare, the agency should be able to name the official who approved it, show that their certificate was current and checkable at source, and show the training behind it covered this kind of case. That answers auditors, appeals and citizens.

Who asks government agencies for proof?

A private company answers mainly to its regulator, its auditors and its customers. A public body answers to all of these and more: public auditors, ombudsmen, tribunals hearing appeals, elected representatives, the press and individual citizens asking why a decision went against them. Public bodies also face additional transparency and fairness duties, and under the EU AI Act some public uses of AI are classed as high-risk, with Article 27 requiring certain deployers to carry out a fundamental rights impact assessment.

These audiences ask different questions, but they converge on one: was the person who made this decision about me competent to judge what the system put in front of them? An agency that can only point to a course completion rate has not answered it.

What should the evidence include?

The second column is the test. If the agency cannot answer each of those questions from records, it should not expect a tribunal or auditor to accept its assurances. The AI agent training plan template for government agencies helps set out the competence standards and roles.

What should officials be trained to catch?

Public sector agents tend to work in benefits, licensing, grants, enforcement and correspondence. The training that produces meaningful evidence focuses on:

• Discretion. Many public decisions involve discretion the agent cannot exercise. Officials must recognise when a draft applies a rule mechanically where policy requires judgement.
• Evidence weight. An applicant's statement is declared, a record from another department may be signed, and a document from the agency's own case system may be sealed. Officials should know how much each can bear.
• Fair treatment. Spotting patterns where an agent's drafts consistently disadvantage a group, and escalating rather than correcting case by case.
• Reasons. Many public decisions must give reasons. An official approving an agent's draft must be able to stand behind the reasons it gives.

The piece on what government agencies get wrong about AI agent training covers the commonest gaps, such as training only the project team.

An appeal example

A licensing authority uses an agent to draft decisions on premises licence renewals. An applicant appeals a refusal, arguing the decision ignored a letter they sent. At the hearing, the authority produces the filed record: the agent's draft, which did not mention the letter; the approving officer's name; and her certificate, current on the day.

The record also shows the officer added a line noting she had read the applicant's letter and why it did not change the outcome. That line, written because her training stressed checking every source the applicant supplied, is what lets the authority defend the decision. Without it, the authority would have had a certified officer and a weak decision.

Which controls make the evidence trustworthy?

• Rights linked to current certificates, so lapses remove approval rights automatically.
• An append-only record that cannot be edited after the decision, which is what a tribunal will expect.
• Retention aligned with appeal windows and public records duties.
• A named senior owner for the evidence, who can explain it to elected members.

Agencies sit alongside regulated sectors with similar needs; see how utilities prove staff are trained and how fintechs do it. The guide to AI literacy and certification covers the general model.