Reconstructing any single agent decision from the record is the internal auditor's most valuable skill, because every other test depends on it. After that, ranked: identifying which bound decided a verdict, testing that approvers were entitled at the time, testing certification evidence, testing change control on limits, and keeping the audit function independent of the agents it audits.
What is different about auditing an agent?
An auditor of a manual process samples transactions and asks people what they did. An agent leaves a fuller trail: every action meets five bounds at once (role scope, certificate, sources, the customer's own limits and the platform guardrails), the lowest wins, and the verdict names the bound that decided. Every verdict is filed in an append-only record.
That changes the work. The question is no longer "can we find evidence?" but "do we know how to read it, and does it show the controls worked?"
Which skills matter most, in order?
- Reconstructing a decision. Pick any action and rebuild it: what the agent saw, with what provenance, which verdict it reached, which bound decided, and who approved if it was held. An auditor who cannot do this cannot test anything else.
- Identifying the deciding bound. If refusals were mostly decided by platform guardrails rather than the customer's own limits, the business may not have set its limits at all. If allowed actions sit just under a limit, the limit may be gamed or mis-set.
- Testing approver entitlement. Held actions need a named person. Was that person authorised for that amount, and, where the business ties approval rights to a current certificate, was the certificate current on that date?
- Testing certification evidence. Depth is earned, not configured. Attest replays the agent against real cases from the business's own operation, across seven gates. The auditor checks that the certificate covered the capability used, had not lapsed (certificates last 90 days), and that a failed run dropped the agent to depth 1.
- Testing change control on limits. Who changed a limit, when, with whose approval, and whether that person also reviews the agent's output.
- Staying independent. Auditors should read the record but never approve held actions, or they audit their own decisions.
What should the auditor test, and how?
| Control | Test | Evidence |
|---|---|---|
| Verdicts follow the bounds | Reconstruct a sample | Verdict record entries |
| Held items approved by entitled people | Match approver to delegated authority on the date | Approval entries, authority matrix |
| Agents certified for the capability used | Compare action date to certificate window | Attest certificates |
| Limits changed under control | Trace changes to approvals | Change log |
What does the first skill look like in an engagement?
Take an insurer's claims payment agent. The auditor selects 30 claim payments made at depth 4 over a quarter. For one, a 9,500 payment, the record shows the agent proposed it, the verdict was held because it exceeded the customer's own auto-pay limit of 5,000, and a claims team lead approved it.
Reconstruction shows the repair estimate was signed by an approved garage, the policy terms were sealed, and the team lead's authority covered 10,000. But the certificate for the payment capability had lapsed four days before the approval, and the agent should have been at depth 1. The finding is not the payment; it is the renewal process that let a lapsed capability keep operating. The fix goes to the agent's owner.
Insurance internal audit has market-specific angles; see internal auditor training in insurance in Mexico and in India. Other reviewing roles, such as caseworkers, read the same record from the inside.
How should an audit team build these skills?
Pair an auditor with an operator for one week of live review before the first engagement. Then reconstruct ten decisions together, including at least one refusal and one withheld action. The case that auditors carry the real control is in why internal auditors are the real control, and the role map is in AI agent skills by role.
Where the Academy fits
Pryme Intelligence Academy's Foundations track explains the decision chain and certification in about four hours, enough for an auditor to plan an engagement. Auditors who test agents often benefit from the Operator track's modules on reading the workspace and when something goes wrong. Certificates issued through VerifyDoc.ai can be checked at source by number or QR code, which auditors can use when testing approver entitlement.
Questions people ask
Can the agent help audit itself?
It can extract and summarise the record for the auditor. The conclusions must be the auditor's, and the extract should be checked against the source.
How large should the sample be?
Use your normal sampling approach, weighted towards depth 4 actions and held items. Refusals deserve a small sample too: an agent that refused too much is a finding as well.
What if the record has gaps?
An append-only record should not have gaps. If one appears, that is a high-priority finding about the platform configuration, not the agent.