How do internal auditors review an AI agent’s decisions?

Internal auditors review an agent's decisions by reperforming them. From the record, they rebuild what the agent saw, which bound decided, and who approved any held action. Then they test each link: was the agent certified for that depth, was the approver qualified at that moment, did the reason support the outcome?

What does internal audit test that first-line review does not?

First-line reviewers judge whether each decision was right. Internal audit judges whether the system of decisions can be relied on. That covers the agent, the people around it and the record. Four questions frame most audits of an agent:

1. Did the agent act within its authority? Its depth should match what its current certification supports and what the business allows.
2. Did the bounds work? Actions that should have been held were held; actions that should have been refused were refused.
3. Were people qualified and attentive? Approvers held a current credential, gave reasons, and did not bulk-approve.
4. Is the record complete and unaltered? Every verdict is filed in an append-only record, so any decision can be reconstructed.

What are the steps?

1. Define the population. All decisions by the agent in the period, by capability and verdict.
2. Select a sample. Random, plus targeted: held items approved fastest, decisions just after a change, refusals that customers complained about.
3. Reconstruct each decision from the record. Inputs, facts used and their provenance, the deciding bound, the verdict.
4. Test the bound. Using the business's own limits and the agent's certification at that date, would you expect the same verdict?
5. Test the human step. Approver identity, certificate status at the time, reason quality, time taken.
6. Test certification. Agent certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day. Check the agent never acted above its certified depth.
7. Conclude and report. Findings with condition, criteria, cause and effect.

What evidence does each test rely on?

An example: an insurer's claims agent

Take an insurer's internal audit team testing a claims agent. In a sample of 60 decisions, every verdict reconstructs cleanly and the bounds behaved as configured. Two findings emerge. First, a claims handler approved four held settlements in the week their reviewer certificate had lapsed; the workspace allowed it because approval rights were not yet tied to certification. Second, a customer limit on settlement value was raised by the claims manager without going through change control.

Neither caused a loss. Both are control gaps. The auditors report them with evidence from the record, and management agrees to tie approval rights to current certificates and route limit changes through change control. Detail on this kind of audit in specific markets is in AI agent training for internal auditors in insurance in France and in Brazil.

What about the audit team's own agent?

Where internal audit uses an agent to sample, draft findings or prepare workpapers, auditors review its drafts as they would a junior colleague's: approve, edit or decline with a reason, and never let it close findings on untested evidence. The hard limits are in what internal auditors should never let an AI agent do.

Training an audit team to do this work is covered in how to train the internal audit team to review AI agent decisions. Whether the role changes is answered in will AI agents replace internal auditors. The role map is in AI agent skills by role.

Where the Academy fits

Pryme Intelligence Academy gives auditors the reading skills these tests depend on. Foundations covers depth, certification and the decision chain; Operator covers human review and reading the workspace. Professional certificates are issued through VerifyDoc.ai, so an auditor testing approver qualification can check any certificate at source, including whether it was lapsed or revoked.