Internal auditors face agents from two sides: as a subject to audit and as an assistant in their own fieldwork. Before either, they should learn to reconstruct a decision from the record, test the evidence behind an agent's depth and certificate, read the minimum chain behind each verdict, and protect their own independence when an agent drafts their work.
Why do auditors need their own preparation?
An auditor's question is always the same: show me how this decision was made, and show me that the control worked. With people, the answer lives in approvals, emails and procedures. With a governed agent, it lives in the record of verdicts, the certificate that granted authority, and the bounds that applied at the moment of action.
An auditor who does not understand those artefacts will either test the wrong things, such as the model's general accuracy, or accept a vendor's description of the controls at face value. Neither gives the audit committee an opinion it can rely on.
What should auditors understand when auditing agents?
- Depth is earned per agent, per organisation. An agent may read at depth 1, draft at depth 2, execute routine work at depth 3 and move money at depth 4. Each agent is certified on the organisation's own cases. The auditor's test is whether the depth in use matches the evidence.
- Certificates lapse. A certificate lasts 90 days. A run that clears no depth drops the agent to depth 1 the same day. Auditors should sample actions and check that each one happened under a certificate that was valid on that date.
- The minimum chain. Every action met five bounds: role scope, certificate, sources, the organisation's own limits and the platform guardrails. The lowest decided, and the verdict names it. If the verdict cannot say which bound decided, that is a finding.
- Four verdicts. Allowed, held, refused and withheld each leave a different trail. Held items should show a named approver. Refused and withheld attempts show where the agent tried to go and was stopped, which is useful evidence that the controls fire.
- Provenance. Facts are stamped sealed, signed, declared or external. An auditor checking a payment decision should confirm it did not rest on an external claim.
What may an agent do in the auditor's own work?
| The agent may | The auditor keeps |
|---|---|
| Read policies, prior reports and evidence files (depth 1) | Scoping the audit and setting the risk focus |
| Draft test scripts, sample selections and walkthrough notes (depth 2) | Setting the sample criteria and approving the sample |
| Compare transaction populations against stated controls | Concluding whether a control is effective |
| Draft findings for the auditor to edit | Rating findings and forming the opinion |
The pattern of drafting without deciding is shared with other evidence-heavy roles; see what paralegals should learn first and what caseworkers should learn first.
What does an audit of an agent look like?
Take an insurer's internal audit team reviewing a claims agent that settles small claims at depth 3. The auditor selects 40 settlements from the quarter. For each, the record shows the verdict, the bound that decided, the certificate number and its validity dates, and the sources consulted.
Two findings emerge. On three settlements the certificate had lapsed two days earlier, and the agent should have dropped to depth 1; the auditor traces this to a configuration fault and confirms it was already fixed. On one settlement the agent relied on a declared statement from the claimant where policy required a signed repair estimate. The auditor raises both, rates them, and asks management for remediation evidence. The agent drafted the walkthrough notes; the auditor wrote every conclusion.
Insurance audit teams in India and France work under different supervisory expectations, from IRDAI to the EU rules; see internal auditor training in insurance in India and in France.
How do auditors stay independent?
Independence is the auditor's core asset, and an agent can quietly erode it. If the agent selects the sample, frames the test and drafts the finding, the auditor may end up confirming the agent's view. Three habits help: set sample criteria before the agent runs, read underlying evidence for every exception rather than the agent's summary, and never use an agent that is itself in scope for the audit.
The case for auditors as the deciding check is made in why internal auditors are the real control, and the broader role map is in AI agent skills by role.
Where the Academy fits
Auditors who need to understand the system without running it start with the Foundations track of Pryme Intelligence Academy: about four hours on why agents must be governed, the decision chain, depth and certification, and data and trust. Those who will test live workspaces regularly can go on to the Operator track, whose module on reading the workspace maps closely to audit fieldwork.
Questions people ask
Should internal audit test the model itself?
Only where model risk is in scope. For most audits of an agent, the controls that matter are the bounds, the certificate and the record, and those can be tested directly.
Can auditors rely on the platform's own certification?
As evidence, yes; as an opinion, no. Auditors should re-perform a sample to confirm certificates, verdicts and bounds match what was reported.
How often should agents be audited?
Base it on risk. Agents at depth 4 that move money deserve more frequent attention than agents that only read and draft.