Independence decides the list. Internal auditors should never let an agent form the audit opinion, rate a finding, set its own sample, confirm remediation is complete, alter reviewed working papers, or help audit a system it belongs to. An agent can gather evidence and draft; the judgement the board relies on must be a person's.
What makes audit red lines different?
In most roles, red lines protect customers or money. In internal audit they protect something less visible: the reason anyone listens to audit at all. The audit committee trusts an opinion because a qualified, independent person formed it on evidence they examined. If an agent quietly shaped that opinion, the opinion is worth less, even when it happens to be right.
That means some acts stay with the auditor even when the agent would do them accurately. Accuracy is not the test. Independence and accountability are.
Which acts stay with the auditor?
| Never let the agent | Why |
|---|---|
| Form or word the audit opinion | The opinion is the accountable act of the chief audit executive and the team |
| Rate a finding's severity | Rating drives board attention and management action; it requires judgement of context |
| Choose the sample or the sample criteria | Whoever picks the sample can steer the result |
| Confirm that remediation is complete | Closure must rest on evidence an auditor examined |
| Edit working papers after review sign-off | The file must show what was known when the conclusion was reached |
| Audit an agent or process it is part of | Nothing can audit itself |
| Share draft findings with the auditee | Timing and wording of disclosure belong to the auditor |
Auditors also sit on the other side of this: when they audit an operational agent, they look for the same red lines in its design. How they test those decisions is covered in how internal auditors review an agent's decisions.
How are the lines enforced?
The audit function sets them as its own limits in the workspace, not as instructions in the agent's brief. An agent asked to "finalise the rating" gets a refused verdict naming the limit. An agent that drafts a sample for review produces a held item that waits for a named auditor. Both are filed in the append-only record.
That record matters twice over in audit. It shows the audit committee which parts of the work an agent touched, and it gives external reviewers of the audit function a way to confirm that judgement stayed with people. Audit teams should be able to answer, for any engagement, "what did the agent draft, and who decided?"
What does a breach of a red line look like?
Take a regional insurer's audit team reviewing claims handling. The agent pulls the population of claims, proposes a sample weighted towards large settlements, and drafts four findings. Under deadline, a senior auditor accepts the sample as proposed and edits the findings lightly.
Nothing in the findings is wrong. But the agent's sample under-weighted small claims settled by a separate automated process, which is exactly where a control gap sat. The gap was found the following year by the external auditor. Had the auditor set the criteria first, including a stratum for automated settlements, the agent would have drawn the right sample. The red line was not about the agent's competence; it was about who decides where to look.
Teams in markets such as the UAE and Brazil add their own supervisory expectations to this list; see internal auditor training in insurance in the UAE and in Brazil.
How should audit functions prepare their people?
Start with a written list of red lines for the function, agreed by the chief audit executive and shared with the audit committee. Then train every auditor on what the agent may draft, what it must hand back, and how to read the record of a held or refused action. Caseworkers in the public sector keep a comparable list for decisions about people, set out in what caseworkers should never let an agent do.
A structured plan for preparing the team is in the training plan template for internal auditors. For the auditors who will test agents in depth, structured certification such as Pryme Intelligence Academy's Operator track gives a common baseline. The cross-role view is in AI agent skills by role.
Questions people ask
Can the agent draft the audit report?
Yes, as a first draft for the auditor to rewrite. The opinion section and finding ratings should be written by the auditor, not edited from the agent's version.
Should audit use the same agent platform as the business?
It can, with separate workspaces, separate limits and no access to change the business's agents. What matters is that audit's agents cannot be configured by the people being audited.
Is it a finding if the business lets an agent close its own exceptions?
Usually yes. Self-closure removes the independent check, and the record will show it.