How do internal auditors handle an action an AI agent has held for them?

Independence comes first. An internal auditor handling a held action asks whether it is audit work, such as an evidence request, a sample or the closure of a finding, or an operational decision routed to them by mistake. Operational items go back unapproved. Audit items are checked for evidence and provenance, then decided with a reason.

What does an audit agent hold?

Internal audit teams use agents to plan, sample, gather evidence and draft. A typical audit agent reads systems and documents at depth 1 and drafts at depth 2; it rarely acts beyond that, because audit outputs carry the function's name. Held items commonly include:

• Evidence requests to auditees, with deadlines.
• Sample selections for testing, before they are fixed.
• Draft findings and ratings.
• Closure of findings after management submits remediation evidence.
• Draft reports before release to management.

Sometimes operational agents also route items to internal audit, for example a held write-off flagged "for audit sign-off". That is a design error. An auditor who approves operational transactions cannot later audit them.

What should the auditor check?

Whether the item is audit work. If it is an operational decision, decline it with the reason "Outside internal audit's mandate; independence" and tell the owner of the routing rule.

Evidence sufficiency. For a finding or a closure, is there enough evidence, of the right kind, to support the conclusion? A sample of five transactions may be enough for a walkthrough and not for a conclusion on operating effectiveness.

Provenance. Remediation evidence from management is declared until tested. A screenshot of a system setting is weaker than a configuration report generated by the system itself, which is closer to sealed. A policy document shows intent, not operation.

Sample method. For held sample selections, check the population is complete and the selection is random or justified. An agent that samples only from records it can read easily may skew the test.

Wording. Findings and reports should state the condition, criteria, cause and effect clearly, without overstating.

How should the auditor decide?

An example: a closure built on a screenshot

Take a retail bank's internal audit team. Management submits evidence that an access control finding is fixed: a screenshot showing that a user group now has read-only rights. The agent drafts a closure and holds it, since closures are always held.

The auditor reads the record and sees the closure rests on one declared screenshot. They decline the closure with the reason "Remediation evidence not independently tested", and ask the agent to draft a request for a system-generated user access report covering the full population. The report shows two users still hold write access through a second group. The finding stays open. Market detail for this kind of work is in AI agent training for internal auditors in retail banking in Germany.

Which controls protect audit's role?

• Operational items routed to audit are declined and the routing fixed.
• Closures require tested evidence, not management statements alone.
• Draft findings are reviewed by an audit manager before release.
• Recurring edits are captured so the agent improves; how to teach internal auditors to bank corrections that improve an AI agent covers the habit.

Other roles handle holds as decision makers rather than assurers; compare how caseworkers handle an action an AI agent has held and how customer support agents do. The role map is in AI agent skills by role.

Where the Academy fits

Internal auditors need to read agent decisions fluently to audit them. Pryme Intelligence Academy's Foundations track covers the decision chain, depth and certification, and the record; the Operator track adds human review and reading the workspace. Auditors who assess regulated, multi-agent deployments may go on to the Architect track.