Are internal auditors accountable for an AI agent’s mistakes?

Not for the operation's mistakes, but yes for their own work. Internal auditors should not be approving an agent's operational actions at all, because that would compromise their independence. Where an agent prepares audit work, such as samples, tests and draft findings, the auditor who signs the work owns its scope and conclusions. The record shows what they relied on.

Where do internal auditors sit in relation to agents?

Auditors meet agents in two places. First, as the subject of audit: the business uses agents in finance, claims or customer service, and audit tests whether the controls around them work. Second, as a tool: the audit team may use an agent to pull samples, run tests across full populations and draft findings.

In the first case, the auditor is not in the decision chain and should stay out of it. An auditor who approves held actions in the operation they audit can no longer give an independent opinion on it. In the second case, the auditor is in the chain for their own work, and accountable for it.

What may an audit agent do, and what stays with the auditor?

• The agent may: extract populations from systems, select samples by the auditor's method, test every transaction against a rule, compare control evidence across periods, and draft findings with references.
• The agent holds: any finding rated high, and any test where the population looks incomplete.
• The auditor keeps: the scope, the method, the judgement on whether an exception is a control failure, the rating, and the conversation with management.

What does the record give an auditor?

Two things. For the operation under audit, the governed record is the best evidence there is. Every verdict the operational agent reached is filed with the bound that decided it, the sources and their provenance, and the person who approved. An auditor can reconstruct any decision without relying on interviews.

For the audit work itself, the audit agent's record shows which data it used and how complete it was. A payments population from the ledger is sealed. A spreadsheet of manual payments sent by finance is declared until reconciled. If the auditor signs a finding resting on a population the record shows was partial, the gap is theirs.

An example: the test that missed manual payments

Take a mid-sized manufacturer. Internal audit asks its agent to test all supplier payments over a quarter for segregation of duties: no one should both set up a supplier and approve a payment to them. The agent tests 2,000 payments from the ledger extract and finds no exceptions. It notes that the extract excludes payments made outside the main run.

The auditor signs a clean finding. Later, a fraud is found among the manual payments. The record shows the note about excluded payments and the sign-off. The auditor is accountable for the scope of the test. The operation is accountable for the fraud itself, and the record of its own agent shows who approved those manual payments.

What should audit teams put in place?

1. Keep auditors out of operational approvals. Grant read access to the record, never approval rights over the agents they audit.
2. Check population completeness before testing. The audit agent should hold any test where its source does not reconcile to the ledger total.
3. Train auditors to read the record. They need to understand depth, verdicts, provenance and the minimum chain to audit agents well. A two-week practicum for internal auditors, logged in a practicum log template for internal auditors, builds this on real cases.
4. Follow the agent's renewals. Operational agent certificates lapse after 90 days, and each renewal changes what audit should test; see how to prepare internal auditors for 90-day recertification.

The front-line view is different; compare collections agents and AI agent mistakes and paralegals and AI agent mistakes. The wider topic is in AI literacy and certification for people who work with agents.