AI-drafted work reads well, and that is the hazard. Fluent text lowers the reviewer's guard. A review routine that starts from the source, not the draft, keeps the partner's attention where it belongs. The routine below takes about twelve minutes for a typical working paper. Scale it up for material areas.
The structural design of review points is in human-in-the-loop review. This post is the reviewer's own checklist.
What changes when AI drafts
Reviewers are used to spotting a junior's gaps: missing steps, thin reasoning, arithmetic slips. AI drafts have a different pattern:
- The structure is tidy even when the substance is thin.
- Wrong statements sound as sure as right ones.
- Citations may be invented or may point to the wrong page.
- Gaps show up as silence, not as visible blanks.
- Two similar items may be merged quietly.
So the usual cue, "this looks careless", is missing. Replace it with deliberate checks.
The twelve-minute routine
Minutes 1 to 2: read the AI-use note. See the AI-use note for working papers. What was the tool used for, what was checked, who owns the conclusion? If the note is missing, stop and send it back.
Minutes 3 to 5: pull three facts and trace them. Choose the three statements that matter most to the conclusion. Open the cited source for each. Do not pick the easy ones. If any fails, widen the sample and tell the preparer.
Minutes 6 to 7: look for what is absent. Ask what a careful person would expect to see that is not there. For a lease paper, perhaps extension options. For a related-party paper, perhaps a director's relative's firm. Silence is the main failure mode, so give it its own step.
Minutes 8 to 9: check the numbers outside the draft. Re-add one total, recompute one ratio, tie one balance to the trial balance. Numbers in AI prose should be tied to the ledger, not accepted because the arithmetic looks plausible.
Minutes 10 to 11: test the reasoning. For any judgement, ask the preparer why. Why that useful life, that materiality, that classification? The reply should come from the preparer, not from rereading the draft. If the preparer can only point to what the AI said, the paper is not ready.
Minute 12: decide and record. Sign off, return with queries or escalate. Note in the paper what you tested.
Questions that cut through
- "Show me the page this comes from."
- "What would have to be true for this to be wrong?"
- "What did you try that the model got wrong?"
- "Where did you disagree with the draft?"
A preparer who never overrode the model has probably not checked it. Some corrections are the normal sign of a real review.
Risk-based depth
Not every paper deserves twelve minutes, and some deserve more.
| Situation | Review depth |
|---|---|
| Formatting, letters, internal summaries | Light read |
| Routine extraction with sample tie-out | Standard routine |
| Material balances, estimates, going concern | Standard routine plus full source check of key items |
| Fraud risk, related parties, legal matters | Partner works from source documents directly |
Where the matter is at the heart of the opinion, the partner's own judgement, from source, cannot be delegated to a model's summary.
Reviewing the final report text
Reports and management letters drafted with AI need a different pass:
- Compare every number and date with the working papers.
- Check names, entity details and year-ends. Models carry over details from earlier prompts.
- Check that statutory references are correct and current. Do not assume a section number is right because it is stated confidently. The hallucination post covers common failure patterns.
- Make sure the tone matches the firm's, and that no wording hedges a finding that should be stated plainly.
Protect the reviewer's time
Review is a scarce resource. Make it easier:
- Ask preparers to submit the source pages next to each key statement.
- Have the preparer mark the lines they changed from the AI draft.
- Keep a short log of recurring AI errors in the firm, such as a model that keeps misreading a certain format, and share it with the team.
- Agree a standard for what the preparer must have checked before a paper reaches review.
Keep the firm's responsibilities clear
The engagement partner is responsible for the quality of the engagement, including direction, supervision and review. Using AI does not move that responsibility. Check the firm's quality management policy and the relevant Standards on Auditing on supervision and review for the exact wording, and record how AI-assisted work is covered. See also AI audit quality management.
Frequently Asked Questions
Should partners review AI output differently from staff output?
The standard is the same. The method changes because AI errors hide behind fluent text, so start from source and look for what is missing.
How many items should a reviewer trace?
Three of the most decision-relevant at minimum, and more if any fail.
Can the manager review and the partner skim?
Depth should follow risk. For high-risk areas the partner still needs direct contact with the evidence.
Last reviewed: 17 October 2026. Facts checked. No ICAI, NFRA or vendor page was opened in this session. The routine is a suggested practice and not a prescribed procedure. References to supervision and review duties are general; confirm against the current Standards on Auditing and the firm's own policy.
Related Articles
Statutory facts on this page are checked against their sources, and the page says where it relied on secondary reporting. How we verify · Report an error