Trace the conclusion to its support
That is the first question I would ask when reviewing AI-assisted audit work. A clear paragraph can hide an unsupported assumption. A confident conclusion can overlook a document that changes the answer.
AI can help organize information and flag possible inconsistencies. It can prepare a draft for review. We still need to evaluate its output against the purpose of the work and the underlying information. A convincing answer alone does not give us sufficient support for an audit conclusion.
The IAASB has published non-authoritative resources on technology overreliance and audit documentation when using automated tools. Read them alongside the applicable standards and your firm's methodology.
Start with one fictional working paper
I would start with a small training exercise. Create a fictional expense schedule and a few made-up invoices. Include an amount that does not match the schedule. Add an invoice from outside the period.
Ask an approved tool to flag differences and suggest questions for follow-up. Have someone check the same documents independently. Then compare both results against the original documents.
Record what the tool found and what it missed. Note any claim that has no support. Count the time spent checking and correcting the output when deciding whether the tool helped.
Keep this exercise separate from an actual engagement file. Use no real client information in the trial.
Before using client data in later work, confirm that your firm approves both the tool and the task. Check who can access the data and how long the provider retains it. Establish whether uploads train models. Follow the relevant confidentiality requirements. Removing a client's name alone may leave identifying details behind.
Five checks before relying on the result
1. What did the tool actually see?
Identify the exact documents and versions it could access. Check whether it read the tables and scanned text correctly. Do not assume it processed every page just because the upload succeeded.
2. Where did each important statement come from?
Trace figures and factual claims to the original information. Open cited references. Evaluate their relevance and reliability. A plausible document title or link does not establish either.
3. What context could change the answer?
Look for missing documents and contradictory information. Identify assumptions that need further work. Avoid repeating a prompt until the tool gives the answer you expected.
4. What did a person verify?
Record the checks performed and discrepancies resolved. Explain why the reviewer accepted or rejected the output. A name beside a review box does not explain the judgment exercised.
5. Could another reviewer follow the work?
Retain the relevant inputs and prompts under your firm's documentation policies. Keep the outputs and review decisions too. Show how the work connects the source information to the conclusion.
Adapt this exercise to your firm's methodology and applicable requirements.
For me, the useful measure is whether the team can explain the work better after using AI. That includes explaining where the tool failed and what still needs checking.
Start with one working paper. Learn from the gaps before expanding its use.
Book an OLi Audit demo for your firm.
Original practical guidance prepared for OLi with AI assistance. It focuses on workflow and product evaluation. It does not replace your organization’s policies or professional advice.
