A model that writes a smooth reason for every variance is a risk, not a feature. In analytical review, the useful model is one that shows the arithmetic, says what it does not know, and asks for evidence instead of inventing a cause. The choice is less about which model is cleverest and more about which one behaves well when the data cannot explain the movement.
Last reviewed: 17 October 2026. Facts checked. We ran a search for the ICAI text of SA 520 and could not open it; only secondary study material on the international standard it is based on came back. The points on setting an expectation and a threshold before computing differences come from that material and general practice, so check them against SA 520 before quoting. No model ranking or benchmark is claimed.
What the task actually is
Analytical procedures compare recorded amounts with an expectation: last year, budget, a related figure such as revenue against receivables, or an industry ratio. A large gap needs an explanation that someone has checked. For the standard's structure, see analytical procedures under SA 520.
A model can help in three separate ways, and each has a different risk:
- Calculating the variances and ratios. Better done in a spreadsheet; a language model can slip on arithmetic.
- Spotting which lines look unusual. Useful, and the user can check it.
- Drafting commentary on the cause. This is where invented explanations appear.
The failure to test for: a confident cause with no source
Give a model a trial balance showing repairs up 62 per cent and it may write that this is "likely due to increased maintenance activity following expansion". Nothing in the data said so. A reviewer skimming the commentary may accept it, and the working paper then holds an unverified management-style explanation.
So the test is not "does the commentary read well". It is "does every stated cause trace to something in the inputs".
A 20-sample bake-off for analytical review
Rankings change, so run your own test:
1. Prepare ten to twenty prior-period files where you already know the answers. Use closed engagements, with identifiers removed. Include ones where the real cause was a reclassification, a one-off, a change in accounting policy, or an error.
2. Give the same prompt and data to each model. Ask for variances above a set threshold, with the calculation shown. Fix the threshold before running, not after seeing the output.
3. Score four things.
- Are the numbers right? Recompute every one in a spreadsheet.
- Are the lines flagged the ones a careful reviewer would flag? Note misses and false alarms.
- Does each cause cite a source in the input, or is it labelled as a question for management?
- Does it avoid explaining away something that was actually an error?
4. Add a trap. Insert a deliberate misposting into one file, such as a capital item booked to repairs. A good model flags the line as unexplained, or raises the possibility. A weak one produces a neat business reason.
5. Repeat once. If the commentary on the same data changes materially between runs, treat the output as a draft prompt for thinking, not as working-paper content.
Prompt design that helps any model
- Tell it to separate "what the data shows" from "possible causes to ask management about".
- Require it to state when information is insufficient.
- Ask for a table: line item, current, comparison, change, percentage, source of any stated cause.
- Forbid causes that are not in the supplied material. Then test whether it obeys.
Prompts only reduce the problem. The scoring is what tells you whether a particular model follows them.
Features worth comparing
| Feature | Why it matters here |
|---|---|
| Ability to work with a spreadsheet or run calculations through a code step | Takes arithmetic away from the language model |
| Handling of a full trial balance in one go | Fewer manual splits, fewer missed lines |
| Willingness to say "not enough information" | Fewer invented causes |
| Consistency across runs | Reviewable, repeatable work |
| Where the file is processed and kept | Client data confidentiality |
Prices and limits change often, so check current terms before choosing, rather than relying on a table written today.
Keep the judgement human
SA 520 expects the auditor to set an expectation and decide how large a difference can be accepted before investigating. A model can suggest, but the expectation, the threshold and the conclusion belong in the auditor's file. Record who reviewed the commentary and what evidence supports each cause, such as an invoice, a board note or a management reply. For the checkpoint design, see human-in-the-loop review.
A practical rule
Use a model to flag and to draft questions. Use evidence to write answers. If a cause appears in the commentary without a document behind it, it is not ready for the working paper.
Frequently Asked Questions
Can AI replace analytical review by the auditor?
No. It can speed up the calculation and the first read, but the expectation and the conclusion are the auditor's.
Which model is best for variance analysis?
No fixed answer is reliable, because models change. Test two or three on your own closed files and score them as above.
Should variance commentary go into the working paper as drafted?
Only after each cause is tied to evidence and a reviewer has signed off.
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