Documentation for audit — keeping yourself defensible

The principle

There is one principle that holds across every workpaper you will ever produce, and AI does not change it: a reviewer should be able to look at the file and see, for any non-trivial figure, where it came from and who decided to accept it. When the work was done by hand, the trace was a ticking pencil. When it was done in Excel, the trace was the formula. When part of the work is done by AI, the trace has to capture three new things: what input went in, what came back, and what a human decided to do with it.

This is not a paperwork exercise to satisfy ICAN. It is the thing that protects you when, in Falgun, the IRD inspector asks why a particular deduction was treated the way it was, or when the engagement partner asks the junior who actually authored the position in a memo with their name at the bottom. If the answer is “the model said so and I trusted it,” the file is indefensible, and so are you.

A practical workpaper note format

You do not need a separate AI logging system. You need three to five lines on every workpaper where AI touched the analysis. The format below is short enough that staff will actually use it, and structured enough that a reviewer can read it in fifteen seconds.

Source data: what you fed the model. (“Client trial balance Magh 2082, sales ledger Jan–Mar, VAT working file.”) Prompt summary: one line on what you asked. (“Asked for reconciliation memo tying sales ledger total to D-03 line 1.”) Model + date: which assistant, which day. (“GPT-class chatbot, 2026-06-14.”) Output kept: which parts of the output ended up in the workpaper, and which parts you rejected. (“Used the reconciliation narrative; rejected the proposed treatment of three credit notes as the model misidentified the period.”) Reviewed by: the human whose name is on the work. (“Reviewed by S. Sharma, senior, 2026-06-14.”)

Five lines. Not a screenshot — screenshots age badly, take up space, and are hard to search. A structured note that a reviewer can read at the top of the workpaper and immediately understand what role the model played and what role the human played.

Things that look harmless but are not

There are three habits that develop quickly in any firm that starts using AI seriously, and each one will eventually cause trouble.

1. Pasting client data into a public chatbot. The PAN, the trial balance, the TDS schedule, the bank statement — none of these belong in a free-tier consumer chatbot. The data leaves your control the moment you paste it. The terms of service of most consumer products allow the provider to use submitted content for training or evaluation. Even where the provider is well-behaved, you have transferred client confidential information to a third party without consent. ICAN’s confidentiality requirements do not include a carve-out for “I was just trying to save time.” Use an enterprise tier with a no-training agreement, or an on-premise assistant, or redact the data before pasting.

2. Accepting the model’s specific citation without verification. This was covered in the VAT and TDS section, and it appears here because it is the single most common cause of a defective workpaper. The model produces “Section 88(1) of the Income Tax Act 2058” with confidence. The senior pastes it into the memo. The reviewer assumes it has been checked. The IRD inspector pulls the Act and the section says something different. The fix is procedural: any citation that appears in the file has a verifier’s initials next to it, and the verifier’s job is to have opened the source.

3. Using AI to “round” numbers in a way the books do not support. A subtler one. The model is asked to produce a summary table from a trial balance and quietly rounds figures in a way that makes the table tidier but no longer ties to the source. The summary looks professional and is wrong. Any number in a workpaper has to reconcile to the books. The model is allowed to round in narrative prose, but the schedules and reconciliations must hold.

A small-firm template policy

You do not need a thirty-page AI policy. For a small or mid-size practice in Nepal, three rules are sufficient, and they should be on a single page that every staff member signs at induction.

1. Client data only goes into approved tools. The firm names the assistants that may be used with client data — typically an enterprise-tier product with a no-training contract — and prohibits all others. Free consumer chatbots may be used for general questions (“explain how a particular VAT mechanism works in principle”) but never with client-identifying data.

2. Every AI-touched workpaper has a five-line note. Source data, prompt summary, model and date, what was kept, who reviewed. The note is in the file before the workpaper is signed off. No exceptions, including for the partner.

3. Every citation, every rate, every legal position the model produces is verified at source by a named human before it leaves the firm. The model can draft; the human signs. The verifier’s initials appear next to the citation in the workpaper.

These three rules are not novel. They are the existing professional standards — confidentiality, working papers, supervision — applied to a new tool. The reason to write them down is that AI is fast enough to outrun habit, and a written rule survives a busy week.

Check your understanding

Quick check

—

When AI has been used in preparing a tax workpaper, what is the single most important item to keep in the file for an ICAN review or IRD inspection?

1. AA full screenshot of the chat conversation, so the reviewer can see exactly what was said.
2. BA note at the top of the workpaper summarising the source data, the prompt, the model and date, what output was kept, and who reviewed it.
3. CA copy of the AI provider’s terms of service to prove the tool was used legitimately.
4. DNo record at all, since the final figures speak for themselves.

Check answer→

What comes next

The compliance chapter ends here, but the demands on your work do not. Tax positions live or die on whether you can communicate them — to the client whose money is at stake, to the partner who signs the engagement letter, to the regulator who reads your submission. The next chapter turns to client communication: drafting clear memos, explaining complex positions in plain Nepali and plain English, and using AI to widen the bottleneck without losing the firm’s voice.