|
EXPLAINER |
AI FOR ACCOUNTANTS · PART 13 OF 16
What an AI audit trail should contain
If you cannot reconstruct exactly what the model saw and who checked it, the output is not supportable.
An AI audit trail is not a log of the fact that AI was used. It is a reconstruction kit - enough information to let a reviewer, a regulator, or a future you reassemble exactly what happened and why the output is trustworthy. That means five things, documented at the time of use: the prompt that was sent, the inputs the model worked from, the version of the model that produced the output, the output itself, and the name of the person who reviewed and approved it. Miss any one of those and the trail is broken.
The prompt matters more than most practitioners expect. Two prompts that look nearly identical can produce materially different outputs when one includes a scope qualifier and the other does not. If a staff accountant asked a model to 'summarize the reconciling items on the October bank rec' and the senior later cannot find that exact text, there is no way to know whether the model was given the right document, the right period, or the right instructions. The same logic applies to inputs: if the model was handed a PDF of the prior-month statement instead of the current one, the output is wrong, and the error will not be obvious unless the source file is captured.
Model version is the detail teams are most likely to skip, because it feels technical rather than accounting. It is not. A model that was updated between the time a workpaper was prepared and the time it is reviewed may produce a different answer to the same prompt. If the version is not recorded, you cannot rule out that out as a source of variance. Some tools stamp this automatically in their output metadata; others require the user to note it manually. Either way, it needs to end up somewhere in the file.
The reviewer field is where AI documentation converges with normal sign-off practice. Someone with appropriate competence has to read the output against the source, not just confirm that output exists. On a lease classification workpaper, that means checking the model's reading of the contract terms against the actual agreement, not just accepting that the model reached a conclusion. The reviewer's name, the date, and a brief note on what was checked complete the trail. Caseware launched its Verity-powered Risk Suggestion Agent for audit teams globally this period, which is a concrete example of AI output feeding directly into risk assessment - exactly the kind of step where a documented human review decision is non-negotiable.
The practical starting point is a one-page AI use memo appended to any workpaper where AI contributed to a conclusion. It does not need to be elaborate: a table with five rows, one for each element above, filled in at the time of use. Firms that build this habit early will find it easier to satisfy peer reviewers and clients who start asking how conclusions were reached. Firms that skip it will be reconstructing from memory, which is not documentation.
WORKED EXAMPLE
In practice
A staff accountant used an AI assistant to draft a summary of reconciling items on the October bank reconciliation workpaper. The bank statement PDF and the GL detail were both available in the client file.
What came back. The model returned a clean list of nine reconciling items with amounts and descriptions, and flagged two items over $5,000. One of the flagged items was a wire transfer the model described as 'unposted in GL' - but on review, the senior found the entry had posted two days after the statement date and was not actually a reconciling item for October. The model had read the GL export correctly but the export had been pulled with a cutoff that excluded the posting.
How it was checked. The senior traced each reconciling item on the model's list back to the bank statement line and the GL entry, ticking them off on the printed statement, and noted the corrected wire transfer treatment in the AI use memo.
A constructed example. The prompt is usable as written; the figures show the shape of a result, not a measured one.
WHEN TO USE IT
| WHEN NOT TO
|
WHAT TO TAKE FROM THIS
| Capture the exact prompt text, not a description of what you asked, at the time of use. | |
| Record the model version - some tools log it automatically; check whether yours does. | |
| Name the reviewer and note what source they checked the output against. |
SPONSORED
QUESTIONS THIS ANSWERS
What does an AI audit trail need to include?
Five things: the prompt sent to the model, the input files or data the model worked from, the model version, the output produced, and the name of the person who reviewed it against the source.
Why does the model version matter for accounting documentation?
A model updated after a workpaper was prepared may answer the same prompt differently. Without the version on record, you cannot rule out the model itself as a source of any variance found later.
Where should AI documentation live in an accounting file?
A simple AI use memo appended to the workpaper where AI contributed to a conclusion works well - a five-row table covering prompt, inputs, model version, output, and reviewer, completed at the time of use.
SOURCES
Where this comes from
What the accounting job market is actually asking for.
GO DEEPER
Go deeper
IN THIS SERIES
Previously: Agents versus automation: the difference that matters for a close
Next: Why you probably do not need to fine-tune a model (coming)
An explainer, not a study: it carries no statistics on purpose. Examples are illustrative.
How accountants are using AI, automation and smarter workflows to close faster, audit cleaner, and free up time for real work.
Accounting Stack · Audit Friendly Data · Accounting & Finance Jobs
Audit Friendly · modernaccounting.ai