|
EXPLAINER |
AI FOR ACCOUNTANTS · PART 5 OF 16
What 'human in the loop' actually means in an audit file
An undocumented AI step is not a shortcut - it is a control gap that a regulator or peer reviewer will find.
Human in the loop is a phrase that gets used as though it means a person was nearby when the AI ran. In an audit file, it means something narrower and more demanding: a qualified reviewer looked at the output, compared it to the source, made a judgment, and left evidence that all of that happened. Proximity is not the same as review, and review without documentation is invisible to anyone who opens the file afterward.
Consider a standard bank reconciliation workpaper. An AI tool reads the general ledger export and the bank statement, matches items, and flags three unreconciled differences. That matching step is useful. But the workpaper does not capture the AI's work as a completed control - it captures whether a preparer checked each flagged item against the underlying transaction, decided it was correctly classified, and signed off. If the preparer accepted the AI output without that check and moved on, the reconciliation looks complete in the file but the review step never happened. The three differences the AI flagged might be right. Or one might be a duplicate entry the model treated as a timing difference.
The same failure mode appears in higher-stakes work. An AI layer that drafts a going-concern assessment or summarizes footnote disclosures is producing a starting point, not a conclusion. Audit standards require the practitioner to form an independent opinion. When the workpaper shows only the AI's summary and a single reviewer sign-off with no evidence that the underlying documents were examined, the sign-off is not evidence of review - it is evidence that someone clicked approve. Peer reviewers and regulators distinguish between these two things, and the distinction tends to surface during inspections.
FloQast's TakeControl 2026 conference this week featured a suite of AI accounting workflow innovations, which is a useful reminder that the tooling is moving faster than most firms' documentation standards.
The practical fix is to treat the AI output as a preparer draft, not as a completed procedure. Workpaper templates should include a field that names the AI tool used, describes what the tool did, and records what the human reviewer independently verified before signing. That last field is the one that matters. 'Reviewed AI output' is not sufficient. 'Traced flagged items to source invoices; confirmed three differences as timing items per supporting wire confirmations' is sufficient. The documentation has to show the reasoning, not just the approval.
This is not a new principle dressed up in new language. Every audit methodology already requires that automated procedures be tested and that the reviewer's judgment be documented. AI tools do not change the standard; they change how easy it is to skip a step without it looking like you skipped it. The audit file should make the human's actual work visible, and if the human's work consisted of reading an AI summary and clicking next, the file is accurately recording a control deficiency.
WORKED EXAMPLE
In practice
A senior associate is clearing an accounts-payable subledger-to-general-ledger reconciliation workpaper for the month just ended. An AI tool pre-populated the tie-out and flagged two unreconciled items totaling a material amount.
What came back. The assistant listed the vendor invoice, the receiving report, and the payment run date for the timing item, and the purchase order and both invoice numbers for the duplicate candidate. It drafted workpaper language that named the documents examined and stated a conclusion for each item. One piece of the drafted language described the duplicate as 'confirmed timing difference,' which was the AI's original label rather than a verified conclusion - the associate had not yet pulled both invoices.
How it was checked. The associate pulled both invoice PDFs from the AP system, confirmed different PO references on each, corrected the workpaper language to 'confirmed valid separate purchases per PO numbers [X] and [Y],' and initialed the correction.
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
| Reviewer sign-off on an AI output is only valid if the underlying source was independently checked. | |
| Workpaper templates need a field for what the AI did and what the human separately verified. | |
| A clean-looking file with undocumented AI steps is a finding, not a completed audit. |
SPONSORED
QUESTIONS THIS ANSWERS
What does human in the loop mean in accounting and audit work?
It means a qualified person reviewed the AI output against the source, formed an independent judgment, and documented that review in the workpaper - not simply that a person was present when the AI ran.
What should an audit workpaper include when AI was used?
At minimum: the name of the tool, a description of what it did, and a specific record of what the reviewer independently verified before signing - not just a generic approval notation.
Is clicking approve on an AI-generated reconciliation a valid review?
No. Approving AI output without tracing flagged items to source documents and recording that trace is not a completed review procedure under standard audit methodology.
SOURCES
Where this comes from
What the accounting job market is actually asking for.
GO DEEPER
Go deeper
IN THIS SERIES
Previously: How retrieval differs from training
Next: Prompt patterns for reconciliations (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