Teach an accounting AI when to ask a question

A practical design for using transaction suggestions without turning an ambiguous bank description into a false accounting fact.

A bank feed arrives with a short description, an amount and a date. It rarely arrives with the business story. A payment to a familiar supplier might buy stock this month and equipment next month. The useful job for an accounting AI is to suggest what deserves attention, then make that attention easier to give.

There are real implementations behind this idea. Intuit announced bookkeeping agents in July 2025. Researchers describe Rel-Cat as a graph approach to transaction categorisation. BIS identifies data quality and model oversight as financial AI concerns. [QuickBooks rollout](https://investors.intuit.com/news-events/press-releases/detail/1258/intuit-introduces-ground-breaking-virtual-team-of-ai-agents-to-fuel-growth-for-businesses), [research](https://arxiv.org/abs/2506.09234), [risk context](https://www.bis.org/publications/fsi-summary-financial-stability-implications-artificial-intelligence-executive-summary).

Consider an invented studio called Cedar Print. Its bookkeeper receives three payments to Paper Lane: 420 for paper, 900 for a printer repair and 2,800 for a new cutter. A merchant-only classifier could treat all three alike. A more useful review screen puts the receipt, proposed category and reason beside each row. The model can suggest supplies for the first entry, request the repair invoice for the second and ask an accountant to review the third. These are example decisions, not tax guidance; the correct treatment depends on the business and its accounting rules.

Before introducing the assistant, Cedar writes a small decision sheet. It lists the categories already used, the evidence needed for each, and which decisions require an accountant. Employees can add a comment without changing the ledger. The assistant sees only the fields approved for this purpose. Personal purchases, payroll records and unnecessary account identifiers are excluded from an initial experiment. That reduces both the exposure of sensitive information and the number of unrelated decisions the test has to explain.

The first run produces suggestions in a separate worksheet. Nothing posts automatically. For each item, the bookkeeper records accepted, corrected or unresolved. A correction also records why: missing receipt, wrong supplier match, duplicate import or unsuitable category. Without those reasons, a high acceptance rate can hide a reviewer who simply stopped checking. Ten carefully explained corrections can teach the team more than a thousand rows marked approved.

Now test an uncomfortable case. The feed contains a 900 payment and a matching invoice for 900, but another employee already entered the bill. The AI should help locate the existing record. It should not create a second expense because the amounts agree. Cedar therefore includes duplicates, split payments, refunds and transfers between its own accounts in the test set. These examples examine the relationship between records, not just the fluency of a sentence.

A second problem is misplaced confidence. An explanation such as regular office supplies may sound convincing while adding nothing beyond the supplier name. Require the proposed evidence link and allow an explicit unknown result. When the receipt is illegible, the interface should ask for a clearer copy and preserve the draft. A model score can help order the queue, but it should not be presented to staff as a calibrated probability unless that interpretation has been demonstrated.

Measure corrected postings per reviewed batch, time spent finding evidence and unresolved items carried into close. Keep the original feed untouched so an accountant can reconstruct the decision. Repeat the test on a later period: an assistant that memorised familiar suppliers may struggle with a new branch or a new purchasing pattern. The meaningful saving is a cleaner review with fewer preventable mistakes, rather than a larger number of suggestions.

Budget for the accounting subscription, any separate AI entitlement, receipt storage and review time. Intuit's announcement described a US rollout with feature differences by product; it does not establish today's availability for every country or plan. Cedar should check its own account before buying around a promised feature. An ordinary export-and-review process remains a sensible starting point.

The experiment earns a wider role only after the team can answer three questions: what evidence supported this posting, who approved it and how can it be corrected? An accounting AI becomes helpful when it makes those answers easier to obtain. A transaction description remains a clue until someone establishes the underlying fact.