How do you use AI with accounting data?

Accounting is unusual among business systems in that the arithmetic is already solved and the classification is not. Totals, balances, tax calculations and reconciliations are deterministic operations that existing software performs exactly, and asking a language model to do any of them introduces error into the one place that has none. The work that is genuinely hard is deciding what a transaction is, matching it to the document it belongs to, and noticing that something looks wrong.

Coding is the strongest use. A bank line reading a supplier name and a reference is a classification problem with a defined output set — your chart of accounts — and history to learn the pattern from. Presented as a suggestion the bookkeeper accepts or changes, this removes most of the repetitive work and keeps the decision. Presented as an automatic posting, it produces a ledger nobody has read.

Matching is the second and is closely related: connecting a payment to an invoice, a receipt to a card transaction, a remittance to several outstanding items. This is a comparison task with checkable output, because the amounts either reconcile or they do not, and a validation on the arithmetic catches the errors that classification alone would miss.

Anomaly flagging is the third and the one to be careful about. Finding transactions that differ from the pattern is genuinely useful and the output is a list to look at, not a conclusion. A flagged item is not evidence of anything, and a system that describes why something looks unusual — a supplier not seen before, an amount outside the usual range, a duplicate reference — is far more useful than one that assigns a suspicion score.

The hard constraints are worth stating plainly. Financial records are the most consequential data in most small businesses and are frequently subject to retention and access obligations, so the question of what leaves the business when a system reads them is a real one and should be answered before connecting anything. And nothing should post an entry without a person releasing it, because a wrong entry is corrected by another entry and the correction is visible forever.

One genuinely useful application that is easy to miss: explaining. A business owner looking at their own accounts and asking why the figure moved, what a line means, or what the difference is between two periods is asking a question the data can answer and the software presents badly. Answering it from the actual records, with the figures traceable, is real help and carries no posting risk at all.