AI in accounting: what it should and should not do
AI is very good at reading a crumpled receipt and very bad at being the final authority on your balance sheet. A serious system uses it for the first and never for the second.
The useful question is not whether AI belongs in accounting. It is which parts of the job are judgement under uncertainty — where a fast, reviewable guess beats a blank field — and which parts are arithmetic and rules, where a guess is unacceptable.
Where it earns its place
- Extraction: reading dates, totals, tax and counterparties off PDFs, photos and spreadsheets.
- Coding suggestions: proposing which accounts a transaction belongs to, learning from what you confirmed before.
- Matching: spotting that 'SQ *BLUE CAF' and 'Blue Cafe Ltd' are the same counterparty.
- Explanation: turning a set of statements into plain English about margin, runway and concentration risk.
- Navigation: answering 'what did we spend on subcontractors in Q2' without you building a report.
Where it must not go
- Posting to the ledger without a human confirming the draft.
- Computing balances. Totals come from the database, never from a language model.
- Overriding a control or clearing an exception because it was asked to.
- Editing history. Corrections are reversals, and reversals leave a trail with a name and a timestamp.
Proposal, then confirmation
The pattern that makes AI safe in a financial system is simple: it proposes, a person confirms, and the audit log records both. You get the speed of automation and the accountability of a manual system, because every posted number can be traced back to a document, a proposal and a human decision.
An assistant that can explain your accounts is valuable. One that can change them without asking is a liability.
Ask it the awkward questions
The best use of a finance assistant is interrogation, not automation: why did gross margin drop two points, which customer is now a third of revenue, what is unsupported in this month's postings. Those are the questions owners actually have, and they are the ones traditional software makes hardest to answer.
