What to check before AI touches accounts payable

If you own payables, someone will soon ask you to approve an AI tool, or you will find out one is already in use. You do not need to understand how the model works to judge it. You need to know what stands between its answer and your ledger. These are the five questions we would ask, and we think they are fair to ask of us too.

Does it see the record, or a paste?

An answer about an invoice is as good as the invoice the tool actually read. Ask where the record comes from. A tool that works from whatever a person pasted into a window will miss the hold, the credit note, or the second invoice with the same vendor reference. A tool that picks the exact record from your own data starts from the right place.

Can you compare its answer with the books?

A paragraph cannot be checked. A list of facts can. Ask whether the tool produces its answer in a form that can be compared with your records: this amount, this date, this status, from this source, proposing this action. If every answer is prose, every check is a person rereading it, and that is the work you were trying to remove.

What checks the answer before anything is written?

This is the question that matters most. Some tools ask the model to check itself, which costs more and fails in the same ways the draft did. Others check nothing. Ask for a check that does not depend on the model: written rules that read the facts and your records, run the same way every time, and can be read by your team. A rule such as no payment on an on-hold invoice should not be a matter of the model remembering.

Who approves, and is it a name?

Ask what happens after a pass. A tool that writes to your accounting system because it judged itself correct has made itself the approver. The safer pattern is that a pass waits for a named person, and that a fail arrives with the reason attached, so the person is reviewing an exception and not redoing the work.

Is there a log an auditor can follow?

Ask to see the record of one item from start to finish: the question, the record used, the draft, each rule result, who approved, and what was written. If that trail exists, month-end questions answer themselves. If it does not, you are trusting a tool that cannot show its work.

How we answer these

Augustine is built around these five questions. It picks the record, takes the model's answer as a checklist of facts, checks that checklist against your books with written rules, waits for a named approver before any change, and logs every step. On 510 practice finance questions in September 2026, the same AI model went from about 1 in 4 right (24%) on its own to more than 8 in 10 right (85%) inside Augustine, and the times it went to make a payment or change a record on a wrong answer fell from 112 to 37. That is practice data, the fall is sharp and not to zero, and no real company's books have run through Augustine yet.

Ask these questions of any tool, including ours. If you would like to ask them in person, book a demo.