Artificial intelligence in accounts payable is real and its useful range is narrower than the marketing suggests. It is strong where an answer exists on a document or in your own history, and weak where the answer is a judgement about the transaction that nobody has written down anywhere. That division predicts fairly well which claims to believe.
Where it genuinely helps
Reading invoices without templates, including layouts it has not seen. Identifying which supplier sent something by matching several signals rather than a name. Suggesting a coding from what similar invoices were coded to before. Flagging an invoice that looks unlike this supplier's usual pattern. All of these draw on something already recorded.
Where it is a suggestion at best
Account coding, because the correct account is a judgement about your organisation rather than a fact about the document. Deciding whether a price difference is acceptable, which depends on a conversation somebody had. Whether an unfamiliar supplier is legitimate. Treating suggestions in these areas as answers propagates past mistakes confidently.
What to insist on regardless
That it says how sure it is, and that low confidence reaches a person on a screen showing the document. And that whatever it did is recorded, so an auditor can see what was decided automatically and on what basis. A confident system with no confidence signal and no record is the shape that produces expensive surprises.
Questions people ask about artificial intelligence in accounts payable
Can it replace the review step?
It can shrink it by being right more often on the easy fields. Removing review entirely means accepting whatever it produces, including when it is wrong.
Does it detect fraud?
It can flag anomalies, which is useful. The controls that actually stop the largest losses are procedural, particularly verifying bank detail changes.
What should we ask a vendor?
How confidence is exposed, what a correction changes, and what record is kept of automated decisions.