Artificial intelligence in accounts payable covers a real capability and a great deal of marketing, and the distinction is worth making carefully because the consequences of getting it wrong are financial. What works well today is extraction and suggestion: reading unfamiliar invoice layouts without templates, proposing a coding based on history, and flagging an invoice that looks unlike the rest. What it should not do unsupervised is decide: approving payment, clearing a matching exception, or writing off a difference. This page separates the two, and sets out the review discipline that makes the first kind safe to use.
What it does genuinely well
Reading layouts it has not seen before, which is the main advance over template-based capture. Suggesting coding from history, which is right often enough to save time and wrong often enough to need review. And flagging outliers, which is a useful second pair of eyes on a population no person reads in full.
Where it should not decide
Approving payment, clearing exceptions and writing off differences. Each is a judgement with money attached, and a model making it silently produces exactly the failure that is hardest to detect: a decision that looks like a process.
The review discipline that makes it safe
Sample what it decided, periodically, and check you agree. Keep suggestions as suggestions a person accepts rather than as defaults that apply themselves. On the worked example on this site, 400 invoices at three touches of six minutes with a 12% exception rate is 134.4 hours a month: $4,300.80 at a $32 loaded rate, $10.75 an invoice and $51,609.60 a year. and the exception hours are precisely where judgement lives.
Questions people ask about artificial intelligence in accounts payable
Is AI coding accurate enough to trust?
It is accurate enough to accept with review and not accurate enough to apply silently, in our reading of how these features behave. Treat a suggestion as a suggestion.
Will it detect fraud?
It can flag anomalies, which is useful and is not detection. The ordinary payables schemes are caught by a person reviewing new suppliers and bank detail changes, and that check remains the highest-yield one.
Should we wait for it to improve?
The extraction is already worth having. The judgement parts are the ones where waiting costs nothing.