OCR in accounts payable reads a supplier invoice and turns the picture into fields. It is genuinely useful and it is sold with a number, usually an accuracy percentage, that tells you almost nothing about how it will behave on your documents. Accuracy depends on your supplier mix: a run of clean PDF invoices from ten regular suppliers reads almost perfectly, and a pile of scanned photographs from thirty occasional ones does not. What separates a tool worth having from one that adds a step is not the headline figure but what happens to the invoices it is unsure about. This page is what OCR reads reliably, what it misses, and how to judge the exception path before buying.
What it reads reliably
The header: supplier, invoice number, date, net, tax and total. These are consistently placed, numerically checkable against each other, and the part of the job that is pure transcription. Removing them from a person is the clear win and it is available from almost any tool.
What it misses, and why
Line detail on unusual layouts, handwritten annotations, multi-page invoices where totals appear once, and anything on a poor scan. None of that is a scandal; it is the nature of reading a picture. The question is whether the tool knows it is unsure.
Judge the exception path, not the accuracy claim
A tool that presents a low-confidence field for correction is safe. One that guesses silently is worse than typing, because a wrong number that looks confident gets posted. Ask to see the interface for a bad scan before you ask for the accuracy figure.
Questions people ask about ocr accounts payable
What accuracy should we expect?
Not a number we will give you, because it depends on documents we have not seen. Run a sample of your own worst invoices through any trial and count. That count is the only relevant figure.
Does OCR remove the need to check invoices?
No. It removes typing. Matching and approval are what check the invoice against what was ordered and received, and no capture tool substitutes for either.
Is OCR worth it at low volume?
Often not on its own. 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 and $10.75 an invoice. Typing is a minority of that, so at small volumes the saving is real but modest, and the case usually rests on what capture enables downstream rather than on the typing.