OCR products for accounts payable have converged enough that reading a clean invoice tells you nothing about which to buy. Three tests, run on your own documents in a fortnight, produce answers that differ sharply between products, and all three concern the cases where an evaluation using vendor samples produces no useful information at all.
Test one: your twenty worst invoices
Pick the documents your team complains about rather than a representative sample, because representative samples are dominated by easy cases you already handle fine. What you need to know is behaviour at the edge, since the edge is where the human time goes and therefore where any saving has to come from.
Test two: what happens when it is unsure
Ask to see the review queue. Software that flags low confidence and puts it in front of a person is doing the right thing. Software that silently fills a best guess is generating work you will not discover until a payment is wrong, and confidence handling is the biggest difference between products that look identical in a list.
Test three: line items on a multi-line invoice
Take an invoice with a dozen lines, several of them partial deliveries against one ordered line, and see what comes out. Header fields are solved and everybody does them well. Lines are what the three-way match needs, and a blank or garbled line grid tells you the product automates keying rather than matching.
And check the input before blaming the engine
Consistent resolution, capture at arrival, and emailed PDFs rather than photographs improve results more than most product differences. If your corrections cluster on poor images rather than on particular fields, the constraint is upstream and changing engines will not move it.
Questions people ask about best ocr software for accounts payable
How long should a trial run?
A fortnight of real documents, including the difficult suppliers. Longer trials mostly add feature comparison, which does not discriminate between products in this category.
What should we measure?
The human minutes still required per invoice afterwards, not the headline accuracy. That number multiplied by your volume is the honest comparison against today.
Does it help with exceptions?
No. Exceptions come from documents disagreeing or being absent, which is a different problem from documents being read, and OCR does not touch it.