What invoice OCR accuracy actually means

A 99% character accuracy rate sounds excellent until you realise that a single wrong digit in a total makes the whole invoice unusable. Accuracy is only meaningful once you know what was counted, and almost every published figure counts the easiest thing.

Character accuracy versus field accuracy

Character accuracy counts individual glyphs. Field accuracy asks a stricter question: is the extracted total exactly equal to the total printed on the page? For accounts payable only field accuracy matters, and it is always the lower of the two numbers.

When you compare tools, insist on field-level figures for the fields you actually post to your ledger — vendor, invoice number, date, tax and total. A tool that will not give you those has told you something.

Why confidence scores matter more than averages

A model that is right 92% of the time and knows when it is unsure beats a model that is right 96% of the time and is confidently wrong. Confidence turns extraction into a review queue: you check the flagged minority and trust the rest.

• Above 0.95 — accept without review in most workflows.
• 0.85 to 0.95 — spot-check on high-value invoices.
• Below 0.85 — always review before posting.

Arithmetic is a better test than confidence

The strongest evidence that an extraction is correct is not the model's own opinion of itself. It is whether the numbers reconcile: line items summing to the subtotal, subtotal plus tax equalling the stated total. A parse that balances is almost certainly right, whatever the confidence score says.

Measure on your own documents

Take twenty invoices that represent your real mix of suppliers, formats and languages, and run them through any tool you are evaluating. Count the fields you would have had to correct. Twenty documents from your own inbox tell you more than any published benchmark, because they contain your suppliers' templates rather than someone else's.