Invoice OCR is discussed as if accuracy were a single number, and it is not. Extraction behaves completely differently on a total printed in a standard place than on a line description that has to be reconciled with what you called the same item on your purchase order. An honest picture of invoice OCR is a picture by field, and it tells you exactly where the human review has to stay.
Reliable: totals, dates, invoice numbers
These appear in predictable positions in predictable formats, and modern extraction handles them well enough that checking each one is usually wasted effort. Supplier identity is similarly strong when the supplier already exists in your system, because the software is matching against a known list rather than reading freely. Spot-check these rather than reviewing them line by line.
Unreliable: line items and anything inferred
Line descriptions vary between what you ordered and what the supplier calls it. Units of measure differ. One ordered line can arrive as several invoiced lines. Beyond that, tax treatment and account coding are not on the document at all: they are judgements about the transaction. Extraction cannot read something that is not there, however good it is.
What to do with the difference
Design the review around the unreliable fields only. That is where the remaining human time should go, and it is a much smaller task than checking everything. It also means the sensible metric for a trial is not overall accuracy but how much review time you still owe after it, measured on your own documents.
Questions people ask about invoice ocr
What accuracy should we expect from invoice OCR?
Ask by field and test on your own invoices. Aggregate figures average easy fields with hard ones and are not comparable between products.
Does invoice OCR do the three-way match?
No. Extraction reads one document. The match compares three, and it is a separate capability that a product may or may not have.
Is it worth it at low volume?
The labour case scales with volume, so at low volume the argument rests on the record and on fewer transcription errors. Do the arithmetic rather than assuming.