Invoice OCR demos beautifully and then meets your actual suppliers. The demo used a clean, digitally-generated PDF. Your suppliers send a photograph of a printout, rotated four degrees, with a coffee ring on the total.
What actually moves accuracy
Not a better model — better constraints. If you know the supplier, you know the layout. If you know the purchase order, you know what the total should be within a tolerance. Extraction that cross-checks against data you already hold beats extraction that reads harder.
Design for the failure
Assume a percentage will be wrong, and make the wrong ones cheap to catch: show the field, the confidence, and the source region of the document side by side. A human confirming ten fields in four seconds is a system that works. A human hunting for which of forty fields is wrong is a system that gets switched off.