What OCR still cannot do
Every OCR vendor shows you the same clean invoice. Here is the honest list of what fails, so you can plan around it instead of being surprised in month-end.
Handwriting
Printed text is a solved problem; handwriting is not. Even the best models struggle with joined-up writing, and small models fail completely. If your documents are handwritten — delivery notes, annotated invoices, prescriptions — expect manual entry and budget for it.
Faint thermal paper
Supermarket receipts fade. A receipt from eight months ago may be grey on grey, and no amount of processing recovers what is no longer there. Photograph receipts the day you get them, not at the end of the quarter.
Merged cells and multi-column layouts
OCR gives you text with coordinates. Deciding that two fragments belong to the same table cell is a layout problem, not a recognition problem — and it is where most automated extraction quietly produces wrong numbers rather than failing visibly. A tool that outputs a plausible-but-wrong table is more dangerous than one that outputs raw text.
Stamps, signatures and stickers over the numbers
If a "PAID" stamp lands on the total, the total is gone. Same for a signature crossing a date field. This is common on invoices and there is no software fix.
Non-Latin scripts and mixed languages
Accuracy drops sharply when a page mixes scripts, or when the model was not trained on the language. English and Chinese are well covered; many others are not. Check before you commit to a workflow.
Tables that span pages
A table continuing onto the next page loses its header, and every extractor will treat the continuation as a new table with unknown columns. The practical fix is to process page by page and repair the joined rows afterwards — which is manual work no vendor advertises.
What this means for your process
The useful framing: OCR converts a typing problem into a checking problem. That is a good trade — checking a table is faster than typing one — but it is not the same as "done".
Design your month-end around a two-minute check: sort by amount, look for outliers, confirm the total against the document's own summary. Two minutes catches almost everything, as long as you actually do it.
How to evaluate any tool in ten minutes
- Feed it your five worst documents, not the clean one in the demo.
- Compare the totals digit by digit.
- Look at what it does when it fails — does it say "low confidence" or silently guess?
- If it uploads, ask where the file goes and read the answer twice.
The last one matters most. A tool that fails visibly can be worked with. A tool that fails invisibly cannot.
Common questions
Is handwriting OCR getting better?
Yes, slowly, with large models. It is still unreliable for business use, and browser-side models are far from it.
Why do clean scans still produce wrong numbers?
Usually layout, not recognition — the digit was read correctly but assigned to the wrong column or row.
Should I use OCR or the PDF text layer?
Text layer when it exists: it is exact and instant. OCR only for photos and scans.