OCR can't read a produce invoice.
AI can.

Pull a produce invoice out of the pile at any restaurant group and look at it honestly. There's a printed table, sure. There's also a hand-written credit in the margin, a substitution the driver noted at the door, a scrawled initial next to it, and a total that no longer matches the printed one. That piece of paper is the actual input to your AP process, not the clean PDF in a software demo.

Template OCR was built for the clean PDF. It's very good at pulling a known field from a known position on a known layout. Feed it the real thing and it does one of two things: it returns the wrong number, or it returns a low confidence score and hands the invoice to a person. Either way you're paying the manual cost.

The problem was never character recognition. It's that reading an invoice requires knowing what the invoice means.

››› The invoice OCR can't handle

Multi-unit operators don't receive one kind of invoice. In a single week a 50-location group will take in broadline distributor sheets running eighty-plus lines, small-farm produce tickets written partly by hand, thermal-paper delivery slips photographed in a walk-in, linen and pest control statements, utility bills for every address, and a stack of vendor emails with the invoice pasted in the body instead of attached.

Each of those arrives in a different format, from a different channel, in different condition. A template-based system needs to be configured for every one of them, and reconfigured every time a vendor changes its billing software.

››› How template OCR actually works

Optical character recognition turns pixels into characters. That's the whole job. Everything downstream, knowing that this string is an invoice number, that this block is the remit-to address, that this column is extended price rather than unit price, comes from a template someone drew over a sample document, or from rules keyed to nearby label text.

Which is why OCR performance falls off a cliff the moment a document deviates. Shift the table down an inch and the coordinates miss. Add a second page and the mapping doesn't carry. Write something by hand and there's no character to recognize in the first place. The software isn't wrong, exactly, it's just answering a much narrower question than the one you need answered.

››› Four failures on real invoices

››› What a language model does differently

A vision-language model doesn't look for a field at coordinates. It looks at the document the way a person does and reasons about it: this number appears next to the word "total" and equals the sum of the lines above it, so it's the invoice total. This handwriting sits on the same row as a crossed-out quantity, so it's a correction to that line. This vendor always bills case quantities in the third column, so a value of 400 in that column is a weight, not a count.

That reasoning is what makes the hard invoices tractable. It also makes the easy ones cheaper, because you're not maintaining a template library to keep them working.

››› AI alone isn't the answer either

Bolting a general-purpose model onto invoice images gets you better extraction and not much else. Extraction was only ever step one. The model still has to know that this location codes produce to 5010 and not 5000, that invoices over $2,500 route to the regional director, that this vendor bills the same delivery twice about once a quarter, and that your fiscal calendar has thirteen periods rather than twelve.

That knowledge is what turns reading into processing. Without it, you've upgraded your OCR and kept your review queue — which is the same trap as assisted AP dressed up as automated AP.

What actually closes the gap is a model trained on restaurant and hospitality documents, grounded in your own coding history, and wired directly into the ERP so the result posts without a human retyping it. That combination is how invoices with handwriting on them go straight through.

››› What to ask your AP vendor

Skip the demo deck and send them your ugliest invoices; the marked-up produce ticket, the three-page broadline sheet, the photo from the walk-in. Then ask:

If the answers involve templates, confidence thresholds, and a review queue, you're buying OCR with a nicer interface. The invoices on your desk deserve something that can actually read them.