Document Processing
The hours that go into paperwork, back in the week.
Invoices, contracts, delivery notes, forms, pages somebody photographed at an angle in a warehouse. Right now a person opens each one, finds four fields and types them somewhere else. We build the layer that reads them as they land, pulls out what matters, checks it against what you already know, and files it with the original attached.
- Invoices and receipts read and posted into the ledger
- Contracts turned into structured data, dates and clauses included
- Hebrew and English documents through the same pipeline
- Handwriting and bad scans flagged for review instead of guessed at
- Extracted values checked against the CRM or ERP before anything is written
- The source file kept beside every record for audit
Why this got easy recently
Classic OCR needed a template per layout, so a supplier redesigning their invoice broke it until somebody wrote new rules. Vision models read a document closer to the way a person does, which means a layout nobody has seen before usually works on the first try. That single change is what moved document work from a big-company project to something a 15-person office can afford.
Confidence, and what happens below the line
Every extracted field carries a confidence score. Above your threshold it writes straight through. Below it, the record waits in a review queue with the original image beside it so a person fixes it in seconds, and the correction feeds the next run. You set the threshold, and most clients start it high and lower it once they have watched the thing work for a month.
Checking, not only reading
Getting a number off a page is the easy half. The useful part is comparing it to what you already know: does this invoice match a purchase order, is this supplier on file, is the VAT arithmetic right, has this document come through before. Those checks are where errors actually get caught, and they are the reason the output is trustworthy enough to post without a person reading every line.
Questions we get asked
Does it handle Hebrew?
Yes, including scanned Hebrew, Hebrew and English mixed on one page, and handwritten notes in the margin. Hebrew documents go through the same pipeline as the English ones. There is no separate, weaker path for them.
How accurate is it?
On clean digital documents, high enough that review becomes spot-checking within a few weeks. On bad scans and handwriting it is lower, which is exactly why the confidence threshold exists. We measure it on your own documents during the build and show you the number before you rely on it.
Where do the documents end up?
In your systems, on your infrastructure. Files stay in your storage, records land in your ERP or CRM, and the source file is linked from the record so an auditor can see where a number came from.
What about sensitive documents?
Access is scoped to the narrowest set that makes the build work and revoked when the engagement ends. Where content cannot leave the country or the building, we will tell you early whether that rules a model out and what the alternative costs.
Tell us what's slowing your business down.
30 minutes. No pitch, no deck. Just listening to your needs and seeing how we can help.
Schedule a discovery call