Paperwork arrives and somebody types it in again.
An invoice comes in as a PDF. Somebody opens it, reads it, and keys the numbers into the spreadsheet. The same document, typed twice, every week, by the person you hired to do something better.
I have not shipped an invoice-OCR system. If that is the whole job, there are specialists further down the page rank who have, and you should probably talk to one of them.
What I have shipped, on my own operations, is the same shape at scale: unstructured stuff in, structured data out, with a gate that admits what it couldn't read. That's below, plus a document prototype you can open.
What I've built
Bank transactions categorised locally
Across 4 legal entities, every transaction is read and categorised by a model running on my own machine. No cloud fallback, deliberately: bank data doesn't leave the building. That rule is written into the code, so nobody has to remember it. The build that survives is the one that knows what it couldn't read and queues it for a person, and that gate is the part I'd price first.
- The paperwork inbox Prototype Invented mail, labelled as invented on the screen. The screen worth your time is the document it couldn't read: below the confidence threshold it flags the page for a human, and no guess gets filed. /demos/mail/ →
First piece is fixed price, typically £750 to £1,500. Then we extend or stop. How pricing works →