Document extraction

OCR turns the file into text lines, then a language model interprets them the way a bookkeeper would. What comes back is vendor, subtotal, tax, total, date, invoice number and individual line items as discrete values you can act on.

The distinction matters. A blob of extracted text still leaves the whole job to do. Discrete fields can be scored, matched, checked and approved one at a time.

The same pass classifies the document — a bill, a receipt, an invoice or a sales receipt — so it is filed correctly without you sorting the pile first.

Formats do not need to be consistent. A crumpled receipt photographed on a phone and a clean supplier PDF are both fine.

Every document is checked against what you have already filed, so the same invoice arriving by email and again as a scan does not become two bills.

This is one of the errors that is cheapest to prevent and most expensive to find later, because a duplicate usually surfaces at reconciliation rather than at entry.

The party named on the document is matched to a vendor or customer you already have — including when the document carries a trading name, an abbreviation, or a slight OCR misspelling of the name you filed them under.

Without this, a supplier quietly becomes three supplier records and your reports stop adding up the way you expect.