Document intelligence
Document intelligence is the step after text extraction: deciding what the extracted text means — which number is the total, which date is the due date, whether a required field is missing. Character recognition has been reliable for years; this second step is where the errors that reach accounting and legal actually originate, and where products differ from one another.
Why extraction is the easy half
Modern OCR on a clean source is above 99 per cent at the character level. On faxed copies, stamps over text, handwritten additions and poor scans the rate drops sharply — and, importantly, without announcing that it has.
What extraction never provides is assignment. A commercial document contains a dozen numbers; which one is the net amount, which the customer reference, which the order number follows from position, surroundings and plausibility. That inference is where a model now sits, and where the errors that matter are made.
The metric that actually predicts usefulness
Not the share correctly read, but the share correctly flagged as uncertain. A misread letter in a supplier name is noticed immediately. A transposed digit in an amount may surface at reconciliation, or not at all.
A system that routes doubt to a review queue rather than guessing is more valuable at 80 per cent throughput than one at 95 per cent with silent errors. Vendors quote the first number; the second is the one to ask for, and the willingness to answer is itself informative.
What the tax rules require, and what they do not
They do not require recognition. They require completeness of the mandatory particulars — in Austria under § 11 UStG, in Germany under § 14 UStG. A document is not bookable because the figures were read but because name, address, VAT identification, date, sequential number, description, consideration and tax amount are present.
Retention is separate again: seven years under § 132 BAO in Austria, eight years for accounting vouchers under § 147 AO in Germany since 2025. It covers the processing logs as well as the original, which turns "what did the system extract, and when" into a record-keeping question rather than a technical one.
What it is not to be confused with
OCR
OCR reads characters; document intelligence assigns meaning to them. Vendors who market both as "OCR" obscure precisely the step where quality is decided.
Enterprise search
Search finds the document. Document intelligence turns its contents into fields a system can act on. The first ends at a link; the second ends at a booking, a validation or a flag.
Frequently asked
How much manual review remains realistically?+
It depends far less on the product than on the incoming material. With consistently electronic invoices the share of untouched documents can be very high; with mixed post containing scans and phone photographs a double-digit percentage stays in review. Anyone quoting a figure without looking at your inbox is quoting a figure from someone else.
Does it work with handwriting?+
Technically a separate problem with materially lower rates. Organisations with substantial handwritten input — timesheets, delivery notes with annotations — should evaluate that separately rather than under one blended number, because the blended number will disappoint in exactly the places it was supposed to help.
Do we still have to keep the paper?+
The retention duty attaches to the document, not the medium. Replacing scanning is permitted but requires a documented procedure: how scanning is done, who checks, how immutability is ensured. That description is what gets asked for in an audit and what is most often missing.
How is this different from a template-based extractor?+
Templates work well and break at the edges: a new supplier layout, a moved field, a second page. A model generalises across layouts and is correspondingly less predictable. The practical answer is usually both — templates where volume per layout is high, model where the long tail is.
Related terms
The test with the worst document
Not the clean sample invoice — the worst copy from last month. What a system does with it, guess or flag, tells you more than any accuracy figure in a proposal.
Request a test