FREEAI
All notes

Document intelligence

Rules and models are not rivals

Why a vision model alone fails on real documents, and where the deterministic layer has to sit for the numbers to hold.

26 Aug 2026 · 6 min read

Rules and models are not rivals

A vision model reads a clean passport page almost perfectly. Then it meets the documents an operations team actually receives: a photograph taken at an angle under a fluorescent tube, a scan with a staple through the date field, two documents on one page, a form that was revised last year and looks nothing like the one in the prompt.

The failures are structured, not random

This is the useful part. Model errors on real documents cluster: the same field fails for the same physical reason. Glare on a laminated surface. A digit that is ambiguous in one typeface. A date order that is unmarked and could be read two ways. Once you know the clusters, you can decide what to do about each of them, which is a very different job from trying to raise a single accuracy number.

Where the deterministic layer sits

  • Format and checksum validation on anything that has structure — document numbers, tax numbers, bank account formats. A model does not need to be right here, because arithmetic can check it.
  • Cross-field consistency inside one document. A date of expiry before a date of issue is not a close call.
  • Reconciliation against what the business already knows. The strongest check is usually not inside the document at all — it is the record the organisation already holds about that person or that transaction.
  • A confidence threshold, and a review queue for whatever falls below it, with the source image next to the extracted value so a person can resolve it in seconds rather than reopening the file.

Why accuracy alone is the wrong bar

Two systems can report the same accuracy and be worth completely different amounts. The one that is confidently wrong on a small number of cases and gives no signal is dangerous. The one that is uncertain on a slightly larger number of cases and says so is useful, because the uncertain cases can be routed to a person and the rest can be trusted.

So the number we agree with a client before starting is not a single accuracy figure. It is a pair: what proportion is handled without a person, and what is the error rate on the part that was handled automatically. Those two together describe the system honestly. Either one alone can be made to look good.

If a document AI proposal quotes one accuracy percentage and does not describe the review path, the review path does not exist.

Working on something similar?

Describe what happens today in a few sentences. We will tell you which part a model can carry, and which part it cannot.

Send a request