Document intelligence
that fails closed

Regulated files are not “unstructured data”

They are structured badly. A Certificate of Analysis has columns; they just have no borders. A vendor data book for an oil and gas handover has a contents list, data sheets, and certificates of conformity that do not share a filename convention. A DICOM governance pack cares about study instance UIDs, retention, and who could export a series — not a pretty gallery. A WCAG conformance evidence pack is a trail: pages tested, criteria (1.4.3 Contrast, 2.4.1 Bypass Blocks), artefacts, residual risk.

Generic “document AI” products optimise for a paragraph summary. That is the wrong artefact. You need rows that survive an audit.

How we build the extractor

Lock a schema with the people who sign the report. Collect a gold set. Try the cheap parsers; they usually fail on borderless tables — we documented that on SI Report Analyser. Then cluster text positions, dual-parse, and emit exceptions. Humans resolve exceptions. The model, if we use one at all, proposes a mapping; it does not write a concentration.

The same pattern applies to funder data import sheets: the destination columns are the schema. Extraction is successful only if the sheet validates.

What a build involves

• One document family (lab PDFs, handover, DICOM logs, WCAG audits, Ofsted safeguarding files).
• A typed schema and an exception queue.
• Role-based access and an audit log of who accepted a cell.
• An export the downstream system already understands.

See also geotechnical CoA screening and funder impact reporting.