Intelligence / Document intelligence
The document becomes data, then a person checks it
Emails, ACORD forms, letters and scans read into structured fields with confidence recorded per field — for intake and for the documents that arrive later.
Extraction confidence is the whole product. A system that hides how sure it was forces a reviewer to check everything or nothing.
Test it on a bad scanWhy show confidence per field instead of one score?
Because a single overall score tells a reviewer nothing actionable. Field-level confidence directs attention: the policy number read cleanly, the loss date was ambiguous. Review effort lands where the uncertainty actually is, which is what makes review fast enough to be real.
Fields, confidence, provenance
What extraction produces
- 01
It handles what actually arrives
Structured ACORD forms, prose letters, forwarded email chains, photographed documents. The realistic input is messy, so messy input is the design case.
- 02
Fields are extracted with confidence
Parties, dates, loss description, policy references and claimant count — each with its own confidence rather than a document-level average.
- 03
Provenance is kept
The extracted value stays connected to the document it came from, so a reviewer can check the source rather than trusting the field.
- 04
Review is the gate
For intake, a reviewer promotes the draft. For later documents, extraction populates and a person confirms before it drives anything.
Where it earns its cost
Documents that arrive after intake too
Intake is the obvious case, but a claim accumulates documents for months and each one is re-keyed by somebody today.
ACORD FNOL form
Structured fields into a draft intake
Police or incident report
Parties, dates, narrative facts for review
Carrier acknowledgment
References matched to the existing claim
Medical or repair estimate
Amounts and dates surfaced for the adjuster to confirm
Multi-claimant document
Claimant count flagged — the split decision stays with the reviewer
Entitlement
Document intelligence, off until bought
It reads. It does not understand.
Extraction is pattern recognition over text. Treating it as comprehension is how automated intake produces confidently wrong files.
No judgment about meaning
It extracts what a document says. Whether what it says is true, complete or consistent with the policy is a person’s assessment.
Poor input, poor output
A faxed photocopy yields low-confidence fields. The correct outcome is more review, and the confidence display is how we make that visible rather than hiding it.
It cannot create or split
No claim is created and no file is split by extraction. Both are reviewer decisions, structurally.