Claims / Agentic intake
Extraction proposes. A person disposes.
Documents arriving by email or upload are read, the loss facts extracted, and a draft intake presented with confidence shown per field.
A draft is not a claim. Nothing is assigned, no clock starts and no reserve exists until a reviewer promotes it — which is what keeps an extraction error cheap.
Bring a messy documentWhat if the extraction is wrong?
Then a reviewer corrects it, which is the designed path rather than an exception. Confidence is shown per field so attention goes where the platform was unsure, and because nothing has been created yet, a wrong extraction costs a correction instead of an amended claim with a started clock.
Three steps, one gate
From document to reviewed claim
- 01
It reads what arrived
An email, a PDF ACORD form, a letter, a scanned report. Parties, dates, loss description, policy references and claimant count are extracted.
- 02
It counts the claimants
A multi-claimant document is flagged as such. The platform counts and suggests; the reviewer decides whether it becomes one claim or several.
- 03
Confidence is shown per field
Not a single overall score. Field-level confidence directs review effort at what was actually uncertain.
- 04
A reviewer promotes it
On confirmation the claim is created through the same door as a typed FNOL — same numbering, same verification, same assignment, same audit.
The boundary
What the draft state guarantees
The value here is not the extraction accuracy. It is that an inaccurate extraction cannot do damage before a person looks at it.
Claim number
Not issued
Assignment
Not run — nobody owns it yet
Statutory clocks
Not started
Reserves
Do not exist
Claimant split decision
Belongs to the reviewer, not the extraction
Entitlement
Agentic FNOL and Document intelligence, both off until bought
What extraction cannot do
It reads documents well. It does not exercise judgment, and we have designed on the assumption that it will sometimes be wrong.
It does not create claims
There is no configuration that lets extraction promote its own draft. The human step is structural.
It does not split files
A multi-claimant document is counted and flagged. Whether that becomes one claim or three is a reviewer decision.
Poor scans stay poor
A faxed photocopy produces low-confidence fields. The honest outcome is more review, not a confident guess.