Insurance
Claims intake from the first notice of loss
Read the first notice and fill the intake the adjuster already opens. A person still decides coverage.
A typical engagement, not a named customer result. The desk still decides coverage. OCR and structured extraction fill the intake in the claims system people already open. We stay until that is how a first notice is handled.
Where it starts
A first notice of loss lands as an email, a PDF, and photos. Before an adjuster can start, someone retypes the same fields into the claims system: who reported it, what happened, when and where, the policy, and the attachments.
What goes into the system
OCR reads the packet. Document understanding extracts the fields a finished notice needs—reporter, loss facts, dates and place, policy identifiers, and attachments—into a structured intake schema. Those values are validated and mapped onto the intake screen the adjuster already opens.
The filled intake lands in the claims system they already use, and the notice is routed to the queue the desk already works. There is no second system to sign into.
Claims packet into intake
Map each beat to a layer before you buy model capacity. A person still clears coverage.
- OCR Read FNOL packet Policy #, loss date, photos → fields
- Classifier Route the claim Auto / property / incomplete
- Retrieval Pull policy facts Coverage clauses on file
- LLM Draft intake note Only the messy narrative
- Human Decide coverage Person still clears the step
Agentic AI enters later — if the system must open the claim screen, attach the packet, and set status with audited tools. Not for inventing coverage.
What a person still decides
Whether the notice is complete enough to open, and every coverage decision. Low-confidence fields and incomplete packets stay in front of a person for review. This engagement does not approve a claim or decide what is covered.
How the engagement runs
Map the work. Sit with the person who retypes the notice. Mark the fields they copy, the photos they attach, and the point where an adjuster takes the file.
Build where the value is. Put the OCR, the structured extraction, and the field mapping into the claims system, on the intake screen they already use. With the team, name what a finished notice is before the build goes further.
Drive it to the result. Stay until a new notice is read, filled, and routed there as the way the desk handles it. The adjuster still opens the file and makes the call. Then the system is theirs.