
The situation
A third-party administrator receives claims by email and portal all day: policy documents, discharge summaries, bills and photographs. Before a claims handler can start, someone must check that the file is complete. Then they sort the file and send it to the right team.
What we deployed
An intake agent reads each incoming claim. It checks the document set against the checklist for the policy type, pulls out the fields claims handlers search by, and routes the file. The manifest lets it read the intake channels and write to the claims system's intake queue, and nothing else.
A file below the confidence threshold waits for a named claims handler, with the evidence gathered. That covers an unreadable scan, documents that do not match, or an unusual policy. The manifest lets only a person reject a claim, so the agent never does.
The five gates at work
- SCOPE
Read the intake channels · write to the intake queue only
- ROLE
Classify and route · never decide or reject a claim
- LIMITS
A spend limit on each run, set in the signed manifest
- REVIEW
Low-confidence files held for a named claims handler
- SEAL
Each routing decision in the audit trail, with its reason
What changed
The first pass now runs around the clock. Each routing decision is saved in the audit trail with its reason.
Each held file is saved with a note of what the agent could not decide, and why.
If your team runs a process like this one, start with a teardown of it. We measure its volume and exception rate, and you keep the document.
More deployments

Invoice matching for a distributor
Invoices needed a three-way match, under a spend limit.

Policy answers with named escalation
Policy questions needed answers, with least-privilege access to the policy library.