Claude Claims Triage Agent
Document-heavy workflows with Claude, humans in control
Claims teams spend much of their time reading unstructured documents and copying fields into systems. This build shows a safer pattern: the model extracts and drafts, plain code decides, and a person approves before anything leaves the building.
Benefits
Explainable routing
Fast track, adjuster review and SIU escalation are decided by deterministic rules with recorded reasons, not by the model.
Human approval by design
Replies are held in an approval queue. Nothing is released without a recorded approval, and every event is audit-logged.
Validated output
Extraction uses a forced tool call checked against a schema, so malformed output is rejected rather than passed on.
Measurable
An evaluation harness scores field extraction and routing on a labelled set, so you can track quality on your own claims.
How it works
- 1.Claim text arrives
- 2.Claude extracts fields via a forced tool call (validated)
- 3.Policy lookup over an MCP server
- 4.Rules route the claim and record reasons
- 5.Claude drafts an acknowledgement (never states approval, denial or payout)
- 6.Reply waits in the approval queue for a human
How to use it
Install
pip install -e .[dev]Run the tests
pytestOffline evaluation (no API key)
python -m claims_triage.cli --evalEvaluate with real Claude
python -m claims_triage.cli --eval --liveQueue one claim for approval
python -m claims_triage.cli --file claim.txtPolicy lookup as an MCP server
python -m claims_triage.mcp_serverOffline demo: one synthetic claim
Captured from a real run with the offline rule-based mock model on synthetic data. It shows the pipeline, routing and approval gate working, not Claude's accuracy.
$ python -m claims_triage.cli --file claim.txt
{
"extracted": {
"claimant_name": "Maria Lopez",
"policy_number": "POL-10001",
"incident_date": "2026-08-03",
"incident_type": "auto",
"loss_amount": 1450.0
},
"policy": { "active": true, "coverage_types": ["auto", "property"], "deductible": 500.0, "limit": 50000.0 },
"decision": { "route": "fast_track", "priority": 3, "reasons": ["covered, active policy, low value"] },
"reply": { "subject": "We received your claim POL-10001", ... }
}
$ python -m claims_triage.cli --eval
claims=8 field_accuracy=100.0% route_accuracy=100.0% avg_latency=0msRead the results honestly
- •The 100% score comes from a rule-based mock on 8 synthetic claims written with the build. It validates the pipeline and gates only.
- •Accuracy, cost and handle-time gains must be measured on your own claims with --live before any figure is quoted.
- •All data is synthetic. Real claims need a data-handling review first.
Reference build demonstrated on synthetic data, not a client deployment. Names of third-party products describe what the build runs on and do not imply endorsement or partnership.
Want this adapted to your environment?
We tailor it to your data, systems and approval process, and measure results on your own cases.
Talk to Us