Skip to content

Insurance

Claims triage

A triage agent that reads each first-notice-of-loss, classifies severity and coverage, and routes it to the right adjuster with the facts already extracted — so claims stop aging in a shared inbox.

  • LangGraph
  • Anthropic
  • Postgres
  • River
-64%

time to first action

adjuster capacity

3×

In short

  • Every claim is classified and routed the moment it arrives, not when someone opens the inbox.
  • Severity and missing-information checks run up front, so adjusters open a claim that's ready to work.
  • Coverage decisions the model isn't sure about go to a human — never auto-denied.

The engagement

Sector
Insurance (P&C)
Company shape
Regional carrier, high claim volume
Engagement
Triage agent + adjuster tooling, ~10 weeks
Region
North America

Anonymised by policy

Every engagement runs under NDA — we protect our clients' confidentiality as a matter of policy, and we'd extend the same discretion to your work. The constraints, the shape, and the outcome here are real; the identity is redacted.

The challenge

First-notice-of-loss arrived by email, web form, and PDF, landed in one shared queue, and waited for a human to read, classify, and route it. Simple claims and complex ones aged at the same rate, and the delay was worst on exactly the severe claims that needed a fast response.

A blunt auto-classifier was a non-starter in a regulated line: a wrong coverage or severity call has real consequences, so anything the model touched had to be explainable and reversible — and a low-confidence coverage decision could never auto-deny a claimant.

The approach

  1. 1
    Extract before you classify

    Each claim is normalised and the key facts (loss type, date, policy, damages) are extracted and cited before any routing decision, so a human can check the inputs.

  2. 2
    Severity and coverage as separate calls

    The agent scores severity and checks coverage independently, and a low-confidence coverage call is flagged for an adjuster rather than decided.

  3. 3
    Route with the work already done

    The claim reaches the right adjuster with facts extracted, missing information listed, and severity attached — so the human starts from a prepared file.

  4. 4
    Reversible and logged

    Every routing and classification decision is logged and reversible; nothing the agent does is a dead end.

The solution

The agent ingests each notice, extracts and cites the claim facts, scores severity, and checks coverage. High-confidence, low-severity claims route straight to the right queue; anything severe or uncertain is escalated to an adjuster with the extracted facts and a list of what's missing. Coverage is never auto-denied — a low-confidence decision is a human decision.

Because every step is logged and reversible, the carrier gets speed without giving up the audit trail a regulated line requires.

The system we built

  1. Intakeemail / form / PDF
  2. Extract + citeloss facts, policy
  3. Severity scoretriage priority
  4. Coverage checkflag low-confidence
  5. Routeto the right adjuster
Abstract by design — the architecture we built, not a client screen. No confidential data is shown.
Illustrative
Walkthrough: intake → extract → severity → route
Illustrative
The adjuster opens a claim with facts extracted and gaps flagged

The results

time to first action

-64%

median, vs. the shared-inbox baseline

adjuster capacity

3×

claims handled per adjuster-day

auto-denied claims

0

low-confidence coverage always goes to a human

Data & safety

Guardrails
Coverage is never auto-denied; low-confidence decisions route to a licensed adjuster.
Traceability
Every classification and routing decision is logged, cited, and reversible for audit.
Data handling
Claim data stays within the carrier's environment; no PII to third-party models.

The stack

  • LangGraph
  • Anthropic
  • Postgres
  • River
Capability: AI Agents & Automation

Questions we get

Coverage is checked as its own step and a low-confidence decision is flagged for a licensed adjuster — the agent never auto-denies a claimant.

A prepared file: the extracted, cited claim facts, a severity score, and a list of any missing information — so they start from a worked-up claim, not a raw email.

Yes — every classification and routing decision is logged, cited to the source, and reversible, which a regulated line requires.

More selected work

// Let's build

Want an outcome like this on your problem?

Tell us where you are. We reply within a day with a concrete next step.