Case study · Meridian Claims Insurance

Five agents replaced 40 humans.

A multi-agent claims triage system at Meridian that auto-resolves 92% of incoming property claims under $25k. Built in 14 weeks. Live for 11 months.

92%AUTO-RESOLUTION
40FTE TRIAGE TEAM REPLACED
$3.8MANNUAL OPEX REDUCTION
6 hrsAVG CLAIM CYCLE TIME
Engagement details
CLIENT
Meridian Claims Insurance
INDUSTRY
Property & casualty insurance
DURATION
14 weeks · ongoing operate
TEAM
5 Mindlytic seniors + 6 Meridian SMEs
STARTED
Q3 2024
STATUS
Live · auto-handles claims under $25k
THE PROBLEM

A 40-person team manually triaging claims.

Meridian's claims triage team spent their days routing claims, requesting documents, checking policy coverage, and assigning to adjusters — work that was rules-based, repetitive, and impossible to scale during weather events.

During Hurricane Mara in 2024, claim cycle time spiked from 4 days to 17. Customers churned. Leadership greenlit an AI build with a hard SLA: cycle time under 24 hours regardless of volume.

WHAT WE BUILT

Five specialist agents, one orchestrator.

Rather than a single 'claims agent,' we built five specialists: an intake agent (parses FNOL), a coverage agent (checks the policy), a fraud agent (scores against patterns), an estimation agent (applies the rate book), and a routing agent (assigns or auto-pays). One orchestrator coordinates them with a written rubric and a human-in-the-loop checkpoint.

Anything over $25k or with a fraud score above threshold gets human review. Everything else is auto-resolved end-to-end with full audit trail.

  • Intake agent: parses FNOL emails, calls, and online forms into structured claim
  • Coverage agent: checks against policy + endorsements + deductibles
  • Fraud agent: 18-feature model + LLM rationale; flags for SIU review
  • Estimation agent: applies rate book + photos + parts catalog
  • Routing agent: auto-pay, adjust, or escalate based on rubric
  • Audit log: every decision queryable for compliance + dispute resolution
WHAT WE FOUND

Specialist agents beat one big agent every time.

Our v1 was a single agent with all five tools. It worked 71% of the time. v2 — the same logic, decomposed into five specialists with one orchestrator — hit 92% on the same eval set. The reason: each specialist had a tighter rubric and could fail fast without poisoning the others.

We also learned that compliance auditors trusted the multi-agent version because they could read each specialist's reasoning independently. Single-agent traces were too tangled to certify.

THE OUTCOME

Cycle time under 6 hours. $3.8M opex out.

Eleven months in, the system auto-resolves 92% of claims under $25k. Average cycle time is 6 hours, including overnight. During Hurricane Lila in 2025, the system held its SLA at 17 hours peak — a 41x improvement over the prior year's manual response.

Meridian retained 8 of the 40 triage staff in higher-skill SIU and complex-claim roles. The remaining $3.8M annual opex was redirected to product and underwriting tech.

Five agents replaced a forty-person triage team. We still double-check the numbers because they seem impossible.
Priya Menon
Head of Claims, Meridian Insurance

What we used.

LLM
Claude 3.5 + GPT-4o
Orchestration
LangGraph + Temporal
Backend
Go + FastAPI
DB
Postgres + S3
Eval
LangSmith + Braintrust
Frontend
Next.js
Infra
AWS ECS
Compliance
SOC 2 controls

Other case studies.

Have an AI project in mind?

Book a 30-minute discovery call. We'll tell you if we're the right fit — and if not, who is.