Case Study 02  /  AI Agents · Insurance · Claim Automation

ClaimsAI

Four AI agents that run the claims desk of an AI-first insurance agency — from the first email about a loss to the final payment, without anyone chasing paperwork by hand.

Role
AI Engineer: architecture, build, deployment
Year
2026
Stack
n8n, AMS 360 API, LLMs, Sheets, Gmail
Status
Live in production
4Agents in production
24/7Inbox & claim monitoring
1Shared source of truth

( 01 ) The problem

Claims don't fail on the big decisions. They fail in the follow-up.

When a client reports a loss, an insurance agency has to classify the claim, find the right carrier for that line of business, send it over with attachments, and log it in its management system, AMS 360.

Then the real work starts: chasing the carrier for a claim number, chasing the adjuster for reserves and payment status, and making sure nothing sits untouched until it becomes a missed deadline. All of it lived in inboxes and in people's heads.

The agency wanted that whole lifecycle to run on its own, with humans stepping in only when a claim actually needs judgment.

( 02 ) The system

Four agents, one tracker, every step of the claim covered.

Each agent owns one stage of the claim lifecycle and hands off through a shared Master Claims Tracker. One agent's output is the next one's trigger, so the system moves a claim forward without a central coordinator.

Claims inboxIMAP · new emailAGENT 1Intake & Routingclassify · prioritiseCarrierclaim + attachmentsMaster Trackershared stateAMS 360claim + documentsAGENT 2 · DAILY 8:00Follow-Up MonitorAGENT 3 · ON CONDITIONPayment TrackerAGENT 4 · WEEKLYAudit & Compliance

Agent 01

Intake & Routing

Trigger — new email in the Claims inbox

  • →Classifies each claim by type: General Liability, Workers' Comp, Auto, Property and more.
  • →Flags "lawsuit", "injury" or "urgent" claims for an immediate high-priority alert.
  • →Matches the sender to the client profile in AMS 360.
  • →Finds the right carrier for the line of business and sends the claim with all attachments.
  • →Creates the claim in AMS 360, saves documents and logs it to the tracker with the email thread ID.

Agent 02

Follow-Up Monitor

Trigger — every morning at 8:00

  • →Scans the tracker for claims still missing a claim number or an assigned adjuster.
  • →Follows up with the carrier inside the original email thread, so the history stays in one place.
  • →Repeats every 7 days until the data arrives and stamps every follow-up.

Agent 03

Payment Tracker

Trigger — claim # and adjuster known, no reserve recorded

  • →Requests reserve amounts and payment status directly from the adjuster.
  • →Records paid amounts and current status in the tracker.
  • →Flags any adjuster who hasn't responded in 14 days.

Agent 04

Audit & Compliance

Trigger — weekly, plus continuous checks

  • →Sends account managers a PDF summary of all open claims every Monday.
  • →Verifies that every reported claim has its documents saved in AMS 360.
  • →Escalates claims with no activity for 30+ days.
  • →Tracks the statute of limitations on unresolved claims.

( 03 ) Data model

Agents coordinate through data, not through each other.

Every field in the tracker has exactly one owner. An agent only acts once the fields it depends on are filled, which keeps the system predictable and easy to debug.

FieldWritten byWhat it holds
Business nameAgent 1Legal name of the insured
Date of lossAgent 1When the incident happened
Claim typeAgent 1GL, WC, Auto, Property…
Claim numberAgent 2Received from the carrier
Adjuster emailAgent 2Contact for ongoing tracking
Reserve / paidAgent 3Amounts and payment status
Audit flagAgent 4Set when data is incomplete

( 04 ) Engineering decisions

Small choices that made it production-safe.

01

Follow-ups stay in the original thread

Agent 1 stores each email's thread ID, so every reminder lands in the same conversation. Carriers see full context, and the agency has one auditable trail per claim.

02

Explicit activation conditions

Agents fire only when their preconditions are met in the tracker. No agent ever asks an adjuster for money before a claim number exists.

03

The right model for each job

A mix of hosted models (Claude, GPT) and a local Ollama model, matched to each task's cost, speed and sensitivity.

04

Humans stay in the loop where it matters

Urgent claims, silent adjusters, stale cases and approaching legal deadlines are escalated to people instead of being handled silently.

05

A tracker anyone can read

Shared state lives in a Google Sheet, so the team can see and correct what the agents did without opening n8n.

06

Secrets out of the workflows

All API credentials, including AMS 360, live in n8n's credential manager, never inside nodes or exported workflows.

( 05 ) Stack

Built on tools the agency already used.

n8nAI Agent nodesClaudeGPTOllamaAMS 360 REST APIGmail / IMAP / SMTPGoogle SheetsSchedule triggersPDF reporting

( 06 ) Outcome

The claims desk now runs itself.

All four agents run in production. Every claim that reaches the inbox is classified, routed to its carrier, logged and tracked through to payment, with the team pulled in only when a case needs a human decision.

Runs itself

Every claim is classified, routed to its carrier and logged automatically — with no manual data entry.

Nothing slips

Carriers, adjusters and legal deadlines are chased and monitored around the clock.

Humans for judgment

The team is pulled in only for urgent claims and decisions that genuinely need a person.