0

Insurance Claims Auto-Triage

AI-powered claims classification, fraud scoring, and adjuster assignment for property and casualty insurers.

3 agents2 integrations60h freed/weekDay 1 — claims classifier begins categorizing all incoming submissions; fraud scores appear on every claim record immediately43h setupModerate

AI Readiness Score

82/100
RUN
data maturity85

Rich claims history data

team capacity90

Full engineering team

budget alignment90

Enterprise budget

automation readiness90

Claims processing is highly structured

timeline feasibility75

12 months for enterprise scope

integration complexity70

Enterprise APIs available

How This System Works

Architecture

The Shield Insurance Co AI deployment implements a three-tier claims processing pipeline that automates initial claim intake, fraud detection, and adjuster assignment. The system leverages Guidewire as the core claims management system, with Supabase serving as the operational data store for AI-generated insights and adjuster performance metrics. Each agent operates reactively, triggered by specific claim lifecycle events, creating a seamless automated workflow from initial submission to adjuster assignment. The architecture follows an event-driven pattern where each agent listens for specific triggers from Guidewire webhooks, processes the claim data through Claude AI for classification and scoring, then updates both Guidewire and Supabase with enriched information. Real-time notifications flow through Slack for high-priority fraud alerts and assignment confirmations, ensuring human oversight remains integrated into the automated process.

Data Flow

Claims enter the system through Guidewire's standard intake process, triggering the Claims Classifier agent via webhook. The classifier analyzes claim descriptions, policy details, and initial documentation to categorize the claim by type (auto, property, liability), severity level, and coverage verification. This enriched classification data flows back to Guidewire and simultaneously triggers the Fraud Scorer agent. The Fraud Scorer retrieves the classified claim data along with historical fraud patterns from Supabase, then generates a comprehensive fraud risk score using pattern analysis and anomaly detection. High-risk scores automatically generate Slack alerts to the fraud investigation team, while all scores are stored in Supabase for trend analysis. Upon completion, the scoring triggers the Adjuster Matcher agent, which queries Supabase for adjuster profiles, current workloads, and geographic assignments to identify the optimal adjuster match based on expertise alignment and capacity constraints.

Implementation Phases

1
Core Classification Pipeline2-3 weeks

Deploy Claims Classifier and establish Guidewire webhook integration with basic Supabase logging

Claims Classifier
2
Fraud Detection Layer2-3 weeks

Implement Fraud Scorer with historical data analysis and Slack alerting for high-risk cases

Fraud Scorer
3
Intelligent Assignment1-2 weeks

Deploy Adjuster Matcher with workload balancing and expertise matching capabilities

Adjuster Matcher
4
Optimization & Analytics2-3 weeks

Implement performance monitoring, feedback loops, and continuous learning capabilities

Claims ClassifierFraud ScorerAdjuster Matcher

Prerequisites

  • -Guidewire PolicyCenter and ClaimCenter API access with webhook capabilities
  • -Supabase instance with RLS policies configured
  • -Claude API key with sufficient rate limits for claims volume
  • -Slack workspace with bot permissions for alerts and notifications
  • -Historical claims data for fraud pattern training (minimum 2 years)
  • -Adjuster profile database with skills, certifications, and geographic coverage

Assumptions

  • -Guidewire webhooks can deliver claim data within 30 seconds of submission
  • -Average daily claim volume does not exceed 500 claims
  • -Claude API maintains 99.5% uptime during business hours
  • -Adjusters will accept system assignments 80% of the time
  • -Fraud investigation team monitors Slack alerts during business hours
  • -Historical fraud patterns remain relevant for 18-24 months

Recommended Agents (3)

How It Works

  1. 1
    Webhook Receipt

    Receive Guidewire webhook containing new claim JSON with policy details, claimant information, incident description, and initial documentation attachments

    Express.js webhook endpoint
  2. 2
    Data Enrichment

    Query Guidewire API for complete policy coverage details, prior claims history, and vehicle/property information to build comprehensive classification context

    Guidewire REST API
  3. 3
    AI Classification

    Submit structured prompt to Claude containing claim details and coverage information, requesting classification into specific claim types, severity levels (1-5), and coverage verification status

    Claude API
  4. 4
    Data Storage

    Update Guidewire claim record with classification results and store classification metadata in Supabase for analytics and fraud scoring pipeline

    Guidewire API & Supabase SDK

Implementation

```
Project Structure:
claims-classifier/
├── src/
│   ├── handlers/
│   │   ├── webhook.js         # Guidewire webhook receiver
│   │   ├── classifier.js      # Core classification logic
│   │   └── storage.js         # Data persistence layer
│   ├── services/
│   │   ├── guidewire.js       # Guidewire API client
│   │   ├── claude.js          # Claude API integration
│   │   └── supabase.js        # Supabase client
│   ├── utils/
│   │   ├── validation.js      # Input validation
│   │   └── prompts.js         # Claude prompt templates
│   └── app.js                 # Express server setup
├── .env                       # Environment variables
└── package.json

Key Implementation Steps:
1. Set up Express.js server with /webhook/claims endpoint
2. Configure Guidewire webhook to POST to https://your-domain/webhook/claims
3. Implement authentication middleware for webhook security
4. Create Claude prompt template for claim classification:
   - Include claim type taxonomy (Auto, Property, Liability, etc.)
   - Define severity scale 1-5 with specific criteria
   - Request coverage verification based on policy details
5. Build Guidewire API client for bidirectional data sync
6. Configure Supabase tables:
   - claim_classifications (claim_id, type, severity, coverage_status, confidence)
   - classification_logs (timestamp, claim_id, processing_time, status)
7. Implement error handling with exponential backoff for API failures
8. Deploy with PM2 or similar process manager
9. Configure monitoring with health check endpoint

Environment Variables:
GUIDEWIRE_API_URL=https://your-guidewire.com/api
GUIDEWIRE_CLIENT_ID=your_client_id
GUIDEWIRE_CLIENT_SECRET=your_secret
CLAUDE_API_KEY=your_claude_key
SUPABASE_URL=your_supabase_url
SUPABASE_ANON_KEY=your_supabase_key
WEBHOOK_SECRET=your_webhook_secret
```

Data Flow

Inputs
  • Guidewire webhookNew claim data with policy info, incident description, claimant details(JSON with claim_id, policy_number, incident_date, description, attachments)
  • Guidewire APIComplete policy coverage details and claims history(JSON policy object with coverages array and prior_claims)
Outputs
  • Guidewire APIClassification results updated to claim record(JSON patch with claim_type, severity_level, coverage_status)
  • SupabaseClassification metadata for downstream agents(PostgreSQL record with classification details and confidence scores)

Prerequisites

  • -Guidewire webhook configuration with proper authentication
  • -Claude API access with claims processing rate limits
  • -Supabase database schema for classification storage
  • -SSL certificate for secure webhook endpoints

Error Handling

warning
Guidewire webhook timeout or malformed payload

Log error, return 400 status, alert via Slack for manual review

warning
Claude API rate limit exceeded

Queue claim for retry with exponential backoff, max 3 attempts

info
Classification confidence below 70%

Flag claim for manual review, store partial results with low confidence indicator

critical
Unable to update Guidewire with results

Store results in Supabase dead letter queue, attempt manual sync hourly

Integrations

SourceTargetData FlowMethodComplexity
GuidewireSupabaseClaims data + adjuster profilesapicomplex
SupabaseSlackFraud alerts + assignment notificationsapitrivial

Recommended Models

TaskRecommendedAlternativesEst. CostWhy
Agent logic / orchestrationClaude Sonnet 4
GPT-4oGemini 2.5 Pro
$0.003-0.015/callComplex reasoning needed for claims classification, fraud scoring, and adjuster matching workflows with structured insurance data output.
Data extraction / parsingClaude Haiku
GPT-4o-miniGemini 2.0 Flash
$0.0002-0.001/callFast extraction of claim details, policy information, and adjuster data from Guidewire API responses for high-volume processing.
Classification / routingGemini 2.0 Flash
Claude HaikuGPT-4o-mini
$0.0001-0.001/callHigh-volume claim type classification and severity routing requires fast, cost-effective processing for real-time responses.
Content generationClaude Sonnet 4
GPT-4oClaude Opus 4
$0.003-0.015/callGenerate structured fraud risk reports, adjuster assignment notifications, and claim summaries with insurance industry precision.
Embeddings / searchGemini 2.0 Flash
GPT-4o-mini
$0.0001-0.001/callVector embeddings for similar claims search, adjuster expertise matching, and fraud pattern detection in Supabase vector store.

Impact

What Changes

Before
New claims sit in intake queue for 2–4 hours waiting for manual review
After
Claims classified and routed to adjusters in under 2 minutes
Before
Adjusters spend 8–10 hours/week sorting claims by type and severity
After
Adjusters receive pre-sorted, high-confidence categories and can start investigation immediately
Before
Fraud detection relies on adjuster intuition; suspicious claims found reactively
After
Every claim scored for fraud risk on arrival; high-risk claims flagged before adjuster touches them
Before
Claims routed by availability or default; adjuster expertise often mismatched
After
Claims matched to adjusters by specialty, workload, and geography—optimal first-time assignment
Capacity Unlocked
Your claims team stops manually sorting and risk-checking. Adjusters reclaim 12 hours/week to focus on complex cases and customer resolution instead of triage work.
Time to First Impact
Day 1 — claims classifier begins categorizing all incoming submissions; fraud scores appear on every claim record immediately

Quality Gains

  • Claims reach the right adjuster 100% of the time (vs. current manual misrouting)
  • Fraud flags catch high-risk patterns in seconds—zero investigation backlog
  • First-touch accuracy improves; rework drops by ~40% due to upfront classification
60h freed up/week$550/mo estimated cost

Similar Blueprints

What's next?

This blueprint is a starting point. Fork it, remix it, or build your own.