AI Insurance Claims Triage & Document Extraction
n8n workflow that processes new insurance claims end-to-end — Claude AI extracts structured data from unstructured descriptions, assigns CRITICAL/HIGH/STANDARD/LOW priority with SLA, routes to the correct specialist queue, and fires adjuster alerts + claimant acknowledgments in under 15 seconds. Zero manual triage.
Overview
A webhook-triggered n8n automation that eliminates manual claims triage for insurance operations teams. The moment a claim is submitted, Claude AI reads the full unstructured description and returns a structured extraction: priority level, claim category, primary cause, affected parties, estimated severity, coverage flags, fraud red flags, missing information checklist, SLA window, recommended specialist, confidence score, and a pre-written claimant acknowledgment message. Critical and high-priority claims are routed to a named specialist with a full adjuster brief email. Standard claims receive an automatic claimant acknowledgment. Every claim is logged to a Google Sheets audit trail and posted to Slack — all within 15 seconds of submission.
Problem
Insurance claims adjusters spend 30–45 minutes manually reading each new claim submission, extracting key fields, assessing priority, identifying red flags, deciding who should handle it, and drafting initial communications — before any actual claims work begins. At 50–200 claims per day, that's 25–150 hours of repeatable cognitive work that follows the same criteria every time. Critical fraud indicators get missed. High-severity claims sit in generic queues. SLA clocks start late. And claimants wait hours for a simple acknowledgment that their claim was received.
Approach
Webhook receives claim payload from any intake form or system. Normalize Claim Input Code node reads from the webhook body and standardizes all fields with safe fallbacks. Build Extraction Prompt constructs a detailed Claude prompt with the full claim context and a precise JSON schema — priority rules, category taxonomy, specialist routing logic, SLA thresholds, fraud indicator patterns, and a claimant message template. HTTP Request calls Claude 3.5 Sonnet via OpenRouter. Parse Triage Result extracts the structured JSON with error fallback to a senior adjuster. IF node branches on CRITICAL/HIGH vs STANDARD/LOW. High path: Assign Specialist Code node maps recommended_specialist to a named adjuster with email and queue. Gmail sends a full adjuster brief (priority badge, SLA deadline, extracted fields, red flags, missing info checklist, confidence score). Slack posts an urgent alert. Google Sheets logs the claim. Standard path: Gmail sends claimant acknowledgment using Claude's pre-written message. Slack posts a routine digest. Sheets logs the claim. Error Trigger catches failures and routes to a Slack error alert.
Tech Stack
Results
3/3 test claims processed correctly: CLM-2026-0847 (Riverfront Properties, $1.85M commercial fire) — CRITICAL, fraud_investigator assigned, 5 red flags detected (prior similar loss, disabled suppression system, suspicious presence), 2h SLA, 0.85 confidence. CLM-2026-0848 (Sullivan Construction, $285K workers comp crush injury) — CRITICAL, workers_comp_specialist assigned, OSHA notification flagged, 4h SLA, 0.95 confidence. CLM-2026-0849 (James Whitfield, $18.5K residential water damage) — STANDARD, property_adjuster assigned, clean claim, 24h SLA, 0.90 confidence. Adjuster emails, claimant acknowledgments, Slack alerts, and Sheets log all firing. End-to-end: 10–14 seconds per claim.