Building ClaimFlow: An AI Autopilot Agent That Processes Insurance Claims End-to-End

By Joshua Damilola Ajisafe (CodedLabs) — June 2026

---

I spent the last week building an AI agent that automates the entire claims processing workflow — from intake to resolution — using two Qwen Cloud models and a human-in-the-loop dashboard. Here's how it went.

The Problem

Claims processing is a $40B+ industry where 70%+ of the work is still manual. A customer sends a message (WhatsApp, web form, email), attaches photos as evidence, and then waits days for a human adjuster to review everything. Most claims are straightforward — the damage is visible, the evidence checks out — but they still sit in a queue behind edge cases and fraud investigations.

I wanted to build something that auto-resolves the 80% of claims that are clear-cut, flags the suspicious ones, and only escalates the genuinely tricky cases to a human.

What ClaimFlow Does

ClaimFlow is a 6-stage Autopilot Agent built for Track 4 of the Global AI Hackathon with Qwen Cloud:

1. Intake — Normalizes messages from any channel (WhatsApp, web, API), detects language, validates input
2. Classify — Qwen 3.7 Max categorizes the claim (damage/warranty/return/refund), identifies the object, assesses urgency
3. Verify — Qwen Vision Max inspects submitted images against the claim text, checking evidence requirements
4. Fraud Detection — A Four Rs framework (Recognize -> Reject -> Reveal -> Route) flags wrong objects, manipulated images, and suspicious user patterns
5. Decide — Auto-approves high-confidence claims, requests more info for medium-confidence, escalates low-confidence or fraudulent ones
6. Respond — Auto-replies in the user's language (English, French, Spanish, Arabic) or creates a human case file

Why Qwen Cloud

I picked Qwen Cloud for two reasons:

First, the free tier is generous. 1,000,000 tokens with no payment method required. For a solo hacker building over a weekend, not worrying about burning through API credits while debugging is a real advantage.

Second, the dual-model strategy. Qwen 3.7 Max handles text classification with strong reasoning — it correctly identified a French-language phone warranty claim and a frustrated customer's refund request from context alone. Qwen Vision Max handles image inspection, comparing what's actually visible in photos against what the user claims.

The Anthropic-compatible endpoint made integration straightforward. If you've used the Anthropic SDK before, Qwen Cloud is a drop-in replacement — just change the base_url.

What I Learned

The hard part isn't the AI — it's the pipeline design. Getting Qwen to classify a claim is one API call. Building a system that handles 6 stages, with decisions cascading from one stage to the next, while keeping a human in the loop — that's where the real engineering happens.

Fraud detection needs vision + context. A photo can show damage, but the same photo submitted by a user with 5 prior claims and 2 fraud flags tells a different story. The Four Rs framework combines image analysis with user history to catch patterns that either alone would miss.

Multi-language support is a small detail that makes a big difference. The agent detects the user's language at intake and responds in it.

Try It Yourself

The code is open source (MIT) at github.com/Starboycoded/claimflow

Built for the Global AI Hackathon Series with Qwen Cloud. Track 4: Autopilot Agent.