Claims FNOL assistant on Azure
A motor-insurance claims agent on Azure, built on the same harness as the support agent.
1 · Purpose
This is a motor insurer's first-notice-of-loss (FNOL) assistant: the moment a policyholder first says "a bus hit my car". It registers the claim, gives a reference, answers status questions and pays small payouts; anything larger, or anyone hurt, goes to a person. It runs on Azure and reuses the support agent's harness library, which supplies about 84% of the code it runs. It proves the harness carries to a second domain and a second cloud, and it is the running example in the agent harness course.
2 · Business benefits
3 · Status and how to join
Good first tasks this week:
- Run the agent on your laptop and report what was unclear
- Draft one exercise for the guided Azure tour
- Add a scenario for withdrawing a claim
4 · Requirements
What it must do
- Registers a motor claim from a conversation and returns a claim reference.
- Answers claim status from the claims system without calling the model.
- Pays payouts up to the limit at once and holds larger ones for a handler on the desk.
- Hands any injury, or a request for a person, to a claims handler.
- Lets policyholders sign in and see only their own claims.
- Politely refuses questions about cover, which are left to an assessor.
How well it must do it
- The claims system verifies a short-lived token on every call and trusts nothing the agent asserts.
- Each app uses its own database role; no app runs as the admin.
- A tool's idempotency key comes from the message, so a crash cannot register a claim twice.
- The payout limit is read from configuration, so it changes without a redeploy.
- Infrastructure is code, with each service's free allowance checked before it is added.
- The same scenarios and four gates run on a laptop before anything reaches Azure.
5 · Architecture, very high level
6 · Tech stack, per platform
| Component | Open stack | Azure | AWS |
|---|---|---|---|
| Chat page and handler desk | own web pages ✓ | Azure Container Apps ✓ | — |
| Sign-in | local test sign-in ✓ | Microsoft Entra ID ✓ | — |
| Agent | Python harness library, Pydantic AI ✓ | Azure Container Apps, App Configuration, Key Vault ✓ | not built |
| Model gateway | Groq called directly ✓ | API Management in front of Groq, Content Safety ✓ | — |
| Claims system | own MCP server ✓ | Azure Container Apps, on-behalf-of tokens ✓ | — |
| Durable waits | DBOS ✓ | DBOS on Azure Database for PostgreSQL ✓ | — |
| State store | PostgreSQL ✓ | Azure Database for PostgreSQL with pgvector ✓ | — |
| Telemetry and alerts | local trace log ✓ | Application Insights, Log Analytics, log alerts ✓ | — |
7 · What needs to be done
Phase 1: on the laptop 4 of 4 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Specification, simulated world and 29 scenarios | done | M | Specification design |
| All four gates pass with a scripted model | done | M | AI testing |
| Chat with it locally with the real model | done | M | Python |
| Checks placed by YAML as plug-ins | done | M | Python |
Phase 1: in Azure 3 of 5 done
| Task | Status | Size | Skill |
|---|---|---|---|
| All Azure resources created from Bicep and azd | done | L | Azure infrastructure |
| Security fixes for money and identity paths | done | M | Security |
| Live sign-in, on-behalf-of tokens and email alerts | done | M | Azure |
| Report a claim through the gateway and see it in Application Insights | doing | M | Azure |
| Deploy from GitHub Actions with a federated login | todo | S | CI/CD |
Evaluation loop and guided tour 0 of 3 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Turn the watch rules into online checks placed by YAML | todo | M | Python |
| A job that reads Application Insights and writes findings | todo | M | Azure, KQL |
| Write the guided tour of every Azure resource, with exercises | todo | M | Technical writing |
Phase 2 0 of 3 done
| Task | Status | Size | Skill |
|---|---|---|---|
| Label about 500 motor-claims messages for a local router | todo | M | Data labelling |
| Answer policy-wording questions from retrieved clauses, with citations | todo | L | RAG |
| Measure cost per conversation and per outcome from real traces | todo | M | Cost analysis |
8 · Roles
| Role | How many | Hours / week | What they do |
|---|---|---|---|
| Azure engineer | 1 | 3-5 hours/week | Owns the Bicep, the deploy pipeline and the Azure checks. |
| Agent engineer | 1 | 3-5 hours/week | Builds claims features and fixes findings, spec first. |
| Insurance domain reviewer | 1 | 1-2 hours/week | Checks the claims rules and wording against real practice. |
9 · What you'll learn
- How one harness library carries to a new domain and a new cloud.
- How to deploy an agent to Azure on free tiers, safely.
- How on-behalf-of tokens stop an agent acting beyond the user's rights.
- How durable waits let a handler approve a payout later.
- How to trace one conversation end to end in Application Insights.
Level-up questions
- Why does the claims system check a token instead of trusting the agent?
- What would you change to move the payout limit per policy type?
- Which checks belong at the gateway and which inside the agent?
- What would it take to swap Groq for a model hosted in an India region?