Report a car accident.
Watch the agents work.
Decide as the handler.
Muster’s claims desk runs on a real Flowable process. A document agent reads the garage’s estimate, one agent drafts the claim, a second checks every field against the documents, guardrails stop anyone from steering the AI, and a person decides. Start a claim, and every step runs in the trial while you watch.
The agents read, draft and check. A person decides.
-
AI agent
1
Read
The document agent reads the repair estimate and proposes the garage, the plate, the date and the total. Nothing goes into the form until the customer applies it.
Document agent -
RulesAI judge
2
Guard
Fixed rules, then an AI judge, read the customer’s text before anything is drafted. A submission that tries to steer the AI goes to a person instead.
Guardrails -
REST lookupAI agent
3
Draft
The policy comes from the insurer’s system over REST. Then the intake agent turns the customer’s words and the estimate into a claim record.
Service registry, utility agent -
AI agent
4
Check
A second agent sees only the sources and the draft, and rates every field as supported, inferred or missing.
Utility agent -
Person
5
Decide
The handler sees the rule checks and the flags first. Approve only works when the rules pass, in the form and again in the process.
User task, gateway
I tried to talk it into a payout 29 ways. Your turn.
Each exhibit is a real attack from the test suite. Pick one, send it as a live claim and watch where it gets stopped. Or write your own.
Tested against the live trial before anyone saw it.
- v8One API request approved a claim on a lapsed policy.
- v9The judge’s reason, which the handler reads, carried the model’s own reasoning.
- v10An honest but impatient customer was flagged as manipulation.
- v11Whoever closed a blocked claim could rewrite why it was blocked.
Rather watch than click?
Built so someone else can give this demo.
Claims systems need every field filled in, so AI that drafts claims will guess. In Flowable, one agent drafts the claim from the customer’s documents, a second agent checks every field against those documents and flags what was guessed, and a person decides.
The fine print
- One person plays every role in the trial. In production the review goes to a claims group, and whoever reports a claim cannot approve it.
- The policy system is a mock on Cloudflare with fictional policies. In production it is the insurer’s own system, behind the same service registry.
- The model layers are probabilistic. The riskiest cases were repeated, and the business rules do not depend on any model.
- This page talks to the trial through a small Cloudflare Worker that holds the access token. Your browser never sees it, and only the live copy of the process can be touched.
- The live demo runs on a 30-day trial until mid-November 2026, with a daily limit. Claims left open are tidied up every night.