Skip to content
Back to work

Core decision

A language model reads each message but can only propose actions from a closed list. Code validates them against the case, and each insurer's rules, which are versioned files, decide where the case goes. Coverage decisions belong to a separate AI adjuster that shows its steps and the sources for each one.

How a message travels

One turn: reporting a crash

  1. turn
  2. input guard

    Recognizes orders to ignore the rules and off-topic messages before any model is called.

  3. agent (Claude Haiku 5.5)

    Understands the message and writes the reply. It can only propose actions from a closed list.

  4. case workflow (LangGraph)

    Code validates every proposed action against the case, along a fixed workflow.

  5. insurer rules

    Versioned files, one per insurer, each rule with its own test. The rules, not the model, decide where the case goes.

  6. policy search (pgvector)

    Cites the exact clause. A verifier checks that it exists, belongs to that policy and is in force.

  7. AI adjuster

    Takes the cases the rules do not close, decides and explains each step. The agent never decides coverage or payments.

  8. output guard

    Checks every message, in both languages, before the person sees it.

Another turn: a question unrelated to insurance

  1. turn
  2. input guard

    Answers with a fixed phrase from the catalog, in both languages, with no second model call.

Two cars with copper and blue headlights drive through the night on a wet road
Cautelia cover art
Cautelia home screen with the three fictional insurers and the buttons to chat, call or view policies
Home: pick an insurer and report by chat or by voice

In numbers and controls

  • More than 4,000 automated tests with an 80% coverage floor.
  • Insurers isolated with row-level security, and a test that fails if any table lacks it.
  • Hard spending limits: prepaid balances, an application-level ceiling and daily limits per visitor.

How it fails

  • Evaluation: 258 cases recorded in the latest full run (October 15, 2026). None failed.
  • Chat latency, from the moment the server receives the message to the first word: p50 1.6 s and p95 2.3 s in Spanish (87 turns); p50 1.5 s and p95 2.2 s in English (85 turns).
  • Voice latency: not measured.
  • It is not connected to any real insurer. Voice works in the browser only, with a daily call limit.