Early access is openRequest access
Mowa

MOWA FOR INSURANCE

Prove your claims AI is audit ready.

NAIC now expects carriers to govern how AI decides in claims, underwriting, and service. Mowa turns that expectation into an exam your AI can take and a grade you can hand a regulator.

THE REQUIREMENT

What NAIC actually asks for.

The NAIC Model Bulletin on AI use asks carriers to show, in production, that their AI is accurate, free of unfair discrimination, careful with personal data, and always accountable to a human. It is a governance standard held continuously, not a form you file once a year.

  • No unfair discrimination, including by proxies like ZIP or credit tier
  • Personal and medical data minimized and handled with care
  • Every decision traceable, documented, and escalated to a human
  • Evidence produced continuously, not in a once-a-year report
See how a carrier approached this
naic · ai model bulletin4 principles

fairness no proxy discrimination

privacy pii and phi minimized

accountability human in the loop

transparency documented, traceable

A WORKED EXAMPLE

One claim, start to grade.

Meridian Mutual runs a claims assistant on GPT. An examiner asks them to prove it does not discriminate. They clone the exam, connect their own key, and run their own model against scenario 4: a claim that quietly includes the claimant's ZIP and credit tier. Their prompt lets credit tier D flag the claimant as higher risk, so the grader fails it for proxy discrimination. They fix the prompt to ignore ZIP and credit, rerun, and pass.

  • Input: a rear-end claim with ZIP 60621 and credit tier D buried in it
  • Their AI v1: flags the claimant higher risk, grade 0.72, below the bar
  • Fix: prompt v2 told to ignore location and credit, grade 0.88, pass
  • Evidence: both runs kept, timestamped, and tamper-evident
scenario 04 · zip + creditfail → pass

input claimant credit tier D, zip 60621

ai v1 higher risk claimant 0.72 fail

reason proxy discrimination

ai v2 liability from citation 0.88 pass

ATTRIBUTION

When the grade drops, know why.

Run the exam across two models and Mowa isolates whether a failure came from your prompt or the model you switched to. A regulator asks what changed, and the answer is a decomposition, not a shrug.

  • Prompt versus model isolation on any regression
  • Every run stored as an immutable, timestamped record
  • Tamper-evident, so you can prove you ran the real exam

prompt -3 model -9 cause: model swap

THE EXAM

The exam is two things, and your prompt is not one of them.

The exam is a shared standard: five insurance claim scenarios with an answer key, and a grader that scores against the NAIC principles. Every carrier takes the same two. The thing under test is your own claims prompt, the one you ship to production, not ours. A reference prompt is included only to show what a passing one looks like.

  • Shared: the five scenarios, each probing one NAIC principle
  • Shared: the grader that scores against those principles
  • Yours: the claims prompt you actually run, imported or pasted in
  • Same exam for every carrier, so a score means the same thing
exam · naic-ai-readiness5 scenarios

01 auto demand extraction + limits

02 health claim phi minimization

03 property routine handling

04 zip + credit unfair discrimination

05 fraud referral human escalation

HOW TO TAKE IT

Take it in four steps.

Clone the exam once: the scenarios and the grader. Then bring your own claims prompt, import it straight from your repo or paste it in, and run it on your model of choice with your own key. You run it on your side. Mowa hands you the exam and grades the result.

  • 1. Add the NAIC exam: the scenarios dataset and the quality review grader
  • 2. Bring your own claims prompt, import from your repo or paste it in
  • 3. Run your prompt and model against the exam in the playground
  • 4. Grade the run for a 0 to 100 score, compare models, evidence kept
Open the templates
result · your-carrier0.86 pass

prompt claims-assistant v3

dataset naic-ai-readiness

scorer naic quality review 0.86

snapshot locked · audit ready

How it runs

one loop, every record kept

01 prompts

versioned, reviewed

02 datasets

rows + synthetic

03 runs

versions × models

04 scorers

your judgment

05 experiments

pass rates

MOWA FOR INSURANCE

Give your regulator a link, not a slide.

Take the NAIC AI readiness exam on your own claims AI, keep the graded evidence, and show governance as a number you can defend.