Risk score and band
A number and a letter the desk can act on.
Risk Assessment
Price AI risk at the time of underwriting, not after the claim.
The application tells you an AI agent exists. Risk Assessment tells you what it does under pressure, what that could cost, and what the insured has to fix before you bind.
Built for real underwriting work.
Simulated conversations built on the AI agent’s own role and permissions, run at volume, plus adversarial testing under the insured’s authorisation. Signals from public sources, the dark web, social channels and metadata feed the same profile.
Every finding is filed under one of six risk categories and rolled into a single score and a band the desk can act on. The number is a risk-selection instrument, not a loss estimate.
Each failure is reproduced and written up with its severity: a refund issued without an entitlement check, reproduced 41 times in 4,000 runs; a customer record changed on an injected instruction; a price misstated under an ambiguous prompt. The transcript sits behind every one.
Reproduced failures become frequency. What the AI agent can do wrong, and to whom, becomes severity. Every assessment adds to the historical record behind the next one, so the numbers pricing needs get firmer over time.
Findings go to the insured with what fixes them. A re-score follows, and the underwriter binds with conditions instead of declining. Risk selection you can defend in a line with no loss history.
Everything around it.
The settings, outputs and controls that come with it.
Conditions
Suggested policy conditions tied to specific findings.
Scenario library
The scenarios run, by category, with pass and fail.
Transcript replay
Read exactly what the AI agent said and did.
Signal sources
Which outside sources contributed, and what they showed.
Frequency and severity
How often it fails, and how much a failure could cost.
Custom scenarios
Add the situations your line worries about most.
Mitigation loop
Findings go to the insured, fixes get re-scored.