Map the AI risk before discussing cover
Refreshed September 30, 2026.
Start with a workflow and a loss. “Uses AI” is too broad to tell an underwriter what can go wrong. A drafting assistant, a clinical recommendation system, and an agent authorised to pay invoices expose different people, assets, and control boundaries. STABLE
The five questions
Section titled “The five questions”| Question | What a useful answer contains |
|---|---|
| What can the system read? | Data classes, identities, tenants, retrieval sources, permissions |
| What can it say? | Intended output, prohibited advice, audience, review requirements |
| What can it do? | Tools, delegated authority, spending limits, changes to external systems |
| Who relies on it? | Staff, customers, clinicians, downstream software; consequence of error |
| What can fail together? | Shared models, tools, identity services, policies, vendors, release channels |
These questions are an academy analysis framework, not a policy form or a regulatory test. STABLE
A practical loss map
Section titled “A practical loss map”The examples below are fictional teaching cases. Coverage depends on actual wording, allegations, causation, exclusions, and applicable law; the class shown is a conversation starter. STABLE
| Workflow | Failure path | Insurance conversation |
|---|---|---|
| Customer support | A fabricated entitlement is relied on by a customer | Tech E&O, contractual liability, reliance |
| Clinical preparation | A summary omits a contraindication; a clinician relies on it | Professional liability, Tech E&O, allocation between vendor and provider |
| Accounts payable | Retrieved instructions redirect a payment to an attacker | Cyber, crime, social engineering, agent authority |
| Claims triage | An unsupported classification influences an adverse decision | Claims practices, discrimination, review and appeal |
| Software delivery | An agent changes production settings beyond approval | Tech E&O, cyber, business interruption, deployment control |
| Hiring | An automated ranking disadvantages a protected group | Employment practices, discrimination, validation and oversight |
Security is one part of AI safety. Reliability, harmful advice, privacy, discrimination, misuse, human reliance, and accountability also matter. NIST’s voluntary GenAI profile supplies a broader risk vocabulary; OWASP’s current LLM and agentic resources focus on application and agent security. GROUNDED · NIST GROUNDED · OWASP Sources checked September 30, 2026: NIST GenAI profile, OWASP LLM initiative, OWASP Agentic Applications 2026.
Say what you know
Section titled “Say what you know”FACT: a named control ran on an observed path. INFERENCE: that control may interrupt a specified loss path. UNKNOWN: how much it changes portfolio loss experience. Keep those statements separate even when they appear in the same pitch. STABLE
Field drill
Section titled “Field drill”Describe one deployment in a sentence: “This system reads ___, produces ___, can act through ___, and is relied on by ___.” Then name the costly failure and the evidence needed to investigate it.
→ Agent authority and tool actions
Drill this page →42 bank questions stand behind what you just read. Check it while it’s warm.