From silent exposure to affirmative wording
Beazley in September 2026 announced it had expanded its cyber and technology errors and omissions proposition with endorsements addressing risks created by a company’s own use of artificial intelligence. The move is important because many policyholders have relied on existing cyber, business interruption or professional liability clauses without knowing whether a loss caused by an AI system would meet the trigger. Affirmative wording can reduce uncertainty, although the detailed terms, limits and exclusions remain decisive.
Voluntary shutdown is a distinctive trigger
One endorsement is intended to address certain costs arising when an insured voluntarily shuts down an AI system. That scenario differs from conventional cyber interruption, where malicious code or an external security event may disable technology. An AI application may still be functioning technically while producing unsafe, discriminatory, misleading or commercially damaging outputs. Management may need to stop it before a regulator, customer or court requires action.
Defining the insured event will be difficult
Coverage will depend on definitions: what qualifies as an AI system, who is authorised to order shutdown, what evidence shows a material risk and which costs begin at the decision point. Policies will also need to distinguish model error, poor data, negligent configuration, unauthorised employee use and third-party platform failure. Overlap between cyber, technology E&O, professional indemnity, product liability and directors’ and officers’ insurance may create allocation disputes if wording is not coordinated.
Insurance cannot replace AI governance
An insurer will expect the policyholder to maintain an inventory of systems, owners, data, suppliers and use cases. High-impact applications should have testing, human oversight, access controls, monitoring and an incident plan. A shutdown guarantee is most credible when the company has predefined thresholds, decision rights and fallback procedures. Otherwise, insurers may struggle to distinguish prudent mitigation from a discretionary commercial decision or an avoidable failure of governance.
Questions for buyers and boards
Buyers should ask whether cover applies to internally developed systems, embedded vendor tools and unsanctioned employee applications. They should test scenarios involving hallucinated advice, discriminatory decisions, corrupted data, intellectual-property claims and an agent taking an unauthorised action. Key questions include waiting periods, sublimits, notification duties, forensic costs, restoration, lost revenue and the interaction with contractual indemnities from technology providers.
A market likely to become more explicit
Beazley’s move suggests that AI risk is progressing from broad discussion to product design. More insurers are likely to develop affirmative grants, exclusions or questionnaires as claims experience develops. That should improve clarity, but it may also fragment coverage across policies. Risk managers and brokers will need to map the full AI loss chain and negotiate coordinated wording. The most valuable outcome may be the discipline created before placement: knowing which systems matter, when they must stop and how the business continues safely.
How to prepare for underwriting and claims
Policyholders can prepare a concise AI risk schedule before renewal. It should list material systems, business purpose, model or vendor, data categories, decisions influenced, human review, maximum tolerable outage and fallback process. For each system, the company should identify realistic loss scenarios and the policies that might respond. Brokers can then compare definitions, triggers and exclusions across cyber, technology E&O and other covers. The same schedule will help after an incident by showing ownership, prior testing and the basis for a shutdown decision. Claims protocols should specify when the insurer must be notified and what evidence should be preserved, including prompts, outputs, logs, model versions and human approvals. This preparation reduces ambiguity while giving underwriters a clearer basis for pricing and risk improvement. It can also reveal uninsured operational costs before a loss occurs.
What to watch next
The market will learn from claims, regulatory action and how policyholders use voluntary shutdown provisions. Buyers should expect wording to change as insurers define acceptable controls and accumulate loss experience. Early purchasers can influence that development by presenting clear scenarios and asking precise questions. A transparent dialogue is more likely to produce useful cover than treating AI as an undefined extension of an existing cyber programme.
Share this:
- More [2]
