One AI Failure, Four Policies: AXA XL Names The Scenarios

Estimated reading time: 7 minutes

A carrier has put the question in writing. AI-related losses “may not fall neatly into one category of risk.”

AXA XL and S-RM published Building Resilient AI on 23 September. The report names four loss scenarios that will test that sentence. A deepfake-enabled payment fraud. An AI-related data breach. A defective output causing third-party loss. An outage at a critical AI provider. Each, the report says, engages different policies. AI risk insurance has no single home.

Why AI Risk Insurance Does Not Sit In One Line

The four AI loss scenarios belong to AXA XL. The report stops short of naming which part of a program each one reaches. Any mapping beyond that point belongs to the reader.

That restraint is itself the finding. A carrier can describe four ways AI produces a loss without saying, in public, which policy pays for each.

Rebiah Bardot-Girard, head of cyber risk consulting services at AXA XL, described the mechanism. “AI risk rarely emerges in isolation,” she said. It amplifies weaknesses in “identity management, data governance, supplier oversight and incident readiness.”

The report goes further in its own voice. It argues that AI should count as an enterprise resilience issue. That issue connects cyber security, AI governance, third-party dependency, operational continuity, regulatory accountability and insurance.

Where AI Liability Coverage Is Heading

AXA XL is the fourth market participant this month to place AI risk insurance questions outside the cyber tower.

Beazley confirmed affirmative AI wording in its cyber and technology errors and omissions policies. That wording addresses AI-driven attacks. RAND reported that most carriers remain silent on AI exposure. Envelop Risk then hired a global head of casualty. The firm said the move would improve its view of how AI risks reach insurers across whole portfolios.

Cyber Insurance News reported earlier that AI exclusions were spreading faster outside cyber policies than inside them. AXA XL’s four scenarios describe why.

The Evidence Problem Behind Every AI Claim

The report’s sharpest passage concerns what happens after a loss.

Joani Green, global technical director for cyber security at S-RM, set out the difficulty. “Traditional logs may not be enough to understand what happened,” she said. The gap widens wherever prompts, outputs, model behavior, retrieval sources or agent actions enter the picture.

The executive summary states the consequence plainly. Organizations that keep no record of prompts, responses, and tool calls lose investigative confidence. Traditional IT events still permit it. AI incidents do not.

An adjuster works from evidence. A forensic firm reconstructs a timeline. Neither task has a settled method when the acting party was a model.

That question runs through our podcast conversation with Mike Nelson of DigiCert on content provenance and claims fraud. It also shapes the episode on MFA and the claim that failed anyway. There a control sat on the application form and proved absent at the moment of loss.

Green framed the practical test. Organizations should work out what they would need to reconstruct an AI-related event. That record must also “support legal, regulatory or insurance processes.”

Take The Cyber Insurance News Survey

Respondents get the results first.

What The Numbers Behind AI Risk Insurance Show

The report synthesizes external research rather than running its own survey.

McKinsey’s 2026 State of AI puts adoption at 88 percent of organizations, up from 78 percent a year earlier. That figure covers AI use in at least one business function. The same research finds 23 percent already scaling agentic AI. A further 39 percent are experimenting with it.

AI governance trails that curve. Deloitte finds only one in five organizations holds a mature governance model for autonomous agents.

One figure deserves attention from anyone pricing professional liability. Stanford’s 2026 AI Index measured hallucination rates across 26 leading models. The range ran from 22 percent to 94 percent.

Pre-deployment scrutiny has improved. The World Economic Forum records 64 percent of organizations assessing AI tool security before deployment. The previous year’s figure was 37 percent. CINI covered the same WEF research on cyber-enabled fraud overtaking ransomware earlier this year.

The report notes that pre-deployment assessment is not enough, because exposure continues throughout deployment and operation.

The AI Governance Risks The Report Names

Data leakage and unauthorized access lead the list. AXA XL cites IBM’s X-Force research finding more than 30 percent of investigated incidents involved theft or misuse of credentials. That marks a shift from software exploitation toward identity-based attack.

Data poisoning follows, occurring during training when attackers inject corrupted fragments to steer outputs or install backdoors. The report warns that such manipulation resists detection after deployment and often requires full retraining on clean data.

Prompt injection, model manipulation and model theft, hallucinations, shadow AI and autonomous agents complete the set. On shadow AI, Green identified the pattern. Formal governance often surrounds strategic AI programs while unmanaged use continues through public tools and embedded software features.

Our earlier reporting on agentic liability covered the last of those directly.

What Underwriters Can Ask

AI risk insurance decisions rest on evidence. The report’s five priorities convert into questions with documentary answers.

Can the insured produce an inventory of its AI deployments? That inventory should name what data each system reaches and where it can take or influence action.

Does the organization retain prompts, responses, and tool calls in a form that would support a claim investigation?

Which AI vendors sit in the critical path, and have they passed third-party due diligence?

Bardot-Girard named the first of those as the starting point. A practical inventory of AI use, data access, and action authority is “now a fundamental starting point for resilience.”

Methodology And Disclosure

Building Resilient AI is a risk-consulting publication rather than original research. It draws on McKinsey, the World Economic Forum, IBM, Stanford, Deloitte, the OECD and OWASP. AXA XL and S-RM practitioners supply the commentary.

AXA XL sells both risk consulting and insurance. It holds a commercial interest in how the AI risk insurance conversation develops. S-RM is a commercial intelligence and cyber security consultancy. The report carries an explicit disclaimer that nothing in it indicates the existence or availability of coverage under any policy.

Jonathan Salter, head of risk consulting at AXA XL, set the frame. AI is moving into everyday systems, “but governance is not always keeping pace,” he said.

FAQ – AI Risk Insurance

What does the AXA XL and S-RM report say about AI risk insurance?

It states that AI-related losses may not fall neatly into one category of risk, and names four scenarios that engage different policies: deepfake-enabled payment fraud, an AI-related data breach, a defective output causing third-party loss, and an outage at a critical AI provider.

Does the report say which policy responds to each scenario?

No. It names the scenarios and says they engage different parts of an organization’s risk and insurance program. It does not specify which line of cover applies to each.

Why is evidence a problem in AI claims?

S-RM’s Joani Green notes that traditional logs may not capture prompts, outputs, model behavior, retrieval sources or agent actions. Without those records, AI incidents cannot be investigated with the confidence traditional IT events allow.

How widely is AI being used?

McKinsey’s 2026 State of AI puts adoption at 88 percent of organizations in at least one business function, up from 78 percent. Twenty-three percent are scaling agentic AI and 39 percent are experimenting with it.

How reliable are AI outputs?

Stanford’s 2026 AI Index measured hallucination rates across 26 leading models ranging from 22 percent to 94 percent. The report argues outputs should be verified rather than assumed accurate.

Is this original research?

No. The report synthesizes work by McKinsey, the World Economic Forum, IBM, Stanford, Deloitte, the OECD and OWASP, with commentary from AXA XL and S-RM practitioners.

Related Cyber Insurance Posts

• Eighty-Eight Percent Have A Plan. Twenty-One Percent Have An Inventory.
• Large Cyber Claims Surge: AXA XL Study Finds 88% of Losses Driven by Major Attacks(Opens in a new browser tab)
• AI Risk Reshapes Cyber Insurance: Key Takeaways From Lockton Re and Armilla’s “Ready or Not”(Opens in a new browser tab)
• Tackling AI-Driven Cybercrime Threats: AXA XL’s New Insurance Coverage Explained(Opens in a new browser tab)
• AXA XL S-RM Acquisition Adds Forensics And Incident Data To Risk Advisory(Opens in a new browser tab)