The insurance market has always adapted to emerging risks after the fact. Asbestos liability was an afterthought until the claims started arriving. Cyber risk lived silently inside property and business interruption policies for years before the industry acknowledged it needed dedicated products. In each case, the delay between the emergence of the risk and the market’s response created a period of acute uncertainty, during which losses accumulated and coverage disputes multiplied.
Artificial intelligence is entering that same interim period, except the underlying exposure is broader, less predictable, and harder to assign to any single line of business.
The concentration problem no one is pricing
Much of the current conversation about AI risk focuses on what models do. The more immediate insurance question may be about who provides them. Enterprise adoption of generative AI has accelerated sharply, but the infrastructure supporting it remains remarkably concentrated. A handful of foundation model providers underpin applications across financial services, healthcare, legal, logistics and public administration. When a bank’s credit engine, a hospital’s triage system and a law firm’s research platform all depend on variants of the same underlying model, they share a common point of failure that none of their individual risk assessments is likely to capture.
This is not a theoretical concern. The cyber insurance market learned this lesson through CrowdStrike in July 2024, when a faulty update to a single endpoint security product caused global IT outages affecting airlines, hospitals and financial institutions simultaneously. The insured losses ran into the billions. That incident involved a conventional software failure. An equivalent event at the model layer, where the compromised system is not merely processing data but actively generating decisions, would introduce a dimension of loss that existing cyber wordings are not designed to address.
The classification problem
Traditional insurance operates through clearly delineated product lines. Cyber covers unauthorised access and data breaches. Professional indemnity covers negligent advice. Product liability covers defective goods. AI cuts across all of these simultaneously, and therein lies the problem.
Consider a financial advisory firm that integrates an AI model into its client-facing research. The model, following a corrupted update from the provider, begins producing subtly inaccurate market analysis. The firm’s advisers, who have come to treat the model’s outputs as reliable, distribute that analysis to clients. Losses follow. The firm’s professional indemnity insurer argues this is a technology failure, not negligent advice. The technology E&O insurer contends the firm failed to supervise its AI tools. The cyber insurer notes there was no data breach and no unauthorised access. Each policy has a plausible basis for declining coverage.
This is the phenomenon the market has begun calling “silent AI,” unintended exposure sitting inside policy wordings that were drafted without artificial intelligence in mind. Research published earlier this year estimated that over 90 per cent of insurers’ total AI liability exposure falls into this uncharted territory, embedded in conventional policies that neither expressly include nor expressly exclude AI-driven losses.
The limits of precedent
The industry’s instinct is to reach for precedent, and the most commonly cited is the silent cyber experience of the late 2010s. When the market recognised that property and casualty policies were inadvertently absorbing cyber losses, it responded with exclusions, affirmative wordings and standalone cyber products. The effort took the better part of a decade and, by most accounts, remains incomplete.
Silent AI is a materially harder problem. Silent cyber was largely contained within identifiable perils: unauthorised access, malware, network interruption. Silent AI touches every policy that covers the consequences of a decision, which is to say virtually every casualty class on a carrier’s book. A flawed AI output can simultaneously trigger professional indemnity, D&O, general liability and regulatory investigation cover. No single exclusion or endorsement can resolve that breadth of exposure.
The terrorism precedent may be more instructive, if less comfortable. After September 11 generated approximately $47 billion in insured losses, the private market did not innovate its way forward. It withdrew capacity. Governments intervened with backstop mechanisms, TRIA in the United States, Pool Re in the United Kingdom, GAREAT in France, because the private market concluded it could not price the tail risk. The question is whether AI systemic risk will follow the same trajectory, and whether policymakers are preparing for that possibility.
The regulatory vacuum
The legal infrastructure for allocating AI liability remains underdeveloped. The EU’s AI Liability Directive, which was intended to create harmonised rules for civil liability arising from artificial intelligence, was withdrawn by the Commission in 2025. Member states are left to apply existing fault-based frameworks, doctrines designed for human decision-making, to systems that operate through statistical inference and opaque optimisation processes.
In the United States, ISO introduced generative AI exclusion endorsements for commercial general liability policies in January 2026, but take-up has been uneven. Large portions of the market remain without clear AI-specific language. The result is that when the first major AI-related coverage dispute reaches the courts, judges will be interpreting policy wordings drafted for a technological environment that no longer exists.
What risk managers are missing
Most organisations approaching AI governance focus, understandably, on data protection, model bias and adversarial robustness. These are important. But they leave unaddressed the question that will matter most when something goes wrong at scale: who absorbs the financial loss?
The answer depends on insurance, and the insurance market currently lacks the tools to provide a reliable one. Few carriers have developed systematic processes for mapping AI model dependencies across their portfolios. Reliable actuarial models for correlated AI losses do not yet exist. The information asymmetry is significant: insurers often do not know which AI models their insureds rely on, how deeply those models are embedded in critical business processes, or what update and security practices the model providers follow.
For CISOs and risk managers, the practical implication is that insurance cannot be treated as a backstop for AI risk in the way it has been for conventional cyber risk. Until the market develops adequate assessment methodologies and dedicated products, organisations deploying AI at scale are carrying more residual risk than their insurance programmes reflect. Recognising that gap is the first step toward managing it.
