AI's Insurance Gap Is Becoming a Constraint on Adoption
Carriers are writing AI exclusions into core liability forms just as enterprise deployment accelerates
Photo by Andrew Dawes on Unsplash
Insurers are carving AI out of standard liability policies as safety incidents pile up. The coverage gap is now a real brake on enterprise deployment.
The backstop is thinning
A run of AI safety incidents is exposing a hole under the industry's feet. Coverage is scarce, and it's getting scarcer. Major carriers including AIG, W.R. Berkley and Great American have sought regulatory clearance for exclusions that let them deny claims tied to the use or integration of AI. Berkley's version is written for D&O, errors and omissions, and fiduciary products, and it bars claims involving any actual or alleged use of the technology.
The standard forms are moving too. A generative AI exclusion for commercial general liability, ISO form CG 40 47 01 26, became available for use in January 2026. A company renewing its program this year may find AI risk sitting in a different box than it did twelve months ago, whether excluded, sublimited or pushed into a separate endorsement.
Why underwriters are flinching
The issue isn't any single loss. Arup's $25 million deepfake fraud and Air Canada's forced refund from a chatbot were survivable. The fear is correlation. One model failure can hit thousands of clients at once, which is the aggregation scenario insurers price worst. Chubb has agreed to cover some AI risks while carving out widespread incidents of exactly that type.
Attribution adds a second problem. The same loss can look like a cyber event, a professional services failure, a media claim or a product defect, and legacy wording was not drafted for that overlap. Underwriters have lived through this before. The absolute pollution exclusion of the mid-1980s followed asbestos and environmental losses that nobody had priced. Cyber went through a similar hardening in 2019 to 2021, when carriers moved to strip out silent coverage and reprice what remained.
What it means for the capex story
The AI buildout has been funded on the assumption that enterprise demand converts into durable spending. Deployment at scale, especially autonomous agents that execute code and make decisions, requires risk transfer. Boards and general counsels will slow rollouts that can't be insured, and lenders financing data center and software exposure will ask who pays when a model fails.
That is a friction, not a verdict. Affirmative AI products exist, and a small specialty market is growing. The tell will be pricing and limits. If capacity stays thin and retentions rise, the cost lands on adopters as higher self-insurance, which is a margin headwind rather than a demand collapse.
Where the analogs point
The post-9/11 terrorism market is the closest precedent for a risk the private market declined to carry alone. Carriers pulled back, and a federal backstop arrived in 2002 through TRIA. Nothing comparable exists for AI, and the policy appetite for one is unproven. The cyber analog is more likely, with a painful two to three year repricing followed by clearer forms and a functioning market.
The counterargument is that the loss data is still thin. Exclusions can be an opening bid ahead of negotiated affirmative cover, and insurers have every incentive to sell protection once they can model it. The sector read is mixed for now. Specialty carriers with AI-specific products have optionality, while broad liability writers are choosing to shrink exposure. The [Sector Rotation dashboard](/sector) is the place to see whether financials and insurers start trading differently from the AI complex.
What to watch over the next 2 to 4 weeks
Watch state regulatory filings for new AI exclusions, particularly any that reach cyber or property forms. Third quarter commentary from large carriers should show whether AI is being discussed as an underwriting risk or a premium opportunity. Look also for any first sizable claim dispute that tests whether an AI loss is covered under an older policy.
The setup changes if a federal or reinsurance backstop gets proposed, or if a single correlated failure produces a headline loss. Either would reset how the market prices the risk. Until then, the insurance gap is a slow constraint on AI adoption, and it deserves more attention than this week's data prints.