Insurance Carriers Are Starting to Decide Who Can Deploy AI at Scale

A growing share of insurance carriers has decided that AI-related losses are simply not their problem. W. R. Berkley, a Fortune 500 commercial insurer, added blanket AI exclusions to its directors and officers, errors and omissions, and fiduciary liability products in late 2025, CEO W. Robert Berkley said on the company's Q4 2025 earnings call. The exclusions bar coverage for any "actual or alleged use, deployment, or development of Artificial Intelligence."

A RAND Corporation report published this year by researchers Sasha Romanosky and Celine Robinson maps a three-way split in the US market: a minority of carriers now affirmatively cover AI losses, a growing number file broad exclusions, and a majority remain silent. The silent middle is where most corporate policyholders actually sit today, and it is also where the worst exposure is hiding.

Standardized forms are doing the heavy lifting. Verisk's Insurance Services Office released three optional generative-AI exclusion endorsements in January 2026: CG 40 47, CG 40 48, and CG 35 08. They cover bodily injury, property damage, and advertising injury arising from AI outputs. ISO's standardized forms appear in more than 80 percent of US property and casualty policies, so whenever a carrier adopts these endorsements, the exclusion reaches almost every commercial line it writes. RAND flags AIG, Great American, and Chubb as carriers moving toward similar positions, though adoption still varies by state and product line.

What is actually being excluded? AI-related losses now show up in at least 11 insurance lines, including technology errors and omissions, professional liability, cyber, D&O, and commercial general liability. Berkley's language covers any "use, deployment, or development" of AI, not just specific model failures or product defects. A company shipping a customer-facing chatbot, an internal summarization tool, or an agentic workflow can find itself outside the policy language without a single loss having occurred.

Underwriters are not doing this because AI losses have spiked. The Artificial Intelligence Incident Database lists 713 incidents drawn from more than 6,000 reports, and roughly 250 US AI-related lawsuits are still working through the courts. Legal and regulatory shock is the most concrete of the five accumulation mechanisms RAND names: dozens of US state laws already cover AI-generated intimate images, automated decision-making, and AI in political advertising, creating loss categories that pre-2020 commercial general liability policies were not drafted to anticipate. Berkley framed the broader question on the earnings call as whether underwriters can "fully understand that risk to control, select, and price for it." For many carriers, the answer right now is no.

There is no hearing, no statute, and no regulator at the center of this shift. Insurance carriers are starting to act as the de facto gatekeepers of enterprise AI. The mechanism is a one-paragraph endorsement that narrows what counts as a covered loss without ever appearing on a public agenda. A September 2025 PYMNTS Intelligence survey of US enterprise CFOs found about 7 percent had deployed agentic AI in live finance workflows and another 5 percent were running pilots. For those companies, the planning question is no longer whether their AI product works. It is whether their insurance policy still applies when it does not.

RAND recommends that state regulators and the National Association of Insurance Commissioners develop a standardized AI Coverage Notice that would require carriers to declare, line by line, whether AI losses are covered, excluded, or silent. Until that arrives, the safest assumption for any company deploying AI at scale is that the silent middle of the market is where the exposure lives, and that contracts, indemnity language, and AI-specific coverage riders are now part of the deployment checklist rather than the post-incident paperwork.