Recently, I had the pleasure of joining a panel at the Bermuda Captive Conference, entitled “Not Another AI Panel: What Captive Owners Need to Know Now.” The premise of the session was to skip the hype that seems to always surround artificial intelligence (AI). Instead, we dug into the practical questions companies are actually facing right now. They’re the same ones each of us on the panel is already being asked about AI in our day-to-day work.

The panel had a diverse mix of perspectives. I shared the stage with Thomas Galbraith, CEO and Co-Founder of Barkr, whose business revolves around AI-powered, insurance-backed valuations for hard assets; Sarah Hopkins, an attorney at Fenwick & West, with specialization in Fintech and Insurtech regulatory solutions; and Federico Candiolo, Senior Corporate Counsel at ASW Law. The session was moderated by Troy Dort, Director, IT Advisory at KPMG in Bermuda.

Through a series of polling questions to the audience, which comprised captive owners and captive insurance professionals, nearly everyone in attendance responded affirmatively that they use AI in their personal lives. A great majority stated they were using it at work. Yet, when we queried the audience, not a single person claimed to be aware of having explicit insurance coverage for AI liability. In fact, the audience posed questions such as, “What even are the risks and perils originating from the use of generative AI?” and “Isn’t that already covered in my existing policies?”

Well, the risk is real, and the commercial market is already reacting to it.

Perils associated with generative AI usage

• Defamatory or disparaging content produced by AI
• Intellectual property infringement arising from AI-generated output that reproduces or closely mirrors protected work
• Intellectual property infringement arising from protected works used to train a model, even if no output reproduces them
• Bodily injury or physical damage resulting from reliance on errors in AI recommendations or instructions
• Model hallucinations that lead a third party to financial loss
• Data breach where protected or confidential information is exposed via AI outputs, even absent any traditional breach or intrusion
• Algorithmic bias in decisions like hiring or underwriting

Exclusions in the standard market

In January 2026, the Insurance Services Office (ISO) introduced three new optional commercial general liability (CGL) endorsements: CG 40 47 (a broad exclusion under both Coverage A for bodily injury/property damage and Coverage B for personal and advertising injury), CG 40 48 (narrower, excluding only Coverage B), and CG 35 08 (which attaches to the Products/Completed Operations Liability Coverage Part, excluding bodily injury or property damage arising out of generative AI embedded in a product or completed work). These exclusions could lead to a lack of coverage within CGL coverage for claims generated from AI outputs. The definition of “generative AI” is strikingly broad and can be wide enough to include any systems that produce text, images, audio, video or code. That means the generative AI perils and coverage exclusions don’t just affect AI companies. The broad definition of “generative AI” could include companies using AI to draft marketing materials, screen resumes, or run a customer service chat bot, to list a few examples.

Beyond CGL, carriers are introducing parallel AI exclusions to address the silent-AI exposure across cyber, E&O, D&O and EPLI forms. Rates for those existing coverages did not contemplate this emerging source of liability that the use of generative AI can create. As the standard market excludes generative AI exposure from these lines of business, that risk doesn’t disappear. The risk lands on the balance sheet of the company that put the tool to use.

Those AI-centric companies whose core business is building their own AI models face a more acute reality. Their exposure isn’t on the margins, but a primary insurance need for their core business function. A handful of carriers have developed insurance products for this industry, but, as with any immature and emerging insurance market, insureds may be stuck paying high premiums or with partially uninsured exposure as the market matures.

Where captives fit in

Both cases are textbook examples of where captive insurance solutions can be beneficial. This is one of the reasons the panel thought the discussion would be compelling at the Bermuda Captive Conference. The captive becomes the vehicle for a sophisticated owner to retain and study the risk rather than paying the standard market’s risk load or going bare. Whether it’s drop down coverage to fill the gaps left from AI exclusions in standard commercial policies, or taking a primary line on the risk itself, the captive can act as the R&D lab by retaining exposure, gathering loss experience and building history. This gives captive owners the information they need to price the risk confidently and accurately, potentially ahead of the commercial market, depending on the nature of the risk. Putting novel risks like AI liability into a captive must be done with prudence. Understanding the appropriate per-claim and aggregate limits the captive should write is a crucial piece to effectively using captive insurance as part of the solution. The captive should not expose itself to the volatility of large limits, but should instead operate in the working layer of the loss distribution to effectively fund and understand the risk.

Actuaries pricing AI risk are faced with the reality that an emergent coverage like this breaks several of the assumptions that classical ratemaking relies on. There is little to no credible loss history, the expectation of a heavy tail to the loss distribution, and potential correlation of exposure to the extent that several points of a company’s supply chain or vendors may rely on the same few underlying foundational AI models. Correlation hidden in the foundational AI models breaks the independence assumption that underpins risk distribution. The actuarial approach shifts towards exposure rating including risk margins. Even non-insurance data sets gathered to help estimate the underlying fundamental drivers of loss activity can be a crucial tool in estimation. Information can also be garnered from adjacent perils to be used as proxy in situations like media/E&O for defamation and hallucination, or cyber for data disclosure. The approach starts with understanding the AI liability exposure needing to be priced, choosing an exposure base and building towards estimates of frequency and severity based on the relevant information.

The challenging part is that the pricing solution isn’t an off-the-shelf, one-size-fits-all approach. It’s bespoke to each individual company and should reflect the underlying AI usage that is driving the exposure.

The closing remarks of the panel discussion in Bermuda revolved around discipline. AI is a powerful technology, and developing or deploying it can fundamentally change what a business is capable of. A comparable level of thought and attention needs to be given to compliance, and governance surrounding these tools must be adopted with their use. That same discipline extends to the insurance industry, which must underwrite and price a risk that is unfolding as fast as the market has ever seen.