Over the past ten days, the debate around AI exclusions has become considerably more consequential. Marsh's public opposition to broad exclusions, supported by Aon and WTW, was followed by comments from Chubb CEO Evan Greenberg suggesting that exclusions, separate coverage and more differentiated underwriting would become part of the market's response to AI liability. Other underwriting organizations have also argued that AI-related exposures can be addressed within existing insurance products rather than being removed wholesale.
I have followed these developments closely but deliberately avoided joining the public conversation until now. As the founder of Armilla, a company dedicated to underwriting AI risk, I have a direct commercial interest in the outcome. It would be easy to treat every exclusion as validation of the need for specialist AI insurance, just as it would be easy to characterize opposition to exclusions as a failure to appreciate the risks. Neither position adequately captures what is happening.
When we founded Armilla in 2023, we were (among) the first voices arguing that AI would fundamentally change the liability landscape and that the insurance industry was not adequately prepared. We anticipated many of the issues now being debated: exposure accumulating silently within existing policies, uncertainty over how conventional coverage would respond, and the eventual need for explicit underwriting and affirmative insurance solutions. At the time, these questions received relatively little attention from the broader insurance industry. Three years later, it is encouraging to see the leadership of the world's largest brokers and insurers publicly engaging with the same issues. We have spent those intervening years developing underwriting approaches and insurance products to address them. I welcome this debate, but having worked on these questions for years, I believe it deserves more than an immediate reaction to the latest headlines.
The fundamental question is not whether AI should be excluded. It is how the insurance industry can preserve meaningful protection for existing liabilities while developing the underwriting capabilities and coverage structures needed to address risks that AI is creating or materially changing.
The current debate began to intensify on October 5, when Marsh Risk CEO Nick Studer and Marsh Re CEO Dean Klisura wrote to insurance and reinsurance executives expressing concern about the adoption of broad AI exclusions. Aon and WTW publicly supported the concerns on October 8. Marsh CEO John Doyle and Studer further explained their reasoning during Marsh's October 15 earnings call, before Greenberg addressed the same question during Chubb's October 21 results.
My first observation is that the executives are not necessarily disagreeing over whether AI introduces risk. Virtually everyone recognizes that it does. The differences concern how substantially it changes established liability exposures, who should bear the uncertainty, and whether the resulting protection should be delivered through existing policies or dedicated coverage.
Marsh's position deserves careful consideration because Nick Studer's explanation on the October 15 earnings call was considerably more nuanced than the headlines suggested, with a few categories being defined.
I think this is a useful framework, particularly the distinction between the first two categories. Consider an architecture firm whose AI-assisted design contains a structural error. If the error results in a professional negligence claim, the firm's duty of care does not fundamentally change because it used AI rather than conventional design software. The insured peril remains professional negligence, and the conventional professional liability policy may already be structured to respond to such a claim.
A broad AI exclusion would potentially remove protection that the insured reasonably expected to retain. It could also introduce considerable ambiguity. If the architect used AI for preliminary research, but an employee subsequently reviewed and approved the design, how would the insurer determine whether AI caused the loss? Would incidental use of AI be sufficient? Would the exclusion apply if several decisions contributed to the outcome?
On this point I agree with Marsh: insurers should be cautious about withdrawing protection for an established liability solely because AI featured in how the insured performed the underlying activity. The precision of the exclusion, and the liability it is intended to address, make a material difference.
Aon has made essentially the same argument. Tracy-Lee Kus has emphasized how pervasive AI is becoming across business activities, questioning the practicality of blanket exclusions. WTW's Andrew Hill has approached the issue partly from a different perspective: commercial insurance must remain useful to its buyers. If an insurer excludes a substantial proportion of modern business activity, customers may question whether the policy still offers value.
There is an equally important reinsurance dimension. Dean Klisura's concern is that exclusions at the treaty level may not align with the policies issued by primary insurers. Where a direct policy responds but a reinsurer excludes the same exposure, the primary insurer can be left retaining a loss it expected to transfer. Reinsurance exclusions do not automatically change the insurer's obligations to its policyholder, but they can materially affect the economics and capacity available to support those obligations.
Where the broker argument needs further development is the assumption that an existing insurance architecture can accommodate every AI exposure through incremental modifications. That may be true for many conventional liabilities, but it does not, by itself, resolve how novel operational risks should be defined, priced or aggregated. Nor does the absence of an AI exclusion establish that a loss satisfies the policy's original insuring agreement.
Evan Greenberg was asked specifically about the debate between preserving current coverage, modifying existing forms and introducing standalone AI liability insurance.
His response was that AI liability is a growing exposure not adequately contemplated in traditional casualty underwriting, and that the industry must understand, manage and address it. He anticipated a combination of coverage clarifications, exclusions and separate AI-related liability products, supported by underwriting and pricing capable of distinguishing between insureds and risk types.
I think some of the reaction has placed too much emphasis on the word exclusions and not enough on what Greenberg said about underwriting. He was describing the need to move from implicit assumptions about exposure toward deliberate evaluation and risk selection. That is precisely what insurers are supposed to do, and it is a stronger argument than simply proposing to remove AI from existing contracts.
Consider two software companies with comparable revenues, customer counts and technology E&O limits. One provides a relatively low-consequence AI document summarization tool. The other operates autonomous agents that approve financial transactions, modify customer records and execute actions within enterprise systems.
Traditional underwriting indicators might suggest similar exposures. In reality, the severity, frequency and aggregation characteristics could differ substantially. The second company's systems may have operational authority, the ability to cause immediate financial consequences, and dependencies on upstream models or connected applications that are not visible through conventional classification and revenue-based underwriting.
The same distinction applies among businesses using AI internally. A healthcare organization deploying AI for administrative support is not assuming the same risk as one permitting autonomous clinical decisions. An AI-generated marketing recommendation and an AI-driven infrastructure control decision are not interchangeable exposures, even if both technically involve generative models.
Greenberg's call for more differentiated underwriting raises an uncomfortable question: what does good AI underwriting actually look like, and how many insurers know what to ask? Much conventional liability pricing relies on established exposure measures, observed claims and an understanding of how losses develop over time. AI presents a much thinner evidence base. Liability claims can take years to develop; the systems at issue may change substantially during the policy period; and it remains difficult to establish which technical or governance controls actually reduce insured loss. Asking whether a business has an AI policy, tests its systems or maintains human oversight is straightforward. Knowing what those answers reveal about claim frequency, severity or pricing is not.
At Armilla, we have argued from the outset that AI exposure needs explicit, differentiated underwriting, not simply a checkbox recording whether a business uses AI. The immediate challenge for the industry is to build and test defensible methods of assessing exposure and controls, using technical evaluation and disciplined assumptions while claims experience matures. Such methods must be refined against actual outcomes; neither conventional classifications nor unvalidated AI risk scores should be mistaken for evidence of insurability. The need for this expertise exists whether coverage is embedded or standalone.
A targeted exclusion may be a legitimate way to establish what the insurer has deliberately chosen not to assume, just as a dedicated policy may provide clearer coverage and limits for an insured seeking specialist protection.
I agree with Greenberg that AI liability requires more deliberate and differentiated underwriting. I also believe that both exclusions and standalone insurance have legitimate and important roles in the market's evolution. Where an insurer cannot adequately assess, price or support a particular exposure, a carefully defined exclusion may be appropriate, particularly where it creates the opportunity for that exposure to be underwritten and insured separately. Equally, standalone AI liability policies can provide dedicated capacity, clearer coverage and protection for exposures that existing insurance products may not adequately address. The question is not whether exclusions or standalone policies are inherently desirable or undesirable, but whether the chosen approach reflects a considered assessment of the underlying risk and provides clarity about what is—and is not—insured.
One of the more consequential questions arising from this debate is whether AI-related liabilities should remain within existing insurance products or migrate toward dedicated coverage. I believe this is a false choice. Much depends on the nature of the underlying exposure, the existing policy's coverage grants, and whether AI materially changes the risk that the insurer has agreed to assume.
Many AI-related claims will continue to arise under familiar legal doctrines, including negligence, professional liability, misrepresentation and intellectual property infringement. Where AI is used in the ordinary course of providing an established professional service, there is a strong argument for preserving protection within existing insurance products. But the fact that a claim falls within an established legal category does not establish that the underlying risk is unchanged. AI may materially alter the frequency, severity, speed and potential aggregation of losses, even where the legal basis of liability remains familiar.
The distinction is particularly important for businesses developing, deploying or assuming responsibility for AI systems. A professional using AI to assist with a routine task may present a very different exposure from an AI vendor whose system makes consequential decisions across thousands of customers. The latter may justify dedicated underwriting, additional limits, explicit coverage for particular liabilities, or a standalone insurance policy. Standalone coverage can also be an affirmative choice by an insured seeking dedicated capacity and contractual certainty, rather than simply a mechanism for replacing protection excluded from another policy.
Any serious assessment should acknowledge the commercial incentives of the organizations advancing these arguments. Insurance is a mechanism for risk transfer, but it is also a competitive market involving brokers, underwriting organizations, reinsurers and capital providers with different economic interests.
Recognizing those incentives is not the same as suggesting that the participants are acting in bad faith. In many cases, their commercial incentives are closely aligned with legitimate responsibilities to clients, policyholders or capital providers.
Brokers have a responsibility to preserve useful protection and competitive terms for their clients; broad exclusions can undermine both. The insurance cycle also matters. Marsh's Q3 2026 Global Insurance Market Index reported average global commercial rates down 5%, the ninth consecutive quarter of declines, although U.S. casualty rates rose 6%. In a broadly soft market, abundant capacity and competition strengthen broker leverage: an insurer seeking to impose sweeping AI restrictions must consider whether another carrier will offer more accommodating terms. The same wording proposal could encounter a different reception in a hard market. This does not invalidate the brokers' concerns, but it helps explain why those concerns carry particular commercial force now—and why current willingness to retain AI exposure should not be mistaken for proof that the exposure has been adequately priced.
Insurers, meanwhile, must pay claims, maintain reserves and protect capital. Long-tail AI liability is challenging precisely because material claims experience remains scarce and the underlying technology is changing faster than conventional actuarial observation periods. An insurer maintaining broad terms without adapting its underwriting may accumulate exposure it has not recognized. Yet imposing exclusions without distinguishing risk characteristics can be an equally blunt response. The real question is how much uncertainty an insurer can responsibly retain, on what assumptions, and with what ability to adjust as evidence develops.
Specialist underwriters, including Armilla, should subject their propositions to the same scrutiny. Their contribution should be demonstrated through the ability to differentiate risk, structure meaningful coverage and deploy underwriting expertise where it adds value. A standalone AI policy should not have to depend on the withdrawal of coverage elsewhere to justify its existence.
Conventional policies already respond to some AI-related losses, depending on their terms. Standalone policies can nevertheless be valuable in their own right, offering additional dedicated limits, clearer affirmative coverage and specialized assessment of consequential exposures. These solutions can coexist and compete on their merits.
The durable opportunity for specialist underwriting lies in making insurance more responsive to emerging AI liabilities, whether the protection is embedded in a familiar contract or provided through a dedicated policy.
Much of the public discussion has concentrated on whether a policy contains an AI exclusion. That is an important question, but not a sufficient one.
An insurance policy does not cover every loss that is not expressly excluded. Coverage begins with the insuring agreement, which specifies the covered events, liabilities or losses. The rest of the contract determines how that grant is defined, limited, conditioned or excluded.
A policy can contain no AI exclusion and still fail to respond to an AI-related loss because the circumstances do not satisfy the relevant coverage trigger.
Equally, a policy with affirmative AI language may provide less protection than the headline suggests if the language only confirms that the use of AI does not invalidate an otherwise covered claim, rather than extending coverage to additional forms of liability.
This is a critical distinction for brokers, insureds and underwriters.
Consider several hypothetical scenarios.
These are illustrative scenarios, not representations that particular policies will or will not respond. Coverage depends on the issued contract, applicable law and claim facts.
Suppose an engineering consultancy uses an AI assistant during the design of a building component. An error passes through the firm's quality-control process and results in a claim alleging negligent engineering services.
The insured's professional obligation has not necessarily changed. The consultancy still owes a duty to exercise appropriate professional care, and its policy may respond to the alleged negligence. If the insurer subsequently excludes any claim arising from the use of AI, it may introduce a significant gap into a traditional risk that the insured reasonably expected to transfer.
Now suppose a company authorizes an AI agent to approve invoices and execute payments within defined parameters. The agent misinterprets a legitimate instruction and transfers funds to the wrong recipient. No attacker has penetrated the network, no credential has been stolen, and the transfer was executed through permissions the company deliberately granted.
The resulting loss may not satisfy the insuring agreement of a conventional cyber policy if that policy requires a qualifying security event. It may also fail to meet the terms of a crime or funds-transfer policy, depending on its wording and whether fraud or another specified trigger occurred.
This is the strongest version of the specialist underwriting argument. The question is not whether the loss involved AI, but whether the insured event has been defined and affirmatively transferred.
Verisk's ISO program has introduced optional generative AI exclusions in general liability and products/completed operations. These include different endorsements affecting bodily injury, property damage and personal and advertising injury. Some exclude a wider range of exposures than others, and their effect depends on whether the insurer actually attaches the endorsement.
An August analysis of U.S. regulatory filings identified 4,078 carrier/program-form-state records through July 31, with 2,369 having reached their filed effective dates.
A carrier may file an endorsement without routinely issuing it. It may apply the exclusion selectively, use it only for particular classes, or offer alternative terms based on underwriting considerations. The underlying filing data does not establish the number of affected insureds, the premium subject to exclusions, or the amount of coverage actually withdrawn.
There is also a material distinction between exclusions being developed for commercial general liability and the affirmative AI treatment introduced by some cyber and technology E&O insurers. It is inaccurate to describe the entire commercial insurance industry as either broadly excluding or comprehensively embracing AI.
That unevenness makes the wording itself increasingly important. Definitions of AI, causation language, the scope of excluded damages, exceptions preserving coverage, and the relationship between underlying and excess contracts can produce materially different results.
We recognized this pattern early in Armilla's development. Drawing on the evolution of cyber insurance, our thesis was that AI liability would initially accumulate largely unexamined within conventional insurance policies. As insurers began to recognize the nature and scale of these exposures, we expected a gradual shift toward explicit underwriting, coverage clarifications, exclusions and dedicated capacity. Importantly, we anticipated that standalone AI liability would emerge as a significant specialty insurance class in its own right, supported by differentiated underwriting, pricing and risk assessment capabilities. This was not simply a prediction that traditional insurers would exclude AI. It was a belief that the nature of AI-related liabilities would eventually justify a dedicated insurance market.
The developments of the past few weeks are consistent with that thesis, although the eventual market structure remains unsettled. We never assumed that every AI-related loss would migrate out of conventional policies, just as cyber insurance did not eliminate cyber-related coverage across other insurance classes. Existing policies will continue to play an important role, while standalone AI liability will provide dedicated limits, clearer protection and specialized underwriting for insureds that need it. The question now is not whether these approaches can coexist, but how quickly the market develops the underwriting expertise, products and capacity to support them.
The diversity of AI-related liabilities does not weaken the case for a standalone insurance class. If anything, it may strengthen it. Cyber insurance evolved into a distinct specialty market despite encompassing several different forms of loss, including first-party financial damage, third-party liability, privacy breaches and business interruption. The common thread was not a single legal cause of action, but a set of technology-related exposures that benefited from dedicated underwriting, affirmative coverage and specialized capacity.
AI liability could follow a similar trajectory. As AI systems become more autonomous and their decisions increasingly consequential, insureds may benefit from dedicated policies that bring together liabilities otherwise scattered across conventional insurance products. Standalone coverage can provide clearer contractual protection, dedicated limits and underwriting that reflects the characteristics of AI systems rather than relying exclusively on traditional industry classifications.
There is also an important distinction between a new legal category of liability and a new underwriting exposure. A claim may continue to be pleaded as negligence, misrepresentation or breach of contract even as AI changes its frequency, severity or accumulation potential. An insurer does not need a completely new legal cause of action to justify differentiated underwriting or a dedicated product; nor does the availability of standalone coverage mean that conventional policies cease to serve their established purpose.
AI accumulation is a credible exposure mechanism, but its insured-loss scale is not yet established. Many unrelated businesses use common foundation models, infrastructure or orchestration tools, creating the possibility of correlated failures. Yet there is no mature claims history from which to estimate the frequency, severity or accumulation of such events. Present AI adoption also varies materially: many uses remain assistive rather than autonomously responsible for consequential operations, although a system need not be revenue-critical to cause a serious liability claim.
A shared-model failure could give rise to claims under several liability lines, but the precise chain from technical failure to covered insured loss depends on contracts, duties, reliance and policy triggers. The existence of a common technological dependency should therefore be treated as an exposure indicator, not as evidence that a large insured catastrophe is inevitable.
The prudent underwriting response is to map meaningful dependencies and concentrations, run plausible scenarios, and update assumptions as operational criticality and claims experience grow. Both insurers and reinsurers need visibility into shared systems across portfolios; they should neither dismiss accumulation as speculative nor present worst-case outcomes as an established loss distribution.
Targeted exclusions can be a prudent component of the response when an insurer cannot reasonably support an exposure with available information or capital. But they should complement, not replace, better identification and management of the exposures the insurer elects to retain.
The public debate is often presented as a choice between two mutually exclusive positions: preserve existing insurance coverage or recognize AI as a distinct liability exposure requiring specialist solutions.
I do not think those positions are mutually exclusive.
I believe exclusions have a legitimate and important role in how the insurance market responds to AI. Insurers must be able to define the risks they are prepared to assume, particularly where emerging exposures introduce uncertainty around severity, accumulation or insurability. In some circumstances, excluding AI-related liabilities from conventional policies may be the appropriate underwriting decision, allowing those exposures to be evaluated, priced and potentially insured through dedicated coverage.
Equally, the scope and wording of exclusions matter. An exclusion that distinguishes a specific emerging exposure from an established liability serves a different purpose from one that broadly removes coverage whenever AI has played a role in a loss. The objective should be clarity over what is insured, deliberate risk selection and appropriate pricing—not a presumption that all AI-related losses must either remain within existing policies or be excluded from them.
Equally, established legal doctrines do not eliminate the case for dedicated AI underwriting or standalone coverage. How a claim is legally pleaded and how its underlying frequency, severity and accumulation are evaluated are distinct questions. Dedicated insurance can be desirable on its own merits, not only as a remedy for an existing coverage gap.
I believe the market will increasingly distinguish among three broad categories of exposure. They are not rigid or exclusive categories; rather, they illustrate why both existing insurance and a substantial standalone AI liability market can develop in parallel.
For AI vendors and organizations relying heavily on autonomous systems, the analysis becomes more complex. The relevant issue is whether the scale, consequences and contractual allocation of AI-driven decisions materially change the risk being insured, and whether a dedicated product offers clearer protection or additional capacity.
For exposures involving unusual contractual commitments or substantial accumulation, specialized structures may be effective. Conventional policies may not have been designed to cover guaranteed model performance, a vendor's assumed indemnity obligations or correlated losses arising across many insureds.
I expect AI liability insurance to evolve through several complementary structures rather than converge on a single product. Standalone AI liability has the potential to become a significant specialty insurance class, providing dedicated limits, affirmative protection and underwriting tailored to emerging AI exposures. At the same time, traditional liability, cyber and professional insurance products will continue to address many AI-related risks, sometimes through explicit extensions or more comprehensive combined coverage.
Businesses adopting AI need confidence that conventional liability insurance does not become unexpectedly ineffective. AI vendors need the ability to stand behind their products and contractual commitments. Their customers need to understand what remedies are available when AI systems fail. Investors, boards and regulators need assurance that organizations can manage and finance the consequences of deploying systems with increasing autonomy.
This is why the emerging discussion around governance, testing and certification is important. Better underwriting should create economic incentives for safer AI development and deployment. Insurers need to distinguish systems with meaningful oversight and risk controls from those with inadequate monitoring, uncontrolled autonomy or poorly understood failure modes.
There are difficult questions that remain unresolved. How should insurers measure concentration around shared models? What evidence is sufficient to distinguish one AI deployment from another? How should responsibility be allocated between model developers, application providers and enterprise customers? How should insurers price liability when claims experience is limited and system capabilities evolve rapidly? What should happen when a business accepts contractual obligations beyond the scope of conventional negligence-based liability?
These are questions the industry needs to address collectively. Brokers understand the needs of insureds and the structure of existing insurance programs. Traditional insurers bring established underwriting capacity and claims experience. Reinsurers bring important perspectives on aggregation and capital. Specialist underwriters can contribute deeper technical evaluation of emerging AI exposures.
The strongest insurance solutions will likely emerge from combining those capabilities rather than insisting that one institutional perspective has the complete answer.
The past ten days have been important because the debate over AI liability is no longer confined to specialist insurance providers or discussions about hypothetical future losses. The CEOs of major brokers and insurers are now publicly addressing how AI should be treated across established insurance markets. That is a welcome development: each side is raising issues the other cannot afford to dismiss.
Marsh is right that broad exclusions could undermine the usefulness of insurance and withdraw protection for liabilities businesses have traditionally insured. Aon is right that the pervasiveness of AI makes sweeping restrictions increasingly impractical. WTW is right to recognize both the commercial importance of preserving coverage and the reality of systemic exposure.
Many AI-related claims will continue to arise under established legal doctrines and can appropriately remain within existing insurance products. But this does not diminish the case for specialist or standalone AI liability coverage. Even familiar forms of liability can exhibit materially different risk characteristics when AI is involved, while other exposures may require dedicated coverage grants, limits and underwriting capabilities that conventional insurance does not provide.
Chubb is right that AI introduces emerging exposures that require deliberate underwriting, differentiated pricing and stronger evaluation of systems, governance and controls. Existing classifications and historic loss experience will not always be sufficient.
I believe the most defensible approach is to preserve meaningful protection for familiar liabilities where appropriate, use precisely drawn exclusions where insurers elect not to assume a particular emerging exposure, develop explicit affirmative coverage where contracts have genuine gaps, and recognize the independent value of standalone policies with dedicated limits and specialized underwriting.
A mature market can support embedded protection, combined AI/cyber/technology liability products and a substantial standalone AI liability class. These forms can address different buyers and exposures, and sometimes coexist within the same insured's overall program. Their effectiveness depends on clarity, coordination and sound underwriting, not uniformity of product structure.
The future of AI insurance will not be determined by exclusions alone. It will be shaped by our ability to define exposures, differentiate risks, build appropriate capacity and provide effective insurance both within existing policies and through standalone products.
That is where I believe the next phase of this market will be defined, and where our collective effort should be directed.
Disclaimer: The insurance products described on this website are offered through surplus lines insurers. Surplus lines insurance is not subject to all the insurance laws and regulations of your state, and such insurers are generally not licensed by or subject to the supervision of the insurance department of your state. Coverage may not be available in all jurisdictions and is offered only through properly licensed surplus lines brokers.
Nothing on this website constitutes an offer to sell or a solicitation to buy insurance. No coverage will be bound or effective unless and until confirmed in writing by an authorized representative and all underwriting requirements have been satisfied. Any descriptions of coverage are for general informational purposes only and do not amend, modify, or supplement any actual policy or contract of insurance.
Please consult with a licensed surplus lines broker to determine availability and eligibility in your state.
Armilla Insurance Services is a Coverholder at Lloyd's. Affirmative AI Liability Insurance is underwritten by certain underwriters at Lloyd's. Not all activities are related to Lloyd's.