Signals and Slips: Brokers put AI uncertainty on the balance sheet

By George Abbott Oct 14 (The Insurer) — Marsh Re’s push against AI coverage gaps has brought a growing tension into focus. Brokers want to preserve the widest possible market for cover, while underwriters need to understand the risks they are assuming as the technology — and the way it is used — continues to evolve. Signals and Slips is The Insurer’s series covering AI as both a technology and a risk factor for the commercial, specialty and reinsurance markets. We examine how the industry uses AI, how its slip-ups can be insured and what the technology means for different classes of business.

Each issue combines an opening analysis with a round-up of recent reporting. In letters sent ahead of the CIAB Insurance Leadership Forum, Marsh Re president and CEO Dean Klisura called for reinsurance to mirror primary coverage for routine AI applications. Where insurers already exclude certain AI risks, he argued, reinsurers should follow those terms without adding restrictions that create gaps. Marsh Risk president and CEO Nick Studer was equally direct.

The broker “cannot support broad or absolute AI, or generative AI, exclusions” that leave protection gaps, he said, warning that clients would be encouraged to seek alternative terms and markets. Aon and WTW have voiced similar concerns. Aon’s Tracy-Lee Kus pointed to how widely AI is already used, while WTW’s Andrew Hill warned that broad restrictions could undermine the value of cyber insurance and drive buyers elsewhere. The brokers’ commercial incentive is straightforward: In a soft market keeping the addressable market as large as possible is paramount.

Clients adopting AI need insurance that remains useful as their operations change. Broad exclusions risk removing the protection those clients increasingly need, and the business brokers can place. An analysis highlighted by Apollo chief economist Torsten Sløk in September found that 69% of S&P 500 companies pointed to a live AI deployment in their second-quarter earnings disclosures, up from 64% in the previous quarter. That is a substantial part of corporate America for insurance to accommodate.

Underwriters, meanwhile, have to turn that demand for protection into a price, a limit and a commitment to pay when it goes wrong. The scale of AI spending gives some context to that task. 7 trillion of worldwide AI spending in 2026, more than 12 times the London Market Group’s projected $219 billion of gross written premiums if current growth continues. Spending and premiums are different measures, but it does indicate the scale of activity insurers are being asked to underwrite.

AI developers have access to vast amounts of capital, direct access to their models, specialist researchers and the strongest incentives to understand their behaviour. Yet in an April 2025 essay, Anthropic co-founder Dario Amodei wrote: “People outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work. " Insurance has always dealt with uncertainty, but Amodei’s admission points to an additional difficulty with frontier AI: uncertainty can extend to the mechanisms themselves. Insurers are essentially being asked to underwrite an airline that uses the Wright brothers' prototype aircraft.

Aviation was a transformational industry, but there were a few crashes along the way. Products Insurers are already identifying specific ways AI can cause loss. The difficulty is deciding how far experience of today’s systems can guide cover for technology that may behave differently by the next renewal. That uncertainty also affects price.

Beazley’s Alessandro Lezzi told The Insurer in September that the wider cyber market was not yet reflecting AI risks in its pricing. His aim was to charge for emerging exposures before a major claims event forced an abrupt correction. “Anticipate the change, bring the price today,” he said. After affirming AI-related coverage in its cyber and technology E&O policies, Beazley added two endorsements covering business interruption from voluntary shutdowns of malfunctioning AI systems and certain AI-specific regulatory proceedings arising from unintentional non-compliance.

Pricing reflects clients’ AI use and governance. Lezzi distinguished those exposures from failure of AI sold as a product, which its current cover does not insure. Coalition launched enhanced AI coverage addressing output-related media and technology E&O risks, including LLM-jacking through service-fraud cover. The endorsement was live in the UK and North America at the September announcement, with further markets planned.

Marsh Re is developing named-peril AI catastrophe protection for defined incidents capable of producing losses across cyber, standalone AI and potentially other portfolios. Capacity, participating reinsurers and detailed terms were not disclosed, and a completed placement was not confirmed. The accumulation concern remains. Everest’s Emily Davis said the industry was still at an early stage in understanding silent AI exposure, including its effect on claims frequency and severity and the potential for losses where organisations use the same model.

Five senior cyber reinsurance market participants said AI was unlikely to materially change purchasing at January 2027 renewals. Gallagher Re’s Johnny Fraser nevertheless warned of potential short-term attritional deterioration and clusters of losses that could support demand for short-window aggregate or surge protection. Even brokers opposing broad exclusions recognise limits. WTW’s Hill warned that AI creates systemic exposure and that clients should not expect every AI risk to be insurable.

Governance Product development leaves a basic question: what, exactly, are insurers pricing? A familiar machine-learning application and an autonomous system taking decisions can present very different exposures. A policy referring broadly to AI still needs to establish which activities and losses it addresses. The Lloyd’s Market Association confirmed in September that it was consulting members on a model AI definition.

Consistent terminology would help insurers describe their exposure and identify mismatches between primary and reinsurance contracts. It would still leave each carrier to determine its appetite and terms. The UK Financial Conduct Authority is gathering evidence on how insurers use AI and how insurance for AI-related exposures could support innovation. Its review covers products, customer service, underwriting and decision-making, alongside the role of risk transfer.

The liability consequences also extend to corporate disclosures. RT Specialty’s Kevin LaCroix reported 22 AI infrastructure lawsuit filings this year as of early September, compared with 16 during all of 2025, highlighting allegations that companies overstated their ability to benefit from the infrastructure boom. Beazley and Gallagher specialists identified a potential D&O exposure where energy-intensive AI strategies conflict with environmental commitments already made to shareholders. Companies need to explain how their AI investment fits with those commitments.

Governance also matters within the insurance industry. Euclid’s John Dexter urged MGAs to document which uses of AI they consider acceptable before implementing the technology. At the Dive In Festival, an opening-session audience poll found that 71% of attendees did not trust leaders to be honest about AI’s impact, underlining the importance of how firms explain changes to their staff. Usage Research commissioned by mea Platform found that 83% of respondents would hand repeatable operational work to AI to capture business currently turned away.

Carriers estimated that operational constraints leave one in nine broker submissions declined or unquoted despite appetite to write them. That enthusiasm came with conditions: 75% said they would trust only an insurance-specific model or a governed hybrid for high-consequence decisions with limited oversight. The opportunity is to handle more business within insurers’ existing rules and controls. Dexter told the European MGA Summit that AI could improve automation and the movement of data, while experienced underwriters remained essential.

His distinction was between helping underwriters process information and replacing their judgment. Distribution is another area of potential change. BofA Securities said Meta’s Muse revived the debate over insurance distribution as AI agents could solicit and compare competing offers. Insurify blocked the agent, warning that automated scraping could strip quotes of information consumers need to make informed decisions.

5 million in a SignalFire-led Series B, following its $17 million Series A in May. The insurance workflow platform also introduced a tool to automate certificate-of-insurance processing. From Ciab And Early October The CIAB Insurance Leadership Forum brought further examples of how AI is changing the risk environment, from claims fraud to the scope of affirmative cover. Crawford & Company’s chief commercial and strategy officer Pat Van Bakel told The Insurer that generative AI was making fraud easier by allowing sophisticated and less experienced actors to create convincing fake photographs and documents.

He warned that fraudulent material could pass unnoticed unless claims organisations develop ways to identify manipulation at the start of the process. Because insurance underpins so much of the economy, he argued, the industry also has a wider stake in improving deepfake detection and responding to the societal risks of AI-generated content. Coalition’s Shawn Ram said the company had affirmatively covered AI risks within its cyber and technology E&O policies since 2024. He described an approach spanning SMEs and enterprise clients across geographies, with enhancements addressing smaller firms’ chatbot use, larger companies’ regulatory exposures and AI tools for brokers.

Ram said brokers had welcomed that clarity as clients navigate data leakage, unwanted automated decisions and evolving regulation. He argued that AI exposures relating to cyber and technology E&O should sit within those existing policies, while distinguishing them from property and general liability risks that cyber insurance is not designed to cover. In a separate interview published on October 8, Coalition’s Michael Phillips warned that cheaply generated claims correspondence could raise handling costs and delay valid claims, describing the pressure as “synthetic inflation”.

Will Gow, now at Brecon Specialty, has described similar friction: AI-assisted demands can trigger legal referrals and complaints processes even where a claim’s merits have not changed. Phillips also argued that shared algorithmic decisions could give groups of claimants a common basis for challenging an outcome, making collective litigation easier. AI’s effect on claims may therefore emerge through the volume and cost of disputes as well as through new causes of loss. Catch up on the previous issue:Insurance for inadvertent hacks.