Traditional insurance carriers have a problem they'd rather not talk about. The algorithms powering everything from hiring tools to medical diagnostics don't fit neatly into the coverage categories they've spent decades refining—and when something goes wrong, the losses can spiral in ways that make underwriters deeply uncomfortable.
So they've been doing what insurers do when uncertainty looms: tightening exclusions, adding carve-outs, and in some cases simply declining to cover AI-related incidents at all. That retreat has created an opening. Now a cluster of startups, backed by Lloyd's syndicates and specialty underwriters, is racing to build an entirely new insurance category from scratch.
Corgi recently unveiled what it calls AI Insurance Coverage, a set of modular add-ons to standard Tech Errors & Omissions policies. The product targets the specific failure modes that keep AI deployers awake: biased algorithms that produce discriminatory outcomes, language models that hallucinate defamatory content, autonomous systems that malfunction, adversarial attacks that corrupt models. It's not groundbreaking in concept—Counterpart rolled out expanded affirmative AI coverage late last year, adding dedicated insuring agreements for AI exposures. Armilla AI launched standalone liability coverage with Lloyd's underwriter Chaucer in April 2025, eventually scaling its capacity to $25 million in early 2026. New Dawn Risk introduced "AI Protect" for both developers and deployers. Cowbell bundled AI risk into its "Prime One" cyber product alongside quantum computing threats in April 2026.
The flurry of launches reflects something more than startup opportunism. It signals a recognition that AI risk doesn't map cleanly onto existing insurance lines—not cyber, not general liability, not quite professional indemnity either. These products are bets on how the market will ultimately categorize and price a whole new class of exposure.
What traditional carriers won't touch
The gap exists because traditional insurers are spooked by correlation risk. A single flawed training dataset could compromise thousands of models simultaneously. A widely deployed algorithm that turns out to discriminate could trigger a wave of lawsuits across industries. Unlike a data breach or a software bug—events insurers know how to model—AI failures can cascade in unpredictable ways.
Aon's recent market reports have flagged what the firm calls "silent AI" exposure: risks embedded in existing policies that weren't designed with machine learning in mind. Some carriers have responded by capping exposure. QBE, for instance, introduced endorsements limiting coverage for EU AI Act fines to 2.5% of policy limits, a detail buried in regulatory summaries from late last year. Chubb has clarified its stance on certain AI incidents while explicitly excluding systemic or widespread events—the kind of correlated losses that could bankrupt a carrier.
That tightening has accelerated recently, perhaps more than the market initially anticipated. Gallagher Re's analysis of the AI insurance landscape, published in May 2026, argues that a significant category of uninsured risk is emerging across professional indemnity, product liability, and casualty lines. The firm maps coverage needs evolving in real time as AI embeds itself deeper into business operations.
Lloyd's appetite for novel risk
Armilla's success in securing Lloyd's backing is worth noting. While standard commercial carriers have retreated, Lloyd's syndicates—with their historical appetite for novel exposures and flexible underwriting structures—have proved more willing to write affirmative AI coverage. Chaucer's partnership in co-developing Armilla's product gave the startup both credibility and distribution, a combination that's hard for an unknown insurer to replicate.
The Lloyd's model suits AI risk in ways traditional insurers can't easily match. Syndicates can write coverage on a modular basis, adjusting terms as new failure modes emerge. They're comfortable with uncertainty in a way that standard commercial carriers, answerable to shareholders and state regulators, often aren't. Armilla's capacity increase earlier this year came as other insurers were explicitly pulling back.
The partnership approach extends beyond Armilla. One80 Intermediaries launched AI warranty coverage with Armilla AI in late 2024, testing the market with a narrower product before scaling to full third-party liability forms. It's a testing-the-waters strategy that's become common in emerging risk categories: start with warranties, move to endorsements, eventually offer standalone policies once the underwriting data becomes clearer.
Endorsements versus standalone products
How these products are structured matters more than it might seem. Corgi's modular endorsements to Tech E&O policies suggest the company believes AI risk will ultimately integrate into existing lines, just with specific terms. Armilla's standalone form, by contrast, treats AI liability as its own category—closer to how cyber insurance evolved from being a niche endorsement in the 1990s to a standalone product line worth billions today.
Counterpart's approach splits the difference. The company frames its offering as expanded affirmative coverage, creating specific insuring agreements for scenarios like biased outputs or hallucinated content while still anchoring the product to traditional Tech E&O structures. The terminology—"affirmative AI coverage"—signals carriers explicitly agreeing to cover these risks rather than leaving ambiguity that could spawn coverage disputes later.
That ambiguity is what drove the market toward these products in the first place. Insurance Business warned brokers early last year that AI risk doesn't equal cyber risk, despite the temptation to treat them as interchangeable. The distinction is playing out in product design now: cyber policies focus on data breaches and network security, while AI liability products address algorithmic failures and model outputs. Fundamentally different exposures, even if they both involve software.
New Dawn Risk's brochure—circulated to brokers in recent months—targets both developers and deployers, a recognition that liability can sit anywhere in the AI supply chain. A developer might face claims if their model is fundamentally flawed; a deployer might be sued even if they used a third-party system responsibly. The products need to accommodate both scenarios, which makes pricing and underwriting considerably more complex than standard Tech E&O.
Governance bundled with coverage
Armilla's partnership with Trustible, announced in October 2025, bundles governance tooling directly with insurance. It's a pragmatic move: underwriters want assurance that controls are in place before they'll write coverage. The partnership essentially makes risk management a condition of insurability, which might become the norm as the market matures.
The logic mirrors how cyber insurance evolved. Early policies were cheap and broadly written; losses mounted; carriers tightened terms and began requiring specific security controls—multi-factor authentication, incident response plans, regular patching. AI insurance appears to be compressing that evolution into a shorter timeframe, skipping straight to the part where coverage comes with mandated safeguards.
An iterative process
What's emerging is less a settled market than an ongoing experiment. Counterpart framed its product launch as an expansion of affirmative coverage, language that implies iteration—carriers learning as claims data trickles in. The products launching now are likely first drafts, destined to be revised as losses materialize and as regulators clarify how liability attaches in AI incidents.
Aon's recent analysis flags AI as an emerging systemic risk in financial lines, with terms tightening across the board. Gallagher Re's latest InsurTech research tracks the theme of carrier-startup partnerships specifically around digital exposures. The big question—whether traditional carriers eventually re-enter this market or whether it remains a specialty line dominated by Lloyd's and startups—remains unsettled.
For companies deploying AI systems, the practical reality is simpler: affirmative coverage now exists, albeit in fragmented form. The harder part is figuring out which version actually matches their exposure and whether the policy will still be available when they need to renew. Given how fast this market is moving, that's not a trivial concern.
