From analyzing ISO endorsement filings across 23 states, the pace at which carriers adopt CG 40 47 versus CG 40 48 reveals which coverage parts they expect to generate AI-related claims first. The answer is unambiguous: carriers are filing the broad exclusion (CG 40 47, Coverages A and B) at roughly three times the rate of the narrow form (CG 40 48, Coverage B only), indicating that bodily injury and property damage from AI systems, not just advertising injury and IP infringement, are already in their loss development assumptions. State regulators have approved more than 80% of these filings. And as CGL policies shed AI exposure at scale, a parallel specialty market is forming where cyber insurance, professional indemnity, and AI liability converge into what Gallagher Re now calls a single “digital risks” product line. This article maps the structural unbundling of AI risk from baseline commercial liability, the investment capital pouring into the convergence zone, and the underwriting discipline required to price exposures with almost no loss history.
The structural unbundling: how AI risk left the CGL policy
The unbundling began on January 1, 2026, when Verisk’s ISO Core Lines Services released three optional generative AI exclusion endorsements. CG 40 47 01 26 is the broadest: it excludes both Coverage A (bodily injury and property damage) and Coverage B (personal and advertising injury) for any claim “arising out of” generative AI. CG 40 48 01 26 is narrower, excluding only Coverage B. CG 35 08 01 26 targets Products/Completed Operations, removing Section I coverage for downstream harm from AI embedded in finished products. All three define generative AI as “a machine-based learning system or model trained on data with ability to create content or responses, including text, images, audio, video or code.”
That definition is broad enough to capture large language models, image generators, code assistants, and any system producing novel output from learned patterns. It does not explicitly reach traditional predictive models that classify or score without generating content, though the boundary between “generative” and “predictive” AI blurs with every model update. The critical phrase, “arising out of,” requires only a causal connection rather than proximate causation. An insured whose vendor-embedded AI tool contributes to a loss may find the exclusion triggered even if the AI component was peripheral to the underlying claim.
ISO forms underpin approximately 82% of U.S. property and casualty policies. When Verisk introduces a new endorsement and major carriers begin filing, adoption follows a predictable S-curve. Within weeks of the January release, W.R. Berkley, Cincinnati Financial, Frederick Mutual, and Philadelphia Insurance filed their own AI exclusion wording with state regulators. Berkshire Hathaway, Chubb, and Travelers requested AI-related exclusions across their GL books. Testudo co-founder George Lewin-Smith predicted that “95% of carriers will immediately employ these exclusions.” Based on filing data through April 2026, that prediction is tracking close to reality in the large-account segment.
| Endorsement | Scope | What It Excludes | What It Preserves |
|---|---|---|---|
| CG 40 47 | CGL Coverage Part | Coverage A + B (BI, PD, personal/advertising injury) | Nothing AI-related |
| CG 40 48 | CGL Coverage Part | Coverage B only (personal/advertising injury) | Coverage A (BI, PD) |
| CG 35 08 | Products/Completed Ops | Section I (BI, PD from AI in products) | Premises/operations exposure |
AI exclusions spread beyond CGL into D&O and E&O
The CGL exclusion cycle was only the opening move. AI exclusion language is now migrating into directors and officers (D&O), errors and omissions (E&O), employment practices liability (EPLI), and fiduciary liability policy forms, creating layered coverage gaps for any enterprise deploying AI across its operations.
AIG, W.R. Berkley, and Great American Insurance Group have sought regulatory clearance for AI-specific exclusions in management liability policies. Berkley’s “Artificial Intelligence Exclusion (Absolute)” for D&O eliminates coverage for any claim “based upon, arising out of, or attributable to” the actual or alleged use, deployment, or development of AI. That language reaches far beyond Verisk’s generative AI definition: Berkley defines AI as “any machine-based system that infers how to generate outputs such as predictions, content, recommendations, or decisions, including systems that emulate input data to generate synthetic content.” This captures not just generative models but traditional ML systems used in pricing, underwriting, and claims triage.
Hamilton Insurance Group has filed a Generative AI Exclusion for professional liability that removes coverage for any claim “based upon, arising out of, or in any way involving any actual or alleged use of generative artificial intelligence,” specifically naming ChatGPT, Bard, Midjourney, and DALL-E. Design professional E&O policies are being targeted with similar language. The absolute exclusions in management and professional liability cover content generation using AI, failure to detect third-party AI-generated content, inadequate AI governance and controls, regulatory investigations involving AI, and AI capability disclosures to investors.
The practical consequence: an enterprise that deploys AI in customer service, marketing, product design, and internal operations could find itself excluded from CGL coverage under CG 40 47, excluded from D&O coverage under Berkley’s absolute exclusion, and excluded from E&O coverage under Hamilton’s generative AI carve-out. Even incidental use of AI tools embedded in enterprise software from Salesforce, Microsoft, or Google may be sufficient to trigger the exclusion. The coverage vacuum is structural, not incidental.
The three-layer liability vacuum
The convergence thesis rests on understanding where liability lands when traditional policies withdraw. Three layers of the AI liability chain have simultaneously shed responsibility.
Layer 1: AI developers disclaim liability. Standard terms of service from OpenAI, Anthropic, Google, and other foundation model providers cap contractual liability at 12 months of license fees and offer no performance warranties. Enterprise exposure from a hallucinating model has no corresponding ceiling. The enterprise holds liability for a model it did not build, trained on data it did not select, using architecture it cannot audit.
Layer 2: Traditional insurers exclude AI. With CG 40 47 in CGL, absolute exclusions in D&O and E&O, and CG 35 08 in Products/Completed Operations, the traditional insurance stack systematically removes AI-related claims from coverage. The exclusion takes effect at the next policy renewal with no transition period.
Layer 3: Standalone AI liability capacity is nascent. The combined capacity of dedicated AI liability providers, including Testudo ($9.25 million per insured), Armilla ($25 million per organization), Munich Re aiSure, Corgi, and Mayflower Specialty, totals a few billion dollars at most. This compares against a CGL market measured in the hundreds of billions. The protection gap between exclusion adoption and standalone product maturation is measured in orders of magnitude.
The result is uninsured AI liability sitting on corporate balance sheets. For enterprises deploying AI at scale, the risk management calculus has changed fundamentally in the first five months of 2026.
Gallagher Re quantifies the convergence
Gallagher Re’s Q1 2026 Global InsurTech Report provides the clearest quantitative signal that the market recognizes this convergence as structural. Total insurtech funding in Q1 2026 reached $1.63 billion, the strongest consecutive quarter pair since Q3 2022 (following $1.67 billion in Q4 2025). AI-focused firms captured 95.2% of the total, or $1.55 billion across 68 deals, at an average deal size of $25.79 million.
Within that total, firms operating at the intersection of AI liability and cyber insurance raised $444.84 million in Q1 2026 alone. Since 2012, digital and cyber risk insurtechs have attracted $5.77 billion across 263 deals. The Q1 2026 figure represents an acceleration: nearly 8% of all cumulative digital risk insurtech funding arrived in a single quarter.
Freddie Scarratt, Gallagher Re’s global deputy head of insurtech, framed the trend directly: “The emerging landscape of third-party AI liability insurance is poised to mirror explosive growth in cyber reinsurance markets.” Andrew Johnston, global head of insurtech, noted that “95.2% of Q1 2026 insurtech funding went to AI-labelled companies, underscoring the industry’s commitment to this transformative technology.”
The convergence thesis, in Gallagher Re’s framing, holds that cyber insurance, professional indemnity, and AI liability are coalescing into one underwriting discipline: digital risks. This is not a prediction. It describes a structural shift already visible in product design, capital allocation, and regulatory filings.
| Metric | Q1 2026 Value | Source |
|---|---|---|
| Total insurtech funding | $1.63 billion | Gallagher Re |
| AI-focused share | 95.2% ($1.55B across 68 deals) | Gallagher Re |
| AI liability + cyber insurtech funding | $444.84 million | Gallagher Re |
| Average AI deal size | $25.79 million | Gallagher Re |
| Early-stage AI insurtech funding | $548.5M (up 36.1% QoQ) | Gallagher Re |
| Cumulative digital/cyber risk funding since 2012 | $5.77B across 263 deals | Gallagher Re |
The standalone AI liability market takes shape
Multiple carriers and MGAs have launched or expanded AI-specific coverage products in 2026, filling the gap that CGL exclusions are creating. The product architectures differ, but they share a common challenge: pricing coverage for exposures with almost no claims history.
Testudo launched in January 2026 as a claims-made product for mid-to-large U.S. enterprises deploying generative AI. Backed by Lloyd’s syndicate Apollo and a participant in Lloyd’s Lab Cohort 14, Testudo initially offered $8.5 million per insured. By March 2026, the MGA expanded capacity to $9.25 million per insured after adding Atrium and QBE to its reinsurance panel. Coverage includes AI hallucinations, intellectual property infringement, physical property damage from AI, unauthorized data disclosures, and third-party claims arising from AI-generated outputs. Co-founder George Lewin-Smith, a former Goldman Sachs VP, described the underwriting approach: “Our technology and data allows us to understand, price and underwrite real-world artificial intelligence risks.” Testudo’s platform uses proprietary litigation data and real-world risk signals rather than historical loss triangles, reflecting the absence of traditional actuarial pricing inputs.
Armilla AI holds the distinction of being the first Lloyd’s coverholder dedicated exclusively to AI liability, backed by Chaucer and Axis Capital. Armilla offers up to $25 million per organization and requires continuous model quality assessments as a condition of coverage. The governance-based underwriting approach, evaluating model inventories, monitoring capabilities, audit trails, and incident response procedures, treats risk management maturity as a proxy for loss frequency when actuarial data is unavailable.
Munich Re pioneered the space in 2018 with aiSure, combining expert-driven due diligence with parametric-style triggers for AI performance errors. In March 2026, Munich Re’s subsidiary HSB introduced dedicated AI liability insurance for small and midsize businesses, covering bodily injury, property damage, and personal and advertising injury for claims stemming from AI-generated content. HSB designed the product to fill exactly the gap that CG 40 47 creates.
Corgi closed a $160 million Series B in early 2026 at a $1.3 billion valuation, launching dedicated AI liability coverage for hallucinations, bias, and training data disputes. AXA XL added a generative AI endorsement to its global cyber insurance program, covering data poisoning, usage rights infringement, and regulatory violations tied to generative AI. QBE introduced an endorsement covering limited fines under the EU AI Act, becoming the first major insurer to explicitly reference the regulation in policy wording.
| Provider | Max Capacity | Backing | Product Type |
|---|---|---|---|
| Armilla AI | $25M per org | Lloyd’s / Chaucer / Axis | Standalone AI liability |
| Testudo | $9.25M per insured | Lloyd’s / Apollo / Atrium / QBE | Claims-made AI liability |
| Munich Re aiSure | Varies | Munich Re balance sheet | Parametric AI performance |
| HSB (Munich Re) | SME limits | Munich Re balance sheet | AI liability for SMEs |
| Corgi | Up to $50M | $160M Series B ($1.3B val) | AI hallucination/bias/IP |
| AXA XL | Global program | AXA balance sheet | Cyber + GenAI endorsement |
The litigation catalyst: 978% growth in GenAI lawsuits
Capital is flowing into AI liability products because the loss events are no longer hypothetical. Generative AI-related lawsuits in the United States grew 978% between 2021 and 2025, with over 700 cumulative filings. Year-over-year increases accelerated from 59% (2023 to 2024) to 137% (2024 to 2025). Filings in just the first four months of 2025 surged 81% compared to the same period the prior year. Stanford’s 2025 AI Index documented a 56% year-over-year increase in AI-related incidents more broadly.
The composition of these lawsuits maps directly onto the coverage parts that CGL exclusions target. Patent infringement accounts for 11.9% of cases, copyright infringement for 11.2%, and personal injury claims including privacy and data misuse for 10.2%. These are Coverage B claims (personal and advertising injury) under the CGL form, precisely the exposure that both CG 40 47 and CG 40 48 exclude.
Average AI-related case settlements run approximately $4 million, according to Testudo’s proprietary litigation database. That figure represents the early, pre-precedent phase of the litigation cycle. As courts establish doctrines around AI-generated content liability, autonomous system failures, and algorithmic discrimination, severity is expected to increase. Munich Re surveys find that 57% of executives identify AI errors, misinformation, and hallucinations as their top risk, followed by legal and reputational risks (56%) and data protection violations (55%).
The litigation trajectory matters for pricing because it establishes the frequency signal that actuaries need. Even without mature loss triangles, the 978% growth rate in filings and the $4 million average settlement provide anchor points for scenario-based severity and frequency assumptions. These are exactly the parameters that early cyber insurance actuaries worked with in 2014 and 2015, when standalone cyber pricing relied on similar proxies before credible loss experience developed.
Pricing without actuarial triangles
The actuarial challenge at the center of the convergence thesis is severe. Applying Berliner’s classic insurability criteria, generative AI risks fail substantially across multiple dimensions, particularly predictability and information availability. The Actuary Magazine, in a 2026 analysis, concluded that historical loss data to model AI-related events remains “scarce,” severity is “unknown (expected to be significant),” frequency is unpredictable, and causation is ambiguous. The creative nature of generative AI outputs means losses can arise from unique and unforeseen interactions between model behavior, user prompts, and downstream applications.
Carriers writing standalone AI liability are deploying several alternative pricing methodologies:
- Parametric triggers. Munich Re’s aiSure uses parametric-style triggers for clearer claims settlement rather than traditional indemnity structures. When an AI system’s performance falls below a contractually defined threshold, the parametric trigger activates, avoiding the causal ambiguity that plagues traditional liability claims.
- Scenario modeling. Without loss triangles, underwriters construct frequency and severity distributions from scenario analysis. AI agents can simulate 25,000 stress test scenarios overnight using Monte Carlo methods across inflation shifts, model failure modes, and cascading deployment errors. The resulting distributions are judgment-informed but at least quantified.
- Real-time litigation data. Testudo’s proprietary platform indexes active AI litigation, tracking filing rates, settlement values, and claim types to build a dynamic view of emerging loss patterns. This is not a substitute for actuarial experience, but it provides real-time calibration for scenario assumptions.
- Governance-based underwriting. Armilla and several other providers evaluate an applicant’s AI governance maturity, including model inventories, continuous monitoring, bias audits, and incident response protocols, as a risk selection mechanism. Enterprises with documented governance programs receive coverage that ungoverned enterprises cannot obtain at any price.
- Synthetic data augmentation. Life insurers are using generative AI to create synthetic datasets that augment sparse historical experience. The same technique applies to AI liability, where synthetic claim scenarios generated from known AI failure modes can supplement the thin tail of actual losses.
Aon’s Kevin Kalinich framed the systemic risk dimension: the industry could absorb “a $400-500 million loss from one company’s misfiring AI agent. What it cannot absorb is an upstream failure that produces a thousand losses simultaneously.” Traditional insurance relies on uncorrelated, diversifiable risk pools. If AI-induced dislocation becomes systemic, because a single foundation model update affects thousands of commercial deployments simultaneously, insurers face unquantifiable aggregate exposures. This correlation structure looks more like catastrophe risk than like traditional liability risk, and it challenges the fundamental diversification assumptions that actuaries use to set capital requirements.
The regulatory overlay accelerates the cycle
Regulatory catalysts are compressing the timeline between AI exclusion adoption and standalone product demand. The EU AI Act classifies AI systems for risk assessment and pricing in life and health insurance as high-risk, with enforcement beginning August 2, 2026. Penalties reach up to 7% of global annual turnover for prohibited AI applications, 3% for other violations. QBE’s endorsement covering limited EU AI Act fines represents the first explicit regulatory-risk AI insurance product.
In the United States, 23 states and Washington D.C. have adopted the NAIC’s 2023 AI Model Bulletin. The NAIC’s 12-state AI Evaluation Tool pilot, running from January through September 2026, covers Colorado, Maryland, Louisiana, Virginia, Connecticut, Pennsylvania, Wisconsin, Florida, Rhode Island, Iowa, Vermont, and California. Colorado’s AI Act imposes a compliance deadline of June 30, 2026, adding a state-level overlay that creates regulatory tension for carriers that exclude AI coverage while simultaneously being required to avoid algorithmic discrimination.
State legislatures are also expanding private rights of action for AI-related harms. New York’s S.B. 6278 creates a private right of action for intimate deepfakes with punitive damages. Michigan’s S.B. 760 allows minors to sue when chatbots encourage self-harm. New York’s A.B. A9396 bans personalized algorithmic pricing with class action provisions and $5,000 statutory damages plus treble damages. Each of these creates a new claim vector that would historically have been absorbed by CGL or E&O policies but now falls into the coverage vacuum.
The regulatory cycle mirrors the cyber insurance precedent precisely. The 2016 U.S. Treasury TRIA determination for standalone cyber policies gave that market regulatory legitimacy. The EU AI Act and NAIC model bulletin are playing an equivalent catalytic role for AI liability, providing a framework that makes AI risk more tractable to underwrite while simultaneously generating the compliance-driven demand that supports premium volume.
The convergence product: what “digital risks” looks like
Gallagher Re’s framing of cyber, professional indemnity, and AI liability coalescing into one underwriting discipline reflects product design choices already visible in the market. AXA XL’s generative AI endorsement sits on its cyber insurance chassis. QBE’s EU AI Act coverage attaches to existing professional lines. Testudo’s claims-made form borrows structural features from both cyber and tech E&O policy architecture.
A fully converged “digital risks” product would cover, under a single policy form: unauthorized data access and breach response (traditional cyber), AI system failures causing third-party harm (AI liability), professional errors arising from AI-assisted advice or output (professional indemnity), regulatory fines and investigation costs for AI non-compliance (regulatory liability), and intellectual property infringement by AI-generated content (media/IP liability). Several London market syndicates are reportedly workshopping manuscript forms along these lines for 2027 renewals.
For the pricing actuary, the convergence creates both challenge and opportunity. The challenge is correlations: cyber losses, AI liability losses, and professional indemnity losses from the same enterprise are not independent. A data breach that exposes training data, causes an AI model to produce biased outputs, and triggers a regulatory investigation generates correlated claims across all three traditional coverage lines. Pricing a converged product requires modeling these joint distributions, not just summing individual peril loads.
The opportunity is that a converged product captures the full digital risk profile of an enterprise in one underwriting assessment. Rather than three separate applications, three separate risk engineering visits, and three separate policy forms, the underwriter evaluates the enterprise’s digital risk posture holistically. This reduces acquisition costs, eliminates coverage overlap disputes, and creates a single point of accumulation management for the carrier and its reinsurers.
Demand signals and the protection gap
The demand for AI liability coverage is not speculative. Munich Re’s 2026 survey found that 63% of executives want to purchase insurance against AI-related risks. The Geneva Association reported that over 90% of businesses expressed a need for specific AI insurance coverage, with two-thirds willing to pay at least 10% more in premiums for it. By early 2025, 71% of firms were using generative AI tools, up from 33% two years prior.
Against that demand, the global AI insurance market was valued at $8.63 billion in 2025 and is projected to reach $59.5 billion by 2033, a compound annual growth rate exceeding 27%. Global cyber insurance premiums, the closest comparable line, are projected to rise from $16 to $20 billion in 2025 to $30 to $50 billion by 2030, according to Munich Re. The cyber protection gap stands at approximately 90%: roughly nine out of ten cyber risks remain uninsured.
The AI protection gap is likely even wider in the near term. Cyber insurance has had over a decade to develop products, build distribution, and accumulate loss data. Standalone AI liability is in its first year. The carriers and MGAs writing AI liability today are, in effect, building the loss database that will eventually enable actuarial pricing to move from judgment to experience. The parallel to standalone cyber in 2014 and 2015 is exact: early entrants write first-mover premium at margins that compensate for uncertainty, and as loss data matures and more carriers enter, combined ratios compress toward equilibrium.
Why this matters for actuaries
The convergence of cyber, professional indemnity, and AI liability into a single digital risks discipline touches every actuarial function.
Pricing actuaries working on GL books must now model the loss-load differential between AI-excluded and AI-silent policies. Patterns we have seen in early-stage loss development on technology E&O and media liability portfolios suggest a near-term AI loss load of 0.5 to 2.0 percentage points of the expected loss ratio for standard commercial GL, varying by insured industry and AI deployment intensity. That range will widen as adoption accelerates and courts establish severity benchmarks. For actuaries pricing standalone AI liability or converged digital risk products, the challenge is constructing frequency and severity distributions from scenario analysis, litigation trend data, and governance-based risk segmentation rather than actuarial triangles.
Reserving actuaries face a bifurcation problem. Historical GL loss development triangles do not isolate AI-related claims. As these claims begin to emerge in the data, they will appear as an acceleration in development factors and an upward shift in frequency and severity trends. Separate development patterns for AI-excluded versus non-excluded segments may be warranted. For standalone AI liability books, initial reserves will rely almost entirely on expected loss ratio methods; chain-ladder techniques require development history that does not yet exist.
Enterprise risk management actuaries must evaluate how correlated AI exposure concentrates across lines. A GL exclusion does not eliminate the risk; it migrates the aggregation to tech E&O, D&O, cyber, or the corporate balance sheet. Whether an enterprise’s total AI exposure is covered, excluded, or uninsured depends on the endorsement status of every policy in the tower. The correlation structure of AI failures, where a single foundation model error propagates across thousands of commercial deployments, resembles catastrophe risk more than traditional liability risk.
From an ASOP compliance perspective, ASOP No. 12 (Risk Classification) governs how AI adoption intensity is reflected in classification relativities. ASOP No. 29 (Expense Provisions) applies to the acquisition cost assumptions for a converged digital risks product. ASOP No. 56 (Modeling) covers any scenario-based AI loss modeling used to set reserves or price standalone products. Actuaries filing rate indications that reflect AI exclusion adjustments should document the basis for the adjustment, the data or judgment used, and the sensitivity of the indication to changes in AI adoption assumptions.
This continues a trend we have been tracking across multiple dimensions: the insurance industry absorbs new technology risk by first excluding it from existing products, then studying the excluded exposure, then pricing it as a standalone line. Cyber followed this path from 2014 to 2019. AI liability entered the exclusion phase in January 2026. The investment data, the litigation growth, the regulatory catalysts, and the MGA launches all point to the same conclusion: the standalone phase is not coming; it has arrived. The actuaries who build the pricing frameworks for converged digital risks products over the next 12 to 24 months will define how the industry absorbs AI liability for the next decade.