Five of the six largest US P&C carriers disclosed production AI deployments on Q1 2026 earnings calls while winning state approval to exclude AI-related damages from the commercial policies they sell. Chubb's CEO named Anthropic's Claude Mythos as triggering a new era of cyber-adjacent risk in the same quarter his company filed AI exclusion endorsements in more than a dozen states.

The number that makes it a strategy rather than a contradiction is the approval rate: regulators cleared more than 80% of the filings.

Key Takeaways

  • 88% accuracy on automated decisions and 4 million data points per commercial risk is AIG Assist's disclosed throughput, at a carrier simultaneously filing AI limitations on the policies it writes.
  • More than 80% of AI exclusion applications have been approved, fastest in Florida, Connecticut and Maryland, with ISO forms sitting under roughly 82% of US commercial general liability policies.
  • 978% growth in US generative AI lawsuits between 2020 and 2025 on more than 700 cumulative filings, with year-over-year growth accelerating to about 137% in 2024-2025 from 59% the year before.
  • One AI failure, five lines. A single system failure can trigger CGL, D&O, E&O, cyber and employment practices claims at once, which is what the multi-line exclusions respond to.
  • $2 million to $50 million is the standalone AI liability limit range now filling the gap, priced without loss triangles.

Both Sides of the Same Ledger

The two activities are running in parallel at the same companies, in the same quarters.

On the deployment side, AIG Assist integrates Palantir Foundry with Claude to process commercial submissions, reporting 88% accuracy on automated decisions and 4 million data points per commercial risk, with human underwriters reviewing edge cases rather than routine binds. Travelers put Claude in front of 10,000 employees against a $1.5 billion technology budget. Allstate built ALLIE in-house, running customer engagement, direct policy sales in three states and claims processing. Chubb created a global claims role with an explicit AI mandate across 54 countries.

On the exclusion side, Verisk's ISO forms went live January 1, 2026. CG 40 47 01 26 removes both Coverage A and Coverage B for loss arising out of generative AI; CG 40 48 01 26 removes Coverage B alone.

Carrier Form / Approach Lines Affected Scope
ISO / Verisk CG 40 47 01 26 CGL (Coverage A + B) Total generative AI exclusion
ISO / Verisk CG 40 48 01 26 CGL (Coverage B only) Personal & advertising injury only
W.R. Berkley PC 51380 D&O, E&O, Fiduciary Absolute AI exclusion
Hamilton HAM-AI-2025 E&O, Cyber Sublimit (25-50% of policy limit)
Berkshire / Chubb / Travelers ISO-based filings CGL State-by-state AI exclusion filings

W. R. Berkley's Form PC 51380 goes furthest, an absolute exclusion reaching D&O, E&O and fiduciary liability for any claim based upon, arising out of, or attributable to the use, deployment or development of AI, covering content generation, failure to identify AI-generated content, inadequate AI policies, chatbot representations and regulatory violations. Coverage requires affirmative warranties that outputs are reviewed by qualified personnel, that verification is documented, and that all AI systems are inventoried and individually risk-assessed.

What the Multi-Line Form Is Actually Pricing

The exclusions that reach across product lines are answering a correlation problem, and that is the part with an actuarial argument behind it.

A single AI system failure does not stay in one coverage part. The same deployment can produce a CGL bodily injury claim, a Coverage B advertising injury claim, a D&O suit over what the company said about its AI capability, an E&O claim from professional reliance on AI output, and an employment practices claim from an AI screening tool, from one root cause. Carriers pricing those lines independently are understating the correlation between them, and removing the exposure from several lines at once is a cheaper response than modelling a dependency structure nobody has data for.

The litigation trajectory is what made that urgent rather than theoretical. Generative AI lawsuits in the US grew 978% between 2020 and 2025 on more than 700 cumulative filings, and the year-over-year rate accelerated to roughly 137% in 2024-2025 from 59% the year before. Skadden reports AI-related securities class actions outpacing other categories, targeting overstated efficiency gains and legacy technology rebranded as AI.

Development is the second half of the problem. Loss development factors need a stable pattern, and the lag between AI deployment and claim emergence ranges from months for content infringement to years for product liability arising from AI-generated advice. Two claim types with different tails sitting in the same accident year, with no history for either, is not a triangle an IBNR estimate can be built on. The exclusion removes the need to try.

Whose Policy Responds When the Carrier's Own Model Is Wrong

The asymmetry creates a question the forms do not answer, and it points back at the filers.

A carrier running an AI system that produces a flawed underwriting decision harming a policyholder has an exposure of its own. Under its E&O cover and under a CGL written on the same market forms it is now filing, AI-related claims are excluded. The carrier deploying most aggressively is therefore also the one whose own AI liability is least clearly covered, and no coverage litigation has tested it yet.

Policyholders face the retail version of the same gap. Renewal packages in Q1 and Q2 2026 attach CG 40 47 or CG 40 48 in the endorsement schedule, which is where endorsements go, and a risk manager who has not read the schedule against last year's has not seen the change. Aon, Gallagher and Lockton have each flagged that companies deploying AI agents may find previously assumed coverage is gone. The categories most exposed are ordinary: marketing teams using AI-generated copy, HR departments running AI screening, product teams embedding generative AI in customer-facing features.

The standalone market absorbing that demand is pricing on the same absent data. Corgi reached a $1.3 billion valuation, Armilla launched in 2025 requiring ongoing model quality assessments, and Testudo opened a claims-made product in January 2026, with limits across the market running $2 million to $50 million. Deloitte projects roughly $4.7 billion of annual global AI insurance premium by 2032. With fewer than 700 cumulative lawsuits through 2025 and no mature triangles, those rate plans rest on scenario modelling and analogy to the cyber development arc, which means the first large AI liability verdict reprices the segment in a single event rather than through experience.

Further Reading

Sources