Chubb's Greenberg Says Casualty Books Have Not Priced AI Liability

Chubb chief executive Evan Greenberg told analysts on October 21 that AI liability "is a growing and emerging exposure that is not contemplated in casualty underwriting really to date," and that the industry will manage it through "a combination of coverage clarifications, ultimately exclusions, and separate AI-related liability coverages," according to the transcript of the third-quarter call. Chubb's North America casualty pricing rose 5.7% in the same quarter.

He was answering Matthew Heimermann of Citi, who asked how to weigh new standalone AI rates and forms against incremental changes to existing forms. "As an industry, we have to put our arms around this," Greenberg said. He added that he could "imagine a regimen around model certification standards emerging," along with best practices in the governance of AI and agentic systems "emerging quickly."

The remarks come six months after Verisk filed the CG 40 47 family of endorsements that let carriers exclude generative AI from commercial general liability policies, which the site covered in April, and a week after Moody's reported that two-thirds of the 58 North American insurers it surveyed run multiple AI applications in production.

Key Takeaways

• Greenberg's sequence runs clarification, then exclusion, then standalone cover, with model certification standards as a possible underwriting basis, on a line he says has no AI exposure priced into it today.
• Chubb's North America commercial pricing split: property down 4.2%, casualty up 5.7%, financial lines up 2.1%, with loss costs that "continue to well outpace general inflation."
• The P&C combined ratio was 84.1% against 81.8%, on $705 million of catastrophe losses and $511 million of favorable prior-period development.
• Moody's found AI at 2% to 10% of IT budgets at most insurers, above 10% at nearly one in five P&C carriers, and only 41% of respondents quantifying and tracking benefits per use case.
• Full credibility for a new liability class needs about 1,082 claims, which an exclusion that removes the exposure from CGL books stops accumulating.

The sequence Greenberg described

"It needs to be understood, it needs to be managed, and it needs to be addressed," Greenberg said of the exposure, noting that the industry has handled emerging liabilities before. The three tools he named map onto what the market has already built. Verisk's CG 40 47, CG 40 48 and CG 35 08 give carriers filed language to exclude generative AI from CGL, and six insurers had filed to use them by April, per the site's coverage of the forms. Munich Re's aiSure line sells performance and error cover for specific models, which the site examined in the spring.

The largest buyers have gone a fourth way. OpenAI secured up to $300 million of AI-risk cover through Aon while Anthropic paid a $1.5 billion copyright settlement from its own balance sheet, and both labs are exploring captives as insurers exclude the risk, according to a CSIS report the site covered in September.

Model certification is the new element. A certification regime would give an underwriter something to rate on before a loss history exists, the role that fire codes and building classifications played in property long before anyone had a credible loss triangle for a given construction type. Greenberg offered it as something he could imagine, not something Chubb is building.

Pricing a quarter without it

Chubb's Form 8-K reports consolidated net premiums written of $15,143 million, up 1.9%, with North America commercial down 2.7% to $5,512 million, or up 1.7% excluding large-account property and financial lines. Core operating income was $2,980 million, down 2.2%, or a record $7.66 per share. The release attributes the decline to unusually light catastrophe losses a year earlier.

Sources: Chubb Form 8-K Exhibit 99.1 for the third quarter of 2026; pricing figures from Greenberg's prepared remarks on the October 21, 2026 call.

On the call, Greenberg said "pricing in numerous areas is failing to keep pace with loss costs," that competition is "spreading to more casualty lines," and that "there is a lot of hungry and extremely naive capacity available." Property pricing was down 11% in shared and layered major-account and specialty business and up about half a point in mid and small business.

The 5.7% casualty figure is the carrier's answer to a loss trend it describes as well above general inflation. By Greenberg's own account it contains no AI component, because the underwriting does not contemplate the exposure. Until an exclusion removes it, whatever AI-related claims arrive under a CGL form sit inside the casualty pick as unidentified trend. Once the exclusion is in, they sit nowhere, which is the problem the site set out in July: a standalone line needs about 1,082 claims for full credibility, and generative AI litigation is up 978% since 2021 on a base too small to rate.

Chubb released $511 million of prior-period reserves in the quarter, of which $14 million came from commercial lines, per the site's reading of the release. A reserve position that thin on the commercial side leaves little room to absorb a new peril arriving unpriced.

The same carriers as AI users

Moody's Ratings surveyed 58 North American life, health, P&C, reinsurance and brokerage companies and found about two-thirds with multiple AI applications in production across core operations, according to Insurance Business on October 18. AI accounts for 2% to 10% of IT budgets at most, with 13% of respondents above 10%; nearly one in five P&C carriers spend more than 10%, against 7% of life insurers.

Measurement lags deployment. Only 41% of respondents quantify and track benefits for individual use cases, 34% quantify but do not track consistently, and 24% describe the benefits as mainly qualitative. Moody's expects the credit effect to be neutral to positive depending on execution, data quality and governance.

Chubb is in the two-thirds. Asked by Meyer Shields of KBW about its own stack, Greenberg described "foundational tech and data and smart layer of technology, algorithmic, large language, and agentic related" tools, and told Gregory Peters of Raymond James that the company runs "multiple layers of detection and defense against third party or external use of agentics and AI."

That puts the certification regime Greenberg imagines on both sides of the ledger. A carrier that excludes AI liability from its CGL policies while running agentic systems in its own underwriting and claims is an operator of the models a certification standard would govern, with its own errors-and-omissions and directors-and-officers exposure attached. The industry Greenberg says must put its arms around the risk is, on Moody's count, still working out how to measure what its own models return.