Verisk CG 40 47 Creates an AI Liability Pricing Gap: From GL Exclusion to Standalone Market

On January 1, 2026, Verisk's ISO Core Lines Services released three endorsement forms carving generative artificial intelligence out of commercial general liability: CG 40 47, CG 40 48 and CG 35 08. Within weeks at least six insurers filed to adopt them or their own variants.

The pricing consequence runs in two directions. A policy carrying the exclusion should carry a lower expected loss load. The excluded exposure does not disappear with it.

Key Takeaways

• CG 40 47 is the total carve-out, removing both Coverage A and Coverage B for any loss "arising out of" generative AI, while CG 40 48 takes Coverage B only and CG 35 08 addresses products and completed operations.
• "Arising out of" needs only a causal connection, not proximate causation, so a claim does not have to be caused by generative AI to be excluded. It has to arise out of its use.
• ISO forms underpin roughly 82% of U.S. property and casualty policies, which is why an endorsement released in January plausibly reaches most commercial GL books by the end of 2027.
• 0.5 to 2.0 percentage points of expected loss ratio is a defensible near-term range for the AI load being removed, and across a 50,000-policy book even one point moves the indicated rate level.
• 74% of small and midsize businesses already use AI programs on HSB's count, with 91% planning to, so the exposure base is growing faster than the loss data that would let anyone price it from experience.

Three Forms, One Phrase

All three endorsements work off the same definition and the same causal language, and the language is doing more work than the definition.

Generative AI is defined across the forms 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 reaches large language models, image generators and code assistants, not predictive models that classify or score without generating content.

Form	Applies To	Excludes	Preserves
CG 40 47	CGL Coverage Part	Coverage A + Coverage 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

CG 40 47 01 26 is the broadest, applying to both occurrence and claims-made CGL coverage parts and excluding Coverage A and Coverage B alike. A manufacturer whose AI-generated product instructions cause injury, a retailer whose chatbot defames someone, a consultancy whose AI-drafted report contains errors: none has CGL coverage under it. CG 35 08 01 26 reaches products and completed operations, so once generative AI is embedded in a delivered product, downstream defect and failure-to-warn theories fall outside Section I.

The operative phrase in all three is "arising out of," which coverage law reads as requiring a causal connection rather than direct or proximate causation. That breadth is what makes the forms hard to price against, because it captures claims where generative AI is one contributing element rather than the cause.

Carrier wording is already going further. W.R. Berkley filed language barring claims related to "any actual or alleged use of AI, regardless of whether the model was company-owned, third-party, licensed, or embedded in software tools," which reaches past the ISO definition of generative AI entirely. Cincinnati Financial, Frederick Mutual and Philadelphia Insurance have filed their own wording, and AIG and Great American have sought clearance across D&O, E&O, EPLI and fiduciary lines.

How Much Loss Load Comes Out

The exclusion poses a question that has no experience-based answer yet, and the rate indication needs a number anyway.

Endorsing CG 40 47 removes AI-related claims from the coverage grant, so those claims come out of the expected loss load used to rate the policy. The difficulty is that GL loss development triangles contain no meaningful volume of AI-specific claims, and the expected loss ratios in current rate levels were fitted to a pre-AI commercial environment.

A near-term AI load in the range of 0.5 to 2.0 percentage points of expected loss ratio is defensible by analogy to technology E&O and media liability, with technology, media and professional services firms at the upper end and manufacturing and construction at the lower. It is judgment, and a rate filing has to say so.

The range is small per policy and not in aggregate. Across 50,000-plus commercial GL policies a one-point reduction moves both the indicated rate level and the underwriting margin, and the load grows with adoption: HSB reports 74% of small and midsize businesses already using AI programs, with 91% planning to.

Cyber gives the timing. ISO introduced CG 21 06 and CG 21 07 in 2014 to strip data-related liability out of CGL, and the risk migrated rather than vanished.

Standalone cyber, barely a line of business in 2013, reached $15.3 billion of global premium by 2024 with $10.6 billion of that in North America, and policies in force grew from roughly 2.2 million to over 3.6 million between 2016 and 2019. Standalone cyber became a mainstream commercial product about four to five years after the exclusions started. The same lag from a January 1, 2026 effective date is the window in which GL books will be shedding load faster than an affirmative market exists to absorb it.

The Exposure the Exclusion Does Not Remove

Two features of the current book mean the excluded risk stays in the portfolio in forms the exclusion was not written to catch.

The first is vendor-embedded AI. Enterprise software from Salesforce, Microsoft, Google, HubSpot and others now ships generative AI features that are often on by default, so an insured may not know it is using generative AI in the sense that triggers the exclusion. That produces selection inside the rated book rather than across it: buyers who understand the exclusion go and buy affirmative cover or self-insure, and buyers who do not stay in the book at a rate that has already had the AI load taken out of it. The exclusion removes the coverage but leaves the second group's expectation of coverage intact until a denial.

The second is the tower. If a primary policy excludes AI and the excess and umbrella layers follow form, the exclusion passes through and the tower is clean. If they do not follow form, the excess layers silently retain exposure the primary carrier has shed, and loss development appears in exactly the layers the primary exclusion was meant to protect. Market wordings do not yet resolve the follow-form question on AI consistently, which makes it a diligence item on every excess placement rather than a treaty-level assumption.

The correlation structure is what makes both of these matter more than the per-policy load suggests. Cyber losses are high frequency and moderate severity, correlated through shared technical infrastructure. AI losses correlate through shared model architectures instead: a single foundation model update that produces defective output propagates across thousands of commercial deployments at once, which behaves more like an accumulation event than like traditional liability. Whether those correlated claims land inside or outside a reinsured portfolio depends on the endorsement status of each underlying policy, one at a time.