ISO CG 40 47 AI Exclusion Forces a GL Rate Adequacy Rethink

From reviewing GL rate filings across multiple program years, the pattern playing out with generative AI exclusions closely mirrors the silent cyber problem that pricing actuaries spent the better part of five years unwinding after 2018. On January 1, 2026, Verisk ISO released three new endorsement forms: CG 40 47, CG 40 48, and CG 35 08. Each allows carriers to carve generative AI exposures out of commercial general liability policies. By April 2026, major carriers including W.R. Berkley, Chubb, Travelers, Berkshire Hathaway, and Cincinnati Financial had filed to adopt these endorsements or proprietary AI exclusion language, with state regulators approving more than 80% of submitted filings (IndependentAgent.com; PYMNTS, May 2026). One industry estimate projects that 95% of carriers will ultimately adopt some form of AI exclusion on their CGL books.

The pricing problem is immediate: 72% of organizations now use AI in at least one business function (McKinsey, 2024), yet 42% of P&C insurers track no AI-specific claims metrics at all (Capgemini, March 2026). When underwriters attach CG 40 47 and exclude an exposure that pervades the insured's operations, the historical loss data underlying the rate indication still contains whatever AI-related claims were paid under the prior policy form's silence on AI. The actuary must quantify and remove that embedded exposure before certifying that the post-exclusion rate is adequate for the narrower coverage grant.

What the three endorsements exclude

All three forms share an identical definition of generative artificial intelligence: "a machine-based learning system or model trained on data with the ability to create content or responses, including but not limited to text, images, audio, video or code." The definition captures large language models, image generators, code assistants, and any system that produces novel output from learned patterns. It does not reach traditional predictive models that classify or score without generating content, though the line between generative and predictive AI grows blurrier in production deployments.

Form	Applies To	Excludes	Preserves
CG 40 47	CGL Coverage Part	Coverage A + Coverage B	Nothing AI-related
CG 40 48	CGL Coverage Part	Coverage B only	Coverage A (BI, PD)
CG 35 08	Products/Completed Ops	Section I (BI, PD from AI in products)	Premises/operations

CG 40 47 is the broadest option. It applies to both occurrence and claims-made CGL forms and excludes both Coverage A (bodily injury and property damage) and Coverage B (personal and advertising injury) for any loss "arising out of" generative AI. CG 40 48 targets Coverage B only, removing personal and advertising injury protection while preserving bodily injury and property damage coverage. CG 35 08 applies to the products/completed operations liability coverage part, excluding downstream product liability theories when generative AI is embedded in delivered products.

The "arising out of" scope problem

Each endorsement triggers on losses "arising out of" generative artificial intelligence. Under established insurance law, "arising out of" requires only a causal connection, not proximate causation. This is one of the broadest causal phrases available in policy drafting. A Policyholder Pulse analysis published April 13, 2026, by attorneys Hadhy Ayaz and Jay Konkel at Pillsbury, warned that if AI exclusions become universally applied, "almost any claim will arguably be connected to AI use," potentially rendering policies illusory under the anti-illusory coverage doctrine (Policyholder Pulse, April 2026).

For pricing actuaries, the scope question is not academic. The effective scope of the exclusion determines the loss elimination ratio. If courts interpret "arising out of" broadly and deny coverage for any claim with even an incidental AI nexus, the amount of historical loss experience that should be attributed to the excluded exposure is larger than if courts narrow the exclusion to claims where AI was the direct, primary cause. Actuaries building a loss elimination factor must model the exclusion's effective scope under the jurisdiction's prevailing interpretation of "arising out of," not just the endorsement's face language.

Several carriers have gone further than ISO's language. W.R. Berkley filed an independent exclusion (Form PC 51380) covering D&O, E&O, and fiduciary liability with language targeting "any actual or alleged use of AI," a scope broader than ISO's generative-only definition. Hamilton Insurance Group filed language excluding claims "based upon, arising out of, or in any way involving" generative AI tools, explicitly naming ChatGPT, Bard, and Midjourney. These proprietary variants create different effective scopes and require different loss elimination assumptions in the rate indication.

Silent AI coverage and the historical loss contamination problem

Before January 2026, standard CGL policies neither explicitly granted nor denied coverage for AI-related losses. This created "silent AI" exposure: carriers paid claims they never specifically priced for, and insureds received coverage they never explicitly purchased. The pattern is structurally identical to silent cyber, which Swiss Re, Lloyd's, and others spent years resolving through affirmative cyber exclusions and standalone products beginning around 2019.

The pricing consequence is direct. Every GL rate indication filed in the past five years was developed using historical loss experience that includes whatever AI-related claims were paid under the silent coverage regime. If a carrier now attaches CG 40 47 and removes generative AI from the coverage grant, the historical experience period's losses are overstated relative to the narrower post-exclusion policy form. The actuary must estimate and remove the AI-related loss component before restating rate adequacy.

This is not a trivial exercise. Claim files from 2020 through 2025 were coded under legacy cause-of-loss taxonomies that predate any AI-specific classification. An advertising injury claim triggered by AI-generated marketing copy that infringed a competitor's trademark would have been coded as a standard Coverage B advertising injury claim. A bodily injury claim arising from AI-generated product instructions would appear as a standard products liability claim. The AI nexus is invisible in the data.

Constructing a loss elimination ratio for AI exposure

The loss elimination ratio (LER) provides the mechanical framework for adjusting historical losses to reflect a narrowed coverage scope. In standard ratemaking, the LER measures the proportion of ground-up losses eliminated by a policy provision such as a deductible. The concept extends naturally to coverage exclusions: when a carrier adopts CG 40 47, the LER for the AI exclusion represents the share of historical GL losses attributable to the now-excluded generative AI exposure.

The formula is straightforward in concept:

LER_AI = L_AI / L_Total

where L_AI is the historical loss dollars attributable to generative AI exposure and L_Total is total historical GL losses. The post-exclusion indicated loss cost is then:

Indicated Loss Cost_{post-exclusion} = Indicated Loss Cost_{pre-exclusion} × (1 - LER_AI)

The challenge is estimating L_AI when carrier claim systems contain no AI-specific coding.

Step 1: Identify AI-adjacent claim narratives

The actuary should work with claims departments to perform a retrospective text search of claim narratives for AI-adjacent keywords: "chatbot," "automated content," "AI-generated," "language model," "algorithm," "machine learning output," "deepfake," and similar terms. This keyword search will not capture every AI-related claim, but it establishes a floor for the AI loss component. Coverage B claims (personal and advertising injury) are the primary target, particularly claims involving intellectual property infringement from AI-generated content, defamation from AI chatbot outputs, and privacy violations from AI data processing.

Step 2: Apply exposure-based frequency estimation

Where claim-level data is insufficient, the actuary can estimate AI exposure frequency using external benchmarks. McKinsey's 2024 survey found that 72% of organizations have adopted AI in at least one business function, with 65% regularly using generative AI specifically. Gallagher's 2026 survey found that one in five insurance professionals reported their insureds had already experienced losses linked to AI risk. The actuary blends these adoption rates with the insured population's industry profile to estimate the proportion of the GL book with active AI exposure, then applies a conditional claim frequency assumption to that subpopulation.

Step 3: Credibility-weight carrier data against industry benchmarks

The Capgemini survey finding that 42% of P&C insurers track no AI-specific claims metrics means that the carrier-specific component of the LER estimate will often carry low credibility. Standard Bühlmann credibility applies:

LER_AI = Z × LER_carrier + (1 - Z) × LER_industry

For carriers with no AI claims coding, Z approaches zero and the LER defaults to the industry benchmark. For carriers that have implemented AI-specific cause codes (even recently), the text-search results from Step 1 contribute to the carrier-specific estimate with a credibility weight that reflects the volume of coded claims. The industry benchmark itself must be constructed from the available survey data, litigation filing trends (701 cumulative GenAI lawsuits filed in the U.S. between 2020 and 2025, per Insurance Intel, growing at 137% year over year), and professional judgment.

Estimated LER ranges by coverage

Coverage	Estimated LERAI	Rationale
Coverage B (personal/advertising injury)	2%–6%	AI-generated content creates direct pathways for IP infringement, defamation, and privacy claims; higher for tech/media classes
Coverage A (bodily injury/property damage)	0.5%–2%	AI-directed physical outcomes remain rare; concentrated in manufacturing and product liability classes
Products/completed operations	1%–4%	Growing as AI becomes embedded in delivered software and physical products; higher for technology E&O-adjacent classes

These ranges are preliminary and will widen or narrow as AI-specific claims coding matures. The key discipline is documenting the method, assumptions, and credibility weights under ASOP No. 25 so that the LER can be updated as data develops, rather than embedding an unexamined assumption in the rate filing.

Rate adequacy implications: CG 40 47 versus CG 40 48

The choice between the broad exclusion (CG 40 47, both coverages) and the narrow exclusion (CG 40 48, Coverage B only) creates materially different residual risk profiles that require separate rate indications.

A carrier adopting CG 40 47 removes both the Coverage A and Coverage B AI loss components. The total LER is the sum of the Coverage A and Coverage B AI elimination factors, weighted by each coverage's share of total GL losses. For a typical commercial GL book where Coverage B represents roughly 15% to 25% of total losses, the combined LER under CG 40 47 might range from 1% to 3% of the overall GL indication.

A carrier adopting CG 40 48 removes only the Coverage B component. Since Coverage B is the primary channel for AI-generated content liability (IP infringement, hallucinated defamation, privacy violations from AI data processing), the LER as a percentage of Coverage B losses is higher than the blended rate, but the absolute premium impact is smaller because Coverage B is the smaller coverage part.

The pricing implication is bidirectional. Carriers adopting the broad CG 40 47 exclusion should see a modest rate reduction for the narrowed coverage, all else equal. The reduction is modest because AI-specific losses remain a small share of the overall GL loss cost in 2026. But the trajectory matters: as AI deployment accelerates across insured operations, the AI-related share of GL losses will grow, and the LER will need to be updated at each rate review.

Reinsurance treaty implications

When a carrier narrows its CGL coverage scope by excluding generative AI, its ceded reinsurance treaties written on the net retained GL portfolio require evaluation. Quota share treaties that follow the fortunes of the primary carrier's coverage grant will automatically reflect the exclusion, but loss corridor or excess-of-loss treaties may require explicit endorsement to confirm that AI-related losses are excluded from the treaty's coverage scope as well.

If the reinsurance treaty does not mirror the primary exclusion, the carrier may face a gap: a claim denied under the direct policy's AI exclusion that the reinsurer also declines to cover because the reinsurance agreement was not updated to reflect the primary form change. Treaty actuaries should review the "follow the terms" and "follow the settlements" clauses in existing agreements and, where necessary, request manuscript endorsements confirming alignment with the ISO AI exclusion forms.

The ceded premium allocation must also be adjusted. If the primary carrier reduces its GL rate indication by the AI LER, the ceded premium to the reinsurer should reflect the narrower loss cost basis. Failure to adjust the ceded premium means the carrier is overpaying for reinsurance on a risk portfolio that has been reduced by exclusion.

The emerging affirmative AI liability market

The excluded exposure does not disappear. It migrates to standalone AI liability products, sits uninsured on corporate balance sheets, or becomes the subject of coverage disputes under pre-exclusion policy years that remain silent on AI. Standalone AI liability insurance is now a small but growing market with three products live and bindable as of April 2026: Armilla AI (Lloyd's coverholder, limits to $25M), Munich Re (standalone AI liability, limits to $50M), and Corgi Insurance (full-stack carrier covering E&O, cyber, and GL from AI errors). Market size projections vary from $4.8 billion by 2032 (Gallagher) to $34 billion by 2034 (DataIntelo, 19.6% CAGR).

Standalone AI liability products lack the actuarial loss cost data that decades of GL experience provide. Rating algorithms for these products will likely start with exposure-based approaches: number of AI models deployed, output volume (tokens generated, images produced, API calls served), governance posture (model inventory completeness, bias audit frequency, human-in-the-loop protocols), and industry classification. As claims experience develops, these exposure-based rates will be credibility-weighted with emerging loss experience, following the same maturation path that cyber insurance took from 2012 to 2022.

What pricing departments should do now

The GL rate filing cycle does not wait for perfect data. Pricing actuaries working on 2027 rate indications for CGL books that will carry AI exclusions should take four immediate steps:

1. Establish AI-specific claims coding. Work with claims departments to add a generative AI cause-of-loss indicator to the claim coding taxonomy. Even a simple binary flag ("AI nexus: yes/no") on new claims begins building the data needed to refine the LER at future reviews.
2. Perform a retrospective claim narrative review. Search historical claim files for AI-adjacent keywords to estimate the silent AI loss component in the current experience period. Document the methodology and results to support the LER selection in the rate filing.
3. Select an initial LER with documented credibility weights. Use the framework above to construct a credibility-weighted LER that blends carrier-specific text search results (where available) with industry adoption and litigation benchmarks. Document the assumptions under ASOP No. 25 credibility standards.
4. Coordinate with reinsurance and treaty pricing. Confirm that ceded reinsurance treaties align with the primary policy's AI exclusion and adjust ceded premium allocations to reflect the narrower coverage scope.

The AI exclusion is still early in its adoption cycle. The loss elimination ratios estimated today will be refined as AI-specific claims data accumulates and courts begin interpreting the "arising out of" trigger in actual coverage disputes. The actuarial discipline is not in getting the initial LER exactly right; it is in building a documented, reproducible framework that can be updated as data and legal precedent develop.