The AI absolute exclusion decoded. ISO CG 40 47, CG 40 48, and the carrier-specific endorsements reshaping CGL in 2026.

What "absolute exclusion" actually means in insurance language

Policy exclusions come in several varieties. A standard exclusion removes coverage for a defined category of loss but may contain exceptions that restore coverage in specific circumstances. A pollution exclusion, for example, often preserves coverage for hostile fire events. A war exclusion may contain a carve-back for terrorism under certain conditions. These partial exclusions leave room for argument about whether the exception applies.

An absolute exclusion contains no such exceptions. It removes coverage for the defined category entirely, regardless of how the loss arose, how minor the causal connection was, or how sympathetic the facts. In insurance drafting, the term "absolute" is not always used explicitly, but the effect is the same when an exclusion is written without any exception language. Several of the AI exclusions now entering the market function as absolute exclusions in this technical sense.

The distinction matters because policyholders have traditionally argued around standard exclusions by pointing to their exception language. Where an exclusion reads "we do not cover X, except in circumstances Y," the insured's argument is that circumstances Y apply. Where the exclusion reads "we do not cover X, with no exceptions," that argument is foreclosed. The ISO CG 40 47 and CG 40 48 forms contain no exception language for generative AI losses. W.R. Berkley's PC 51380 endorsement explicitly adds "in any way involving" language, which courts have read as foreclosing even tangential causal connections. Both approaches function as absolute exclusions in practical effect.¹

The word "absolute" also signals something about the carrier's intent. When a carrier files an absolute exclusion with a regulator, it is communicating that it does not want to price AI risk into the general policy, it does not want to litigate AI causation questions at the claim stage, and it does not intend this to be a negotiable position at the renewal table. That intent shapes how the exclusion will be administered and how coverage disputes will be argued in court.

The ISO standard endorsements: CG 40 47 vs CG 40 48 vs CG 35 08

In January 2026, Verisk's ISO Core Lines Services made three new endorsement forms available to member carriers across the United States as part of a multistate filing.² The forms address generative AI exposures in commercial general liability coverage. Each form is optional: carriers choose whether to adopt them, and buyers should not assume that their carrier has or has not adopted them without checking the endorsement schedule of their current policy.

CG 40 47 01 26: the full CGL exclusion

CG 40 47 is the broadest of the three forms. When attached to a standard ISO CGL policy, it removes Coverage A and Coverage B simultaneously for losses arising out of generative artificial intelligence. Coverage A covers bodily injury and property damage. Coverage B covers personal and advertising injury, which includes defamation, wrongful eviction, copyright infringement in advertising, and reputational harm.

The practical scope of CG 40 47 is wide. A customer-facing AI agent that provides incorrect medical information, causing a user to be physically harmed, would previously have been a Coverage A claim. Under CG 40 47, that claim is excluded. An AI marketing tool that generates advertising copy containing a third party's copyrighted phrase, exposing the business to an infringement claim, would previously have been a Coverage B claim. Under CG 40 47, that claim is also excluded.³

CG 40 47 is applicable to both the occurrence version and the claims-made version of the ISO CGL form. The endorsement attaches at the policy inception date of the renewal. Any loss that occurs after the endorsement attaches is subject to the exclusion, even if the AI system was deployed before the endorsement took effect.

What CG 40 47 does not exclude are losses that have no causal connection to generative AI. A premises liability slip-and-fall has nothing to do with AI output and is not affected. A traditional advertising injury claim arising from a human copywriter's work is not affected. The exclusion is triggered by the causal connection to generative AI, not by the mere presence of AI technology in the insured's business.

CG 40 48 01 26: Coverage B only

CG 40 48 is the narrower form. It removes Coverage B only, leaving Coverage A intact. A carrier attaching CG 40 48 rather than CG 40 47 has made a deliberate underwriting decision: it will still respond to bodily injury and property damage claims with an AI causal link, but it is not prepared to cover content-related liability arising from AI outputs.

For operators running generative AI in content-heavy contexts, CG 40 48 represents a significant gap. Coverage B is where defamation claims, copyright infringement claims, and reputational damage claims live. These are among the most common categories of claim arising from AI-generated text, images, and code. A business that generates customer-facing content using AI tools and holds a CGL policy with CG 40 48 attached has effectively lost coverage for the most likely category of AI-related claim.⁴

At the same time, CG 40 48 preserves more coverage than CG 40 47 for operators whose AI risk is concentrated in physical or operational contexts rather than content production. A logistics company using AI to route vehicles has a different risk profile from a marketing agency using AI to produce copy. CG 40 48 is a better outcome for the logistics company than CG 40 47 would be.

CG 35 08 01 26: Products and Completed Operations

CG 35 08 operates on the ISO Products/Completed Operations Coverage Part, which is a separate coverage part from the main CGL form. It excludes Section I coverage for bodily injury and property damage arising from generative AI in the products and completed operations context. This means that if a product your company designed, manufactured, or supplied incorporates generative AI and that product causes harm after it leaves your premises, the Products/Completed Operations coverage will not respond to the AI-related element of that claim.⁵

CG 35 08 is particularly relevant for technology companies, software developers, and manufacturers building AI into their products. A software vendor whose AI-enabled product causes property damage to a client's systems after installation would normally look to Products/Completed Operations coverage. Under CG 35 08, that coverage is removed if the damage arose from the AI component.

The three ISO forms can operate together or independently. A carrier could attach all three to a single policy, producing a CGL that excludes AI from Coverage A, Coverage B, and Products/Completed Operations. Or a carrier could attach only one, depending on its underwriting appetite for different segments of AI risk.

Carrier-specific endorsements vs the ISO standard

The ISO forms represent the industry's attempt to standardise AI exclusion language across the CGL market. But standardisation is not universally adopted. Several major carriers had already filed their own AI exclusion language before the ISO forms were released, and that language does not always align with the ISO standard definitions or scope.

W.R. Berkley Form PC 51380

W.R. Berkley filed Form PC 51380 in June 2024, making it one of the first major carriers to introduce an absolute AI exclusion across management liability product lines.⁶ The form covers D&O, E&O, and Fiduciary Liability products. Its exclusion language eliminates coverage for any claim arising from the actual or alleged use, deployment, or development of artificial intelligence by any person or entity.

The PC 51380 scope is notably broader than the ISO CGL forms in several respects. It covers not just AI-generated content but also the insured's failure to detect AI-created materials produced by third parties. It covers inadequate AI governance policies and training as an independent trigger, meaning a D&O claim arising from poor board oversight of an AI programme is excluded even if the AI itself never caused a loss. And it explicitly covers statements or disclosures about the company's AI capabilities or business plans, which creates an exclusion for D&O claims arising from investor communications about AI strategy regardless of whether any AI system was involved in the underlying facts.⁷

PC 51380 defines artificial intelligence broadly, without limiting itself to generative AI. It covers "any machine-based system that infers how to generate outputs," which reaches automated decision-making, predictive analytics, and algorithmic systems that predate the generative AI wave.

Hamilton Insurance Group

Hamilton Insurance Group filed a Generative Artificial Intelligence Exclusion for professional liability policies that takes a narrower approach than PC 51380.⁸ The Hamilton form removes coverage for claims "based upon, arising out of, or in any way involving" the use of generative AI, defined as systems producing text, imagery, audio, or synthetic data in response to user prompts. The definition specifically names ChatGPT, Bard, Midjourney, and DALL-E as examples.

The named-examples approach in the Hamilton form creates an interpretive issue. As new generative AI tools enter the market, insureds may argue that a tool not on the named list falls outside the exclusion. The counter-argument is that the definition also contains "including, but not limited to" language, which courts have generally read as making the named examples illustrative rather than exhaustive.

AIG and Great American

AIG filed a proposed AI exclusion with US state regulators, noting at the time that it had "no plans to implement" the exclusion immediately but wanted the option available as AI claim frequency increased.⁹ Great American Insurance filed equivalent language for its E&O and management liability lines. Both carriers' exclusion language targets claims arising from autonomous AI decision-making and AI-generated content, consistent with the general direction of the market.

The AIG filing is notable for its regulatory transparency. By filing the form in advance of implementation, AIG signalled that the question is not whether it will use an AI exclusion but when. Operators whose AIG policies do not currently contain an AI exclusion should confirm the position at the next renewal rather than assuming the current wording will persist.

Chubb

Chubb has taken a different approach from the absolute exclusion model. Its AI exclusion carves out losses from widespread or systemic AI events while retaining some coverage for isolated AI incidents.¹⁰ The distinction between a systemic event and an isolated incident is not precisely defined in publicly available Chubb language, but the practical effect is that a business suffering a single AI agent error may still have coverage under a Chubb CGL where it would have none under a policy with the ISO CG 40 47 form attached.

Chubb's approach reflects a reinsurance market concern: AI system failures could be correlated across thousands of policyholders simultaneously in a way that resembles catastrophe exposure. The systemic exclusion is Chubb's mechanism for managing that accumulation risk, not a general statement that AI losses are uninsurable.

Beazley and QBE: the sublimit approach

Beazley and QBE have moved in a different direction from the outright exclusion model, introducing sublimits that cap AI-related payouts at approximately 10 per cent of total policy limits rather than eliminating coverage.¹¹ On a cyber policy with total limits of GBP 3.8 million, this produces an effective AI sublimit of roughly GBP 380,000.

The QBE sublimit approach also specifically addresses LLMjacking, the practice where attackers compromise corporate AI accounts to exploit usage credits, capping payouts on those incidents at roughly GBP 188,000 on a GBP 5 million policy. Beazley's head of cyber underwriting management Aidan Flynn confirmed that the AI sublimit wording was in development as of April 2026 and had not yet been applied to in-force policies.¹²

The sublimit model is materially better for insureds than an absolute exclusion for businesses whose AI losses are likely to be modest and individual. But for businesses with significant AI exposure where a single AI agent failure could produce a seven-figure loss, a sublimit capped at 10 per cent of policy limits may provide less effective coverage than a standalone specialist AI product.

The five trigger phrases to look for in your renewal

AI exclusion language does not always use the phrase "artificial intelligence" in the heading. It may appear under headings like "Emerging Technology Exclusion," "Digital Content Exclusion," or "Automated Systems Exclusion." The substance of the exclusion is determined by the operative language, not the heading. These are the five phrases that signal the presence of an effective AI exclusion.

What survives an absolute exclusion

An absolute AI exclusion is not the end of the coverage analysis. Courts apply consistent interpretive principles to exclusion disputes, and several of those principles can preserve coverage even where the exclusion language appears to remove it entirely.

The first principle is narrow construction. Courts read exclusions narrowly and against the insurer as the drafter of the policy language. An exclusion that could be interpreted broadly or narrowly will be interpreted narrowly. An exclusion that could reach a particular loss or not will be read as not reaching it if the facts support that reading.

The second principle is independent causation. Where a loss has multiple causes, one of which is excluded and others of which are not, coverage may attach if the non-excluded causes are independently sufficient to produce the loss. In an AI-related claim where the plaintiff alleges both autonomous AI action and human negligence in deploying the AI, the human negligence may be an independent basis for the claim that survives the AI exclusion. This argument has been used successfully in analogous contexts in pollution exclusion litigation.¹³

The third principle is the mixed-claim analysis. A complaint that alleges AI-related wrongful acts alongside non-AI-related wrongful acts creates a mixed claim. The insurer may have a duty to defend the entire action if the non-AI allegations independently trigger coverage, even if the AI allegations do not. The duty to defend standard is lower than the duty to indemnify: it requires only that the complaint contain allegations that could potentially fall within coverage, not that those allegations will succeed at trial.

The fourth consideration is the attachment date. An AI exclusion takes effect when the endorsement attaches, which is typically the renewal date of the policy. A loss that occurred before the endorsement attached is governed by the prior policy's wording. If the prior policy was silent on AI, the loss may be covered under the prior policy even though it would be excluded under the renewed policy. Operators managing ongoing AI incidents that straddle a renewal date should review the attachment date carefully.

What does not survive an absolute exclusion is a claim where AI is the primary or sole cause, the loss occurs after the endorsement attaches, and the complaint does not include non-AI allegations that could independently trigger coverage. In that scenario, the exclusion will apply and coverage will be denied.

The negotiation room: rare but real

Buyers who accept AI exclusions at renewal without negotiation leave potential coverage on the table. The exclusions currently in the market are carrier filings, not statutory requirements. Carriers have discretion to modify them for specific accounts, and some are prepared to do so for accounts that present a favourable risk profile.

The practical factors that create negotiation leverage are: account size and premium volume, the quality of the buyer's documented AI governance programme, the quantified scope of AI use within the business, and the existence of an AI risk register that the carrier can use to price the residual exposure. A buyer that presents a board-approved AI governance policy, a named AI officer or responsible AI team, a documented scope of all AI deployments, and a quantified estimate of maximum probable AI-related loss is in a fundamentally different negotiating position from a buyer that has not addressed AI governance at all.¹⁴

The negotiation outcome, where it is achievable, is a buyback endorsement that reinstates coverage for defined AI losses on specific terms. The buyback should be defined by form number and should specify exactly which coverage parts are reinstated, which categories of AI loss are covered, which exclusions from the buyback apply, and the sublimit that applies to reinstated AI coverage. A general description in a renewal letter is not enforceable in the same way as a policy endorsement.

For most SMEs, the more practical path is not negotiation but supplementation. A specialist AI insurance product from a carrier that writes affirmative AI coverage on defined terms fills the gap that the generalist carrier's exclusion creates, without requiring the insured to negotiate against a carrier that has already filed its position with the regulator.

When the exclusion is silent vs explicit: why both matter

Before the Verisk ISO forms and the carrier-specific exclusion filings, most standard CGL, E&O, and cyber policies were silent on AI. A silent policy is one that neither explicitly grants nor explicitly excludes coverage for AI-related losses. The silence created an ambiguous coverage position that could be argued in either direction when a claim arose.

The ambiguity cut both ways. Insureds could argue that a loss arising from AI activity was covered because the policy did not exclude it. Carriers could argue that the same loss was not covered because the policy was written before AI existed as a coverage category and the parties could not have intended to cover it. The outcome depended on the specific facts, the jurisdiction, and the court's interpretation of the policy language.¹⁵

Reinsurers pushed primary carriers to resolve this ambiguity. Munich Re and Swiss Re, among others, began requiring cedents to clarify their AI positions: either add an affirmative AI coverage endorsement that explicitly granted coverage on defined terms, or add an exclusion that removed AI losses from the standard policy. The ISO Verisk forms are the standardisation of this push at the primary market level.

For operators, the shift from silence to explicit exclusion is a significant development. A silent policy preserved the argument for coverage. An explicit exclusion eliminates it. A policy renewed in 2026 that contained an AI exclusion the insured did not notice has a fundamentally different coverage position from the policy it replaced, even if all other terms are identical. This is why reading the endorsement schedule at renewal is not optional in 2026.

At the same time, a policy that remains silent in 2026 should not be read as an affirmative grant of AI coverage. The market has moved far enough that carriers are administering silent policies conservatively: an insurer whose policy is silent on AI will argue that the policy was not intended to cover AI losses and will look for causation arguments to decline the claim. Silence in 2026 is better than an explicit exclusion, but it is not the same as affirmative coverage.

How to map an exclusion to a specialist coverage gap fill

The structured response to an AI exclusion in a standard policy is a gap analysis followed by a specialist coverage placement. The gap analysis identifies which categories of AI loss are excluded from the standard policy. The specialist placement fills the identified gaps with affirmative coverage on defined terms from a carrier that is pricing AI risk rather than excluding it.

The categories of AI loss that most commonly need gap-filling after CG 40 47 or equivalent exclusions attach are: Coverage B losses arising from AI-generated content (defamation, copyright infringement, reputational damage); Coverage A losses arising from AI decisions that cause physical harm or property damage; E&O losses arising from AI agent errors in professional services contexts; and regulatory penalty exposure under the EU AI Act and the revised Product Liability Directive.

The carriers currently writing affirmative AI coverage for these categories include Armilla (Lloyd's coverholder, up to USD 25 million per company), Munich Re aiSure (parametric AI performance coverage), HSB (Hartford Steam Boiler, Munich Re subsidiary, for bodily injury and advertising injury from AI use), Counterpart (professional liability with affirmative AI coverage), and Corgi (full-stack AI-native coverage for technology companies). The full Carrier Comparison Matrix, which maps each carrier's product against the coverage categories created by the exclusion wave, is available at agentinsured.eu.

The mapping exercise should produce a coverage architecture document that shows, for each category of potential AI loss, which policy is intended to respond and the limits available. Gaps in that architecture are uninsured exposure. The document also serves as evidence of a documented AI risk management programme, which is relevant both for negotiating with generalist carriers and for the EU AI Act's Article 26 operator obligations.

See the Coverage Audit tool to run a structured gap analysis against your current policy schedule. The tool maps your policy terms against the exclusion categories created by the ISO forms and produces a gap register formatted for broker submission.

The five-step renewal review

The following five steps represent the minimum review that any operator running AI agents should complete at the next CGL, E&O, or cyber policy renewal.

1. Collect the full policy document, not just the schedule. The endorsement schedule, definitions section, conditions, and exclusion schedule together determine the actual scope of coverage. Requesting only the policy summary or the declarations page is insufficient. Ask your broker for the full policy wording including all endorsements and their attachment dates.
2. Search the endorsement schedule for the five trigger phrases. Use the annotated checklist above. Run a text search for each phrase. If any of the phrases appear, locate the endorsement in which they appear and read the full operative clause, including the definition section. Confirm whether the endorsement is CG 40 47, CG 40 48, CG 35 08, or a carrier-specific form like PC 51380.
3. Map your AI use against the excluded categories. List every AI tool or agent your business currently uses. For each one, identify whether it produces content (Coverage B exposure), takes autonomous actions (Coverage A and E&O exposure), is incorporated into a product delivered to clients (Products/Completed Operations exposure), or is described in investor or board communications (D&O exposure under PC 51380-type exclusions). Match each use case against the exclusion categories you found in step two.
4. Identify the uninsured exposure and quantify it. For each category of AI use that falls outside your current coverage, estimate the maximum probable loss if that use case produces a claim. Use the Moffatt v. Air Canada refund quantum as a reference for customer-facing agent errors, and the Mata v. Avianca sanctions cost as a reference for AI-generated content in professional services contexts. The quantification does not need to be precise but it must be defensible enough to support a broker submission.
5. Initiate the specialist coverage conversation at least 90 days before renewal. The specialist AI insurance market is not yet commoditised. Placement takes longer than standard commercial lines, underwriters require detailed application information about AI governance, and the carrier options are fewer than in mature product lines. Starting 90 days out gives enough time to complete underwriting, negotiate terms, and bind coverage before the renewal date. Starting two weeks before renewal gives you none of those options.

Frequently asked questions

What is ISO CG 40 47?

ISO CG 40 47 01 26 is the Exclusion: Generative Artificial Intelligence endorsement published by Verisk (Insurance Services Office) with an effective date of January 2026. It is an optional endorsement for ISO Commercial General Liability Coverage Part policies in both occurrence and claims-made versions. When attached, it removes Coverage A (bodily injury and property damage) and Coverage B (personal and advertising injury) for losses arising out of generative artificial intelligence. Generative AI is defined as a machine-based learning system or model trained on data with the ability to create content or responses, including text, images, audio, video, or code.

What is the difference between CG 40 47 and CG 40 48?

CG 40 47 01 26 is the broader form, excluding both Coverage A and Coverage B for generative AI losses. CG 40 48 01 26 is narrower, excluding Coverage B only. Coverage B covers personal and advertising injury, including defamation, copyright infringement, and reputational damage from AI-generated content. Coverage A, covering bodily injury and property damage, remains intact under CG 40 48. A carrier attaching CG 40 48 will still respond to physical harm claims traceable to AI but will not cover content-related liability.

What does ISO CG 35 08 exclude?

CG 35 08 01 26 applies to the Products/Completed Operations Coverage Part. It excludes bodily injury and property damage arising from generative AI in the completed operations context. If a product incorporating generative AI causes harm after delivery, the Products/Completed Operations coverage will not respond to the AI-related element. The form operates independently from CG 40 47 and CG 40 48: a policy could carry all three simultaneously.

What does the W.R. Berkley PC 51380 exclusion cover?

W.R. Berkley Form PC 51380, filed in June 2024, is an absolute AI exclusion for D&O, E&O, and Fiduciary Liability. It excludes any claim arising from the actual or alleged use, deployment, or development of artificial intelligence. The scope covers AI-generated content, failure to detect AI-created materials, inadequate AI governance, chatbot misrepresentations, regulatory investigations, and statements about the company's AI capabilities. The last category reaches D&O claims from investor communications about AI strategy, even where no AI system was deployed.

What is an AI absolute exclusion in insurance language?

An absolute exclusion removes coverage entirely for a defined category of loss, with no exceptions. It differs from a standard exclusion, which may contain carve-backs or exceptions that restore coverage in specific circumstances. The ISO CG 40 47 and CG 40 48 forms contain no exception language for generative AI losses. W.R. Berkley's PC 51380 adds "in any way involving" language that forecloses even tangential causal connections. Both function as absolute exclusions in practical effect.

Which trigger phrases should I search for in my CGL policy?

Five phrases most commonly activate AI exclusions: (1) "arising out of" generative AI, (2) "based upon, arising out of, or attributable to" AI, (3) "in any way involving" AI, (4) "automated or algorithmic decision-making," and (5) "machine-based learning system or model." The phrase "arising out of" requires only a minimal causal connection. "In any way involving" is even broader. Finding any of these in an endorsement schedule requires reading the full operative clause to determine which coverage parts are affected.

Does an absolute AI exclusion remove coverage for employee misuse of AI?

Most current absolute AI exclusion language does not distinguish between autonomous AI decisions and employee misuse of AI tools. The ISO CG 40 47 definition covers any generative AI system, including standard tools like Microsoft Copilot, ChatGPT, and Gemini used by staff. If an employee uses one of those tools, produces a defamatory output, and that output causes a third-party claim, the exclusion applies even though the AI did not act autonomously. The key question is whether the loss "arose out of" the AI system, not whether the AI acted independently.

Can I negotiate an AI exclusion off my policy?

In some cases, yes. Buyers with documented AI governance programmes, named AI officers, board-level AI oversight, and a quantified AI risk register have successfully negotiated buyback endorsements reinstating coverage for defined losses. The negotiation is more viable for accounts with significant premium and for carriers that have not yet adopted the ISO standard forms. For SMEs, a specialist AI insurance product typically fills the gap more efficiently. Any buyback should be defined by form number, not by a general renewal letter description.

What survives an absolute AI exclusion?

Courts interpret exclusions narrowly and against the insurer. Three categories may survive: mixed claims where the complaint alleges both AI-related and non-AI wrongful acts that are independently actionable; claims where the causal connection to AI is too remote to satisfy the "arising out of" standard; and claims where the loss predates the endorsement's attachment date. What does not survive is a claim where AI is the primary cause, the loss occurs after attachment, and the complaint contains no non-AI allegations that could independently trigger coverage.

What is the difference between a silent AI policy and an explicit AI exclusion?

A silent AI policy neither grants nor excludes AI coverage, leaving claims in an ambiguous middle ground. An explicit AI exclusion resolves that ambiguity against the insured. Reinsurers including Munich Re and Swiss Re pushed carriers to resolve silent AI exposure, leading to the Verisk ISO forms and carrier-specific endorsements. For operators, a silent policy preserves the argument for coverage while an explicit exclusion eliminates it. A 2026 policy that remains silent should not be read as an affirmative grant of AI coverage: carriers are administering silent policies conservatively.

Related reading

• AI policy exclusions: what SME operators must review before renewing - the four exclusion types that matter most for AI agent operators, with worked examples from Moffatt v. Air Canada and Mata v. Avianca.
• Does my business insurance cover AI errors in 2026? - a coverage category map for CGL, cyber, E&O, and professional indemnity policies, with a plain-language guide to where AI agent risk currently sits.
• Does your insurance cover AI mistakes? - the foundational explainer on AI coverage gaps for operators new to the coverage question.
• The 90-day AI agent insurance compliance playbook - a sequenced action plan from coverage gap analysis to binding specialist cover, calibrated to the EU AI Act August 2026 deadline.
• Coverage Audit tool - a structured self-assessment that maps your current policy wording against the exclusion categories created by the ISO forms and produces a gap register for broker submission.