CG 40 47 on Your Global Liability Tower: Finding the AI Exclusion Before Your FY27 Renewal Does

In January 2026, Verisk's ISO Core Lines Services made three new endorsement forms addressing generative artificial intelligence available to member carriers across the United States in a multistate filing, as reported by Claims Journal on 20 July 2026. Each form is optional. No carrier is required to adopt any of them, which is exactly why the change is easy to miss: adoption happens carrier by carrier, renewal by renewal, with no single announcement that a risk manager in Mumbai or Bengaluru would see.

The signal that adoption is accelerating came on 17 August 2026, when Insurance Journal's magazine issue reported that the three ISO exclusions have garnered more interest from carriers, in an article titled Insurer Interest in AI Coverage Exclusions Growing as Risk Becomes Omnipresent. Interest from carriers in an optional exclusion translates, over the following renewal cycles, into the exclusion appearing on endorsement schedules.

For an Indian group with a US subsidiary carrying its own Commercial General Liability policy, or a global liability programme whose US local policies follow ISO wording, this is not an abstract US market story. It is a wording change that can attach to your tower at the next renewal without anyone flagging it. The broker's cover note will summarise limits, premium, and headline terms. The exclusion will sit in the list of form numbers on the endorsement schedule, where nobody reads unless they are looking.

The two forms most relevant to a liability tower split along the CGL's own internal structure, according to the Verisk form descriptions reported by Claims Journal on 20 July 2026:

• CG 40 47 excludes bodily injury, property damage, and personal and advertising injury arising out of generative artificial intelligence under the Commercial General Liability Coverage Part. That reaches both Coverage A (bodily injury and property damage) and Coverage B (personal and advertising injury). The form carries its own definition of generative AI, so the scope of the exclusion turns on that defined term, not on what your engineers would call generative AI.
• CG 40 48 limits the generative-AI exclusion to personal and advertising injury only, that is Coverage B alone. Coverage A claims survive.

The difference between the two forms is the difference between losing advertising-injury cover for AI-generated marketing content and losing liability cover for any injury or damage a court connects to a generative-AI system anywhere in your operations. On a manufacturer whose design team uses generative tools, or a services business whose client deliverables pass through an AI drafting step, CG 40 47 is a materially different policy from CG 40 48 at the same premium.

A typical Indian multinational's liability programme has a master policy written in India, Singapore, or London, with locally admitted policies issued in each operating country. The US local policy is usually written on ISO forms by a US admitted carrier, precisely because that is what the admitted placement rules require. ISO endorsements enter the programme through that local policy.

Three features of this structure keep the change quiet:

1. Local renewals are delegated. The US subsidiary's finance team or the broker's US network partner handles the local renewal. What travels back to group risk management is a confirmation of bound terms, rarely the endorsement schedule itself.
2. The cover note summarises, the schedule governs. A new exclusion added by form number does not change limits, premium, or the named insured, so it does not surface in the summary documents a group risk manager actually reads.
3. Difference-in-conditions logic may not respond. Groups often assume the master policy drops down for anything the local policy excludes. Whether it does depends on the master's own wording, and if the master follows or adopts an equivalent AI exclusion, the DIC/DIL layer has the same hole as the local policy. This is the same silent-gap mechanism we described for Indian domestic placements in silent AI exclusions in commercial insurance, now arriving with an ISO form number attached.

This audit does not require coverage counsel to start. It requires the documents and a text search.

1. Collect the endorsement schedules, not the cover notes, for every US local policy in the programme, the global master, and every umbrella and excess layer. For policies bound but not yet issued, ask the broker for the list of forms and endorsements on the binder.
2. Search the schedules for the form numbers: CG 40 47, CG 40 48, and any other CG 40-series form referencing artificial intelligence (the January 2026 filing contained three forms, so check for a third AI-related form number on the schedule as well). ISO form numbers appear with or without spaces, so search both CG 40 47 and CG4047.
3. Search the policy text for the defined term. Where a carrier has not adopted the ISO form, a manuscript equivalent will still define its trigger. Search for "generative artificial intelligence", "artificial intelligence", "machine learning", and "AI" in the definitions and exclusions sections. A manuscript exclusion with a broad AI definition can be wider than CG 40 47 itself.
4. Map each hit to Coverage A or Coverage B. Record, for every policy in the tower, whether the exclusion reaches bodily injury and property damage, or only personal and advertising injury.
5. Check the excess layers' follow-form position. An umbrella that follows form inherits the primary's exclusion. An umbrella with its own AI exclusion can be narrower or wider than the primary's, which creates a mismatch in the middle of the tower: a loss covered at primary but excluded at excess, or the reverse.

The CGL's two main insuring agreements respond to different fact patterns, and the choice between CG 40 47 and CG 40 48 decides which of those patterns you keep.

Coverage B scenarios, gone under either form: marketing copy drafted by a generative model that a competitor alleges disparages its product; an AI-generated advertising image that allegedly infringes another party's rights in a way that falls within personal and advertising injury; a chatbot-published statement alleged to be defamatory. These are the claims the narrower CG 40 48 is built to exclude.

Coverage A scenarios, which survive CG 40 48 but not CG 40 47: a component whose design process included a generative design tool fails in the field and injures a user; AI-generated installation instructions contribute to property damage at a customer site; an AI-assisted maintenance recommendation precedes an equipment failure that damages third-party property. Under CG 40 47, if the claimant pleads that the injury arises out of generative AI as the form defines it, the carrier has an exclusion to point at.

The phrase to focus on in any version of the wording is "arising out of". US courts have generally read that causation language broadly. A group whose workflows embed generative tools at many points should assume a claimant's pleading, or a carrier's reservation of rights, can connect more losses to AI than the group's own engineers would. Mapping where AI-connected losses would land across CGL, cyber, professional indemnity, and D&O is a separate exercise, and one worth doing alongside this audit; we set out the method in the AI gap map across liability towers.

Where a US local carrier insists on attaching one of the exclusions, the negotiation moves to carve-backs. Four positions, in descending order of value to the insured:

1. CG 40 48 instead of CG 40 47. If the carrier's underwriting concern is AI-generated content (defamation, advertising injury), the Coverage B-only form addresses it while preserving bodily injury and property damage cover. This is the single highest-value swap on the table.
2. A human-in-the-loop carve-back. Wording to the effect that the exclusion does not apply where the AI-generated output was reviewed and approved by a natural person before publication or use. For a group whose SOPs already require human sign-off on AI-assisted deliverables, this converts an existing control into recovered coverage.
3. Scheduled-use carve-backs. List the specific AI-assisted workflows the group runs (design assistance, document drafting, customer-service triage) and carve them back by schedule. Narrower than a general carve-back, but easier to obtain because the underwriter can price what is listed.
4. Causation tightening. Replace "arising out of" with "solely caused by" or "directly resulting from" in the exclusion trigger. This keeps mixed-causation losses, where AI is one contributing factor among several, on the covered side of the line.

Where the exclusion cannot be softened, the residual exposure becomes a candidate for affirmative AI liability cover placed alongside the CGL; the emerging options are covered in affirmative AI liability insurance for Indian groups and GCCs. Buying back a named exposure is more expensive than never losing it, so the carve-back conversation comes first.

Indian groups tend to anchor programme renewals to 1 April, but US local policies frequently run on calendar-year or placement-anniversary terms, so exclusion adoption can reach your tower at several points between now and the FY27 master renewal. The audit is worth running now rather than at renewal, because a form attached at a mid-cycle local renewal is already live.

A workable division of labour:

• Group risk management owns the document collection and the form-number search, and maintains a one-page register: policy, carrier, AI exclusion present or absent, form number or manuscript, Coverage A or B scope, follow-form status of each excess layer.
• The broker confirms, in writing, each US carrier's adoption position on the January 2026 ISO filing and whether the carrier intends to attach any of the three forms at the next renewal. Ask the question explicitly; silence in a renewal presentation is not an answer.
• Legal or coverage counsel reviews the defined term for generative AI in any exclusion found, against an inventory of the AI tools and workflows the group actually runs. The exclusion's definition, not the group's, controls.