Silent AI Exclusions Hit Indian Commercial Policies in 2026: Where GL, D&O, Cyber and Tech E&O Now Fall Short

The change that reset the market did not originate in India. On 1 January 2026, ISO (the Verisk forms house whose language sits inside most commercial liability wordings globally) put three generative-AI endorsements into circulation. CG 40 47 is the broad form: it strips bodily injury, property damage and personal and advertising injury "arising out of generative artificial intelligence" from Coverage A and Coverage B of the standard commercial general liability form. CG 40 48 removes personal and advertising injury alone under Coverage B. CG 35 08 carves the same exposure out of products and completed-operations cover. Together they convert what underwriters call silent AI, exposure that was neither priced nor excluded, into an explicit exclusion a carrier can attach at renewal.

Adoption moved fast. By April 2026, W.R. Berkley, Chubb, Travelers, Berkshire Hathaway and Cincinnati Financial had filed to use these forms or their own proprietary AI language, and US state regulators had approved more than 80 percent of the filings on record.

Why should an Indian broker care about a US forms development? Three reasons. First, most large Indian corporates buy D&O, cyber and Tech E&O on global or controlled master programmes where the master wording follows international forms, so an AI exclusion attaches upstream and flows into the Indian local policy through a difference-in-conditions link. Second, Indian primary carriers reinsure liability and cyber offshore, and treaty wordings are already carrying AI exclusion clauses that push down into facultative and primary terms. Third, IRDAI's use-and-file route under the IRDAI (Insurance Products) Regulations, 2024 lets Indian insurers introduce endorsements without prior approval, so an equivalent Indian carve-out can land on a renewal quote with little warning. The wording, not the marketing brochure, decides whether an AI loss is paid.

Start with the general liability form, because it is where buyers assume a stray AI output would land and where they are most often wrong. A commercial general liability policy responds to bodily injury or property damage caused by an occurrence, defined as an accident. When a generative model produces a defamatory line, an infringing image, or a wrong instruction a customer relies on, the loss is usually financial rather than bodily injury or physical property damage, so it fails the trigger before any exclusion is reached. Coverage B, personal and advertising injury, is the one head that could respond, covering libel, slander or copyright infringement in an advertisement. That is precisely why CG 40 48 targets it.

Once CG 40 47 or CG 40 48 attaches, even that narrow Coverage B route closes. The operative phrase, "arising out of generative artificial intelligence", is broad by design. In Indian and English liability law "arising out of" imports a wide causal test, so a claim only loosely connected to an AI tool can be pulled inside the exclusion. A marketing team that used a generative tool to draft a campaign later alleged to be misleading could find the advertising-injury cover gone even though a human approved the final copy.

For Indian public-liability and product-liability buyers the practical effect is a shrinking grant. The base form already excluded most pure financial loss; the AI endorsement now removes the residual advertising-injury and product-defect routes an AI-authored output might have used. What looked like a general liability question is now a specialist-lines question.

Directors and officers cover is where the AI exclusion does the quietest damage, because the claims it triggers are exactly the ones boards now face. A shareholder or regulator alleging that directors failed to supervise an AI system, overstated AI capability in a prospectus, or ignored documented model-risk warnings brings a classic management-liability claim. On a D&O tower that would ordinarily run through Side A (non-indemnifiable loss of individuals), Side B (company reimbursement) and, for listed entities, Side C securities cover. If the D&O wording now carries an AI carve-out, the insurer can argue the wrongful act arose out of AI use and decline.

The trap is the interaction of two clauses. D&O forms contain a failure-to-supervise or oversight grant meant to respond to exactly these claims. An AI exclusion sitting alongside it can swallow the oversight claim the grant was sold to cover. Under the SEBI (LODR) Regulations, 2015 and the BRSR reporting framework, boards must disclose material technology and governance risks, so an AI-related misstatement in an annual report or an offer document is a live exposure, not a hypothetical. IPO-bound companies buying public-offering securities insurance (POSI) face the same question on the prospectus.

For mid-market and start-up boards the fix is a negotiated write-back that preserves cover for supervision and disclosure claims, even where the underlying operational AI act is excluded. Without it, the very digitisation the board is being praised for becomes an uninsured personal exposure for its directors.

This is the gap where most AI losses actually sit, and where the two relevant policies each point at the other. A cyber policy responds to a security failure: unauthorised access, data breach, network interruption. A technology errors-and-omissions policy (Tech E&O, the professional indemnity form for technology providers) responds to a failure to perform a technology service to the required standard. An AI failure that is neither a hack nor a clean professional error can fall between them. Two common patterns: a model that drifts over months and starts producing wrong outputs, and an automated decision that denies a loan or a claim with no human review.

Insurers are now writing exclusions on both sides. Cyber wordings are adding "AI event" or "artificial intelligence" exclusions that remove drift and output-error losses. Tech E&O forms are adding an AI carve-back that restores some cover, but usually with a sublimit, a higher self-insured retention, and conditions precedent such as documented human-in-the-loop review and model-change logs. Miss the condition and the carve-back lapses.

The Indian regulatory overlay sharpens this. An automated decision that produces a wrong or discriminatory outcome can trigger data-fiduciary liability under the Digital Personal Data Protection Act, 2023, and a security event still carries CERT-In six-hour reporting duties. A single incident can therefore generate a DPDP claim, a cyber notification and a Tech E&O professional-negligence allegation at once, with each insurer arguing the loss belongs on another tower.

The broker's job is to confirm that the drift-and-decision scenario is affirmatively granted somewhere in the programme, not merely absent from every exclusion. Silence in three wordings is not cover; it is three declinatures waiting to happen.

The one place an AI loss turns physical is the plant floor, and property wordings were not drafted with algorithms in mind. If an AI-driven control system mis-operates a boiler, a robotic line, a furnace or a grid-dispatch system and causes physical damage, the claim presents as a machinery breakdown, boiler explosion or fire loss rather than a liability loss. Material damage responds, and the resulting business interruption or loss of profits follows the reinstatement of the damaged item.

The question that decides the claim is proximate cause. Indian fire and engineering wordings, like their marine cousins, pay for loss proximately caused by an insured peril and exclude loss proximately caused by an excluded one. If the physical peril (fire, explosion, sudden mechanical failure) is the proximate cause, the property policy should respond even though an AI decision set the chain in motion. If an underwriter has added an AI exclusion to the property or machinery-breakdown section, the fight becomes whether the algorithm's instruction or the physical event was the dominant, effective cause.

Concurrent-cause reasoning matters here. Where an insured peril and an excluded cause combine, Indian courts generally test which cause was proximate in efficiency, not merely in time. An AI exclusion drafted with "arising out of" or "in connection with" language tries to widen that test in the insurer's favour.

Property AI exclusions are still rarer than liability ones, but they are appearing on technology-heavy risks: automated warehouses, lights-out manufacturing, and power-dispatch operators. For those buyers the electronic-equipment and machinery-breakdown sections deserve the same word-by-word reading now applied to the liability tower. Assuming the property policy is untouched by the AI debate is the mistake.

Treat the AI exclusion as negotiable wording, not fixed print. The 2026-27 renewal is the moment to shape it, using IRDAI's use-and-file route to seek a manuscript endorsement where the insurer will support one. Work through six points on every tower.

1. Definition scrutiny. Read how "generative AI" or "artificial intelligence" is defined. Push to narrow it to autonomous or content-generating systems and to exclude ordinary embedded software, so the carve-out does not swallow everyday tools.
2. Affirmative write-back. Rather than accept a bare exclusion, negotiate to restore cover for named use cases (for example, an internal drafting tool with human sign-off) under a stated sublimit.
3. Causal language. Resist "arising out of" and "in connection with". A tighter "directly caused by" test keeps loosely connected claims inside cover.
4. Conditions precedent. Where a carve-back attaches human-in-the-loop review, model logging or change control as a condition, make sure the client can actually evidence it, because a breached condition precedent voids the grant.
5. Cross-tower consistency. Lay the GL, D&O, cyber and Tech E&O wordings side by side and check that the same AI event is not excluded on all four, which would leave a total gap. The goal is that at least one tower affirmatively responds.
6. Expiring-versus-proposed comparison. The switch from silent to explicit happens quietly at renewal. Compare the proposed endorsement schedule against the expiring wording line by line and flag any new AI language in writing to the client before binding.

Document the negotiation as you go. If a carve-back is declined, a file note recording that the broker raised it protects the broker's own professional indemnity position if the client later suffers an uninsured AI loss. This discipline is slower than accepting the quoted terms, but the alternative is a client discovering the gap at claim stage, when nothing can be changed.

The pattern across all five lines is the same. The AI exclusion is invisible until you read the endorsement schedule, and the true gap only appears when the GL, D&O, cyber, Tech E&O and property wordings are laid side by side and tested against one incident. A single AI failure can be excluded on the liability form, sublimited on Tech E&O, silent on cyber and argued over on property, and no one policy document reveals that on its own.

This is where searchable wordings intelligence earns its place in the renewal file. Sarvada indexes insurer policy wordings and endorsements so a broker can search AI-exclusion and carve-back language across GL, D&O, cyber and Tech E&O forms in one place, compare an expiring wording against a proposed one, and see where a carve-out on one tower has quietly reopened an exposure on another. It turns a manual, error-prone read across five PDFs into a structured comparison the whole placement team can check, and it flags the definitional differences (how each insurer scopes "generative AI") that decide whether an everyday embedded tool is caught.

For Indian brokers the value shows up at two moments: when quoting, so the client sees the gap before binding rather than at claim stage, and when defending the broker's own file if a loss later falls between towers. If you are preparing 2026-27 renewals and want to map where AI cover falls short across a client's programme before you bind, Request Access to see how Sarvada surfaces the exact clauses that decide these claims.