Silent AI Exclusions in Commercial Policies
Insurers are quietly cutting AI coverage through exclusions that don't change policy limits.
Commercial policies are losing “silent AI” coverage through exclusions and sublimits that can change protection without changing the headline policy limit. For companies deploying autonomous or AI-driven hardware, the conclusion is direct. The next renewal must establish what each policy affirmatively covers, because silence about AI is no longer a sound basis for transferring the risk.
What "silent AI" means and why it existed in the first place
A claim involving an autonomous machine may go unpaid even though the company bought insurance for bodily injury, property damage, or technology errors. The company may learn that it lacks protection only when the insurer looks at how AI helped cause the loss. Until then, the policy’s familiar coverage categories can conceal an unresolved question about whether they apply to autonomous operation.
That uncertainty is what “silent AI” describes. Policies written before generative AI became part of everyday work did not say whether they covered or excluded AI-related losses, so policyholders assumed that the coverage they had would pay. Insurers had not deliberately priced and granted a distinct AI benefit. For autonomous hardware, the gap was also structural, because standard forms assumed human operation and did not describe autonomous systems. Silence left room to argue for coverage.
How ISO's January 2026 endorsements mechanically remove that coverage
The Insurance Services Office (ISO) made three generative-AI exclusion endorsements available for commercial general liability policies on January 1, 2026. Insurers can attach these standard forms at renewal without negotiating bespoke wording. This is a change in the contract. A company can renew the same type of insurance but lose protection for losses involving technology it uses in its work.
Cloud Security Alliance’s August 2026 research note describes the three forms, CG 40 47, CG 40 48, and CG 35 08 [1][2][3]. CG 40 47 excludes bodily injury, property damage, and personal and advertising injury arising out of generative AI under the CGL Coverage Part. CG 40 48 applies the exclusion only to personal and advertising injury under Coverage B, while CG 35 08 applies it to the products and completed-operations hazard. Their scope is not interchangeable. Recognizing the form number therefore starts the analysis, but understanding which part of the policy it modifies determines where protection has been removed.
The operative phrase is “arising out of,” not “caused by.” In insurance law, that phrase is read broadly, so an AI system need not be the proximate cause of the loss for the exclusion to apply. Its place in the causal chain can suffice. For autonomous hardware using a generative component in navigation, perception, or decision-making, this reaches beyond claims that identify the AI system as the immediate source of harm. The exclusion’s name may sound narrower than its effect.
But all three forms apply only to generative AI. Predictive models, pricing algorithms, and other non-generative AI remain in silence unless the carrier has written broader language. So a mixed robotics stack can have generative components that are excluded and other components whose coverage is still unclear. Joe Lam, a director at Verisk, ISO’s parent, has described the endorsements as “very essential in the marketplace,” and carrier adoption filings have grown through 2026. That expanding CGL response is only the beginning of the change.
Why the CGL exclusions are only the visible edge of a broader retreat
Most enterprise AI exposure was never concentrated in CGL. Financial harm can instead engage directors and officers liability, errors and omissions, cyber, and fiduciary liability, making the spread of restrictions into these lines especially consequential. Berkley has added an absolute AI exclusion to its D&O, E&O, and fiduciary liability products. Those restrictions also affect the policies companies count on for claims about how they do their work, run the business, and oversee it, beyond claims for physical injury and property damage.
Cyber policies are taking a different route. Beazley and QBE have introduced AI sublimits near ten percent of overall cyber policy limits. Coverage can remain while meaningful capacity shrinks. A policy may expressly mention AI protection yet cap an AI-driven loss at a fraction of the limit the buyer associates with the policy. With those caps, LLMjacking, in which an attacker hijacks AI systems, could use up all the coverage for the loss long before reaching a multi-million-dollar cyber limit.
Mayflower Specialty’s October 2026 guide notes that at least one insurer introduced a broad AI exclusion in 2025 that could reach AI-washing securities suits and oversight claims. Those are potential consequences of a high-profile autonomous system incident, distinct from the underlying injury claim. Many older D&O forms remain silent on AI. Whether renewal preserves that wording or replaces it with an exclusion is therefore a substantive coverage question.
The liability backdrop is already concrete. The Authors Guild reports that Bartz v. Anthropic received approval on July 20, 2026, for a $1.5 billion copyright settlement over Anthropic’s buying and storing pirated books for AI training without permission. Mobley v. Workday concerns potential vendor liability for employment decisions influenced by AI screening tools, with the case still in discovery and no final liability finding. Neither AI-mediated conduct nor a disputed allocation of responsibility makes determining whether insurance covers the resulting claims a theoretical exercise.
For an autonomous system incident, the resulting fragmentation can span several policies at once. CGL may exclude the loss, cyber may set a sublimit for it, D&O may have new exclusion wording, and coverage under Tech E&O may depend on the carrier’s own terms. Each line can leave the claim to another. Autonomous hardware makes that fragmentation especially difficult because a single failure can combine financial harm with physical injury, damaged equipment, and interrupted production.
Why autonomous and AI-driven hardware faces a harder version of this problem than software-only AI
An autonomous machine operating in the field creates simultaneous exposure to bodily injury, property damage, product liability, cyber loss, and production loss. Standard policy architecture separates those risks into coverage lines. The machine does not fail within those boundaries. Its software, sensors, physical components, and operating environment can contribute to the same incident, while each resulting loss encounters a different insuring agreement or exclusion.
The attribution of liability makes the division harder. Current liability and equipment programs are built around driver-liability regimes meant for human operation, while autonomy shifts the risk toward software failures, sensor errors, and system defects. Those categories sit between product liability and Tech E&O. A defective third-party sensor that triggers a robot failure can put the claim on the company that integrated it. The integrator need not have written defective code to face the claim.
The type of loss then determines the next coverage obstacle. Bodily injury and property damage may encounter CG 40 47 where generative AI forms part of the causal chain. Standard CGL coverage typically pays nothing for production losses when a robot shuts down an operation without physical damage. An AI sublimit may cap coverage when a cyberattack takes over the robot’s controls, while Tech E&O or inland marine may cover a software update or sensor failure that takes it offline, depending on what the policy says. Removing an exclusion from one policy would not resolve all those boundaries.
Underwriting must consequently account for the company’s use of AI, the authority the system has, the controls that constrain that authority, the approvals that precede execution, and the external systems or assets it may modify. These distinctions define the risk being transferred. Yet standard applications have no fields for supervision models, operating envelopes, geofencing, emergency stops, or safety cases, leaving the actual operation of the machine poorly described.
What a renewal audit for AI-deploying hardware companies must actually check
A company deploying autonomous hardware should treat renewal as a live audit, starting with the endorsement schedule to supplement the declarations page. Check for CG 40 47, CG 40 48, and CG 35 08 on every CGL policy, then identify the coverage each attached form removes. Absence is not an affirmative grant of coverage. A policy without those forms may still contain broader carrier wording or leave autonomous operation unresolved.
The same audit should examine AI-specific language in D&O, E&O, and fiduciary liability, where exclusions are arriving one insurer at a time. Compare the expiring wording with the renewal wording rather than assuming that the policy name establishes continuity. For cyber, check the declarations for AI sublimits as well as the exclusions. The amount of the stated limit available for the AI-related loss matters even when the policy mentions AI coverage.
For the hardware itself, confirm that the insuring agreements respond when no human is at the controls. Check whether bodily injury and property damage coverage applies when the machine runs that way and whether equipment or inland marine coverage follows it off-premises to the sites where it works. Production loss needs a separate answer. Check the policy’s scope for a software or sensor failure that stops an operation, including any requirement for physical damage.
Customer obligations must then be compared with the protection that remains. Enterprise contracts often require minimum Tech E&O and cyber limits, and government contracts can require more. When an AI exclusion removes protection that a contract requires, the company may have a certificate but lack the coverage it needs to meet the contract’s terms.
The GSA’s draft clause GSAR 552.239-7001 illustrates why government obligations require their own review. The draft includes incident-reporting obligations and makes prime contractors responsible for downstream Service Provider compliance. It is still a draft, separate from any contract term that already applies, but it shows why companies need to check contract terms as well as insurance. For active or planned field pilots, that comparison also needs to happen early. Autonomy placements take longer because fewer carriers have appetite, so waiting until a pilot is scheduled can leave the company unable to produce a certificate in time.
What underwriters now require to write affirmative AI coverage — and how to build a submission that gets it
Affirmative AI coverage depends on evidence that an underwriter can evaluate, rather than a general assurance that the company manages its technology responsibly. The CSA note identifies an evidentiary bar that includes documented human override capability, a human-in-the-loop inventory of consequential AI decisions, data-provenance documentation, a named accountable executive, and deepfake-resistant authentication. Governance must be visible in the submission. A company unable to document those controls can be denied coverage that a company with a usable record can access.
For autonomous hardware, the submission must also provide architectural proofs that deterministic safeguards are in place against runaway execution cycles. Without them, the likely result is denial or a prohibitive premium. The underwriting task is to distinguish a bounded system from one whose execution cannot be adequately constrained, and the submission must provide evidence for that distinction.
To make that evidence usable, a company should build the submission around how its machines operate in the field. It should describe the authority delegated to AI, what the system cannot do, and what happens when it fails, supported by the supervision model, operating envelope, geofencing, emergency stops, and safety case. A generic commercial application has no way to describe those features well enough. Supplying them gives the underwriter a basis for pricing affirmative protection rather than treating the machine like rented construction equipment and compensating for uncertainty through restrictive terms.
Market access is the other requirement. Fewer carriers have appetite for autonomous hardware, and those willing to consider it are not reached through standard broker channels. A specialist broker who regularly places autonomy risks can identify current appetite and avoid carriers that have quietly exited. Sending the submission to the wrong carrier consumes time and reduces leverage. Broker commissions are already built into the premium, so the choice between a specialist and a generalist concerns submission quality and access.
The renewal objective is therefore a documented match between the machine’s delegated authority, its safeguards, the losses it can produce, and the coverage the insurer expressly agrees to provide. That is the basis for affirmative risk transfer.
