AI policy exclusions. What SME operators must review before their next renewal.

The bifurcation of the insurance market

The insurance market's response to AI agent risk is not a single event. It is a structural shift happening at different speeds across different product lines and geographies. Understanding the shift requires distinguishing between two simultaneous developments: specialist carriers building AI-specific products, and mainstream carriers adding AI exclusions to general policies.

AIUC began writing the first AI-specific policies in the United States in 2025, with ElevenLabs as its first named insured under an AIUC-1-backed policy in February 2026. Munich Re aiSure has been writing AI performance risk since 2018 and extended to LLM coverage in 2019. Armilla, operating as a Lloyd's coverholder from Toronto, offers coverage up to USD 25 million per company for performance errors, hallucinations, data leakage, and regulatory violations. These are affirmative, AI-specific products.

Simultaneously, AIG, Great American Insurance, and WR Berkley filed endorsements with US regulators in late 2025 to add AI exclusions to their standard errors and omissions, directors and officers, and cyber policies. The effect of those endorsements, when they flow through to renewal terms, is to remove AI-related losses from policies that SMEs have relied on for general liability coverage. European carriers observed the US filings and are incorporating equivalent exclusions at renewal. The net result is that an SME that did not review its policy wording at the most recent renewal may now be carrying a coverage gap it does not know about.

The four exclusion types to review

Policy exclusions are not written in plain English. They appear in endorsement schedules, in definition sections that modify the scope of coverage, and in conditions that look like procedural requirements but function as coverage limitations. The four exclusion types that matter most for AI agent operators are described below, with guidance on where to find them.

1. Autonomous action exclusions

An autonomous action exclusion removes coverage for losses arising from decisions or actions taken by an automated system without human approval at the point of execution. The wording varies. It may appear as an exclusion for "loss arising from automated or algorithmic decision-making," "loss resulting from artificial intelligence operations," or "loss caused by any system operating without human authorisation at the time of the relevant act."

This is the exclusion that would most directly affect an SME running a customer-facing agent. In the Moffatt v. Air Canada case, the chatbot issued a refund commitment that no human had authorised. Under an autonomous action exclusion, that commitment would be an uninsured loss. The carrier would argue that the agent acted without human approval at the time of the relevant act, triggering the exclusion, and the insured would be left with a judgment to pay from its own funds.

Where to find it: in the definition of "covered act" or "professional services," and in the exclusion schedule of technology E&O and cyber policies. Look for language that limits coverage to "acts performed by natural persons" or that requires "human review and approval prior to execution."

2. AI-generated content exclusions

An AI-generated content exclusion removes coverage for intellectual property, defamation, and professional liability claims arising from content produced by an AI system. The practical scope of this exclusion is larger than it appears. Marketing copy generated by an AI agent, technical documentation drafted by an AI, legal summaries produced by an AI research assistant, and code written by an AI coding tool are all potentially caught.

In the Mata v. Avianca case, heard in the Southern District of New York in 2023, an attorney submitted a brief containing fabricated case citations produced by an AI research tool. The attorney faced sanctions. Under an AI-generated content exclusion, the professional indemnity exposure arising from that filing would not be covered. The insurer's argument would be that the loss arose from AI-generated content, and the exclusion removes it from the policy.

Where to find it: in media liability endorsements, in technology E&O policy definitions of "professional services," and in the content-related exclusions of general liability policies. Look for language that removes coverage for "content produced by automated means" or "outputs generated by artificial intelligence systems."

3. Scope creep exclusions

A scope creep exclusion removes coverage for losses that arise when an AI agent acts outside its documented authorised scope. This exclusion targets the operator rather than the technology. An agent that has been configured to answer customer service queries but begins making booking changes without authorisation has exceeded its scope. An agent that has been scoped for read-only database access but executes a write operation has exceeded its scope.

The practical challenge with scope creep exclusions is that they require the operator to have defined the agent's scope in writing. An operator who cannot produce a written scope definition cannot demonstrate that the agent acted outside it, which means the insurer has a basis for declining to cover any loss and requiring litigation to establish the facts.

Where to find it: in conditions precedent, in definitions of "authorised use," and in policy schedules that require the insured to maintain documented operational controls. The condition may not use the words "scope creep" but will achieve the same effect by requiring the insured to have maintained a documented scope definition as a condition of coverage.

4. Governance exclusions

A governance exclusion removes coverage where the operator cannot demonstrate that it maintained documented oversight, audit telemetry, or incident handling procedures at the time of the loss. This is the newest class of exclusion and the one most directly shaped by the EU AI Act and the AIUC-1 standard.

Under the EU AI Act's Article 26 obligations, deployers of high-risk AI must maintain oversight registers, logging schedules, and incident protocols. Insurers writing EU market coverage in 2026 have begun incorporating these obligations as policy conditions. An insured that cannot produce its Article 26(6) log retention schedule when a claim arises will find that the insurer has a basis for denying coverage under the governance exclusion.

Where to find it: in conditions precedent to coverage, in warranties within the policy schedule, and in audit rights clauses that permit the insurer to request governance documentation at any time during the policy period. An audit rights clause that the insured cannot satisfy at the time of a claim is the functional equivalent of a governance exclusion.

How to review your current policy

The review process begins with collecting the current policy documents in full, not just the schedule and the summary. The exclusion schedule, the endorsements, and the definitions section together determine the actual scope of coverage. For SMEs whose policies were last renewed before 2025, the definitions of "professional services," "covered act," and "automated systems" should be read carefully to determine whether they were written in a pre-AI context.

A useful test is to apply the policy wording to a hypothetical scenario. Suppose your customer-facing AI agent issued a refund of EUR 300 to a customer it was not authorised to refund. Run the policy language against each element of that loss: the agent's action, the amount, the proximate cause, the lack of human approval. Where the language is ambiguous, that ambiguity will be resolved by the insurer in a claim context, and experience with the Moffatt case suggests that ambiguity does not resolve in favour of the insured when the incident involves autonomous AI action.

The outcome of the review is a coverage gap analysis: a list of the scenarios that your current policy does not cover. The gap analysis is the document you bring to your broker or to a specialist carrier when you initiate the conversation about dedicated AI coverage.

What the EU AI Act and Product Liability Directive add

From 2 August 2026, the EU AI Act creates direct regulatory liability for deployers of high-risk AI systems who do not meet the Article 26 obligations. From 9 December 2026, Directive (EU) 2024/2853 on liability for defective products creates strict product liability for AI software that causes damage to users. Both statutory liability regimes create exposure that standard SME policies were not written to cover.

The AI Act's penalty regime reaches EUR 15 million or 3 per cent of worldwide annual turnover for deployer breaches. The Product Liability Directive's scope of compensable damage includes destruction or corruption of digital data and medically recognised psychological harm, categories that appear in neither standard cyber nor standard professional indemnity wordings.

These statutory liabilities are not theoretical. They are the legal basis on which a regulator will open an enforcement inquiry and on which an affected person will bring a civil claim. Standard SME policies are not designed for them, and the exclusions being added in 2025 and 2026 make the gap wider, not narrower.

For the full regulatory picture, the Why It Matters page covers the EU AI Act and Product Liability Directive in plain language. For the structured pathway from gap analysis to coverage, see the coverage pathway. For the European institutional context on what AI insurers are building, agentinsured.eu tracks active carriers and coverage frameworks.

Frequently asked questions

Do standard cyber policies cover AI agent incidents?

Most standard cyber policies written before 2024 contain wordings that either do not mention AI or have begun to be amended with explicit AI exclusions. The key test is whether the policy covers losses arising from automated decisions or autonomous actions taken without direct human authorisation. Most standard wordings do not. SMEs should request written confirmation from their insurer that AI agent activity is covered before the next renewal.

What are the most common AI exclusions being added to UK and European policies?

The most common exclusions being added in 2025 and 2026 are: losses arising from autonomous AI decision-making without human approval, losses from AI-generated content that infringes intellectual property, losses from AI recommendations in professional services without qualified human review, and losses from AI systems operating outside their documented scope. AIG, Great American Insurance, and WR Berkley all filed exclusion endorsements with US regulators in late 2025, and European carriers have followed.

If my existing policy excludes AI, what are my options?

Three options exist. First, negotiate an AI endorsement with the existing carrier, though many are declining to write AI cover as part of general policies. Second, approach a specialist AI carrier such as AIUC, Armilla, or a Lloyd's syndicate offering AI-specific policies. Third, restrict the AI agent to tasks that fall within the existing policy scope and accept the coverage gap for autonomous actions.

What should an SME look for when reviewing AI policy exclusions?

SMEs should look for four categories of exclusion language: autonomous action exclusions, AI-generated content exclusions, scope creep exclusions, and governance exclusions. Each can eliminate cover for the most likely AI agent incident scenarios.