Insurers Are Writing AI Out of Liability Policies

Plenty of AI programmes still treat insurance as a matter for someone else's renewal meeting. Product ships the assistant, security argues about permissions, finance funds the pilot and legal reviews the vendor terms. Then an output goes wrong, and a customer decision is steered by an agent that sounded confident and was wrong. The failure modes are familiar by now: a false claim in a sales deck, a privacy spill through a chatbot, a defamatory summary, a hallucinated contract clause.

The board's first instinct is often the old industrial one: existing liability, cyber, and errors-and-omissions cover will handle it. That instinct is ageing badly. In the 2026 renewal cycle, carriers and rating organisations have made generative AI an explicit exclusion on commercial general liability (CGL) forms, and other lines are following with absolute or expansive AI carve-outs. A few lines of technical wording in a policy endorsement can leave a large hole in the profit-and-loss statement.

Once a liability form excludes generative AI, a bad output becomes a loss the company pays itself, unless someone has already agreed to carry it.

Insurers are rewriting the liability forms

When a loss class becomes visible and hard to price, insurance markets reclassify it. Verisk's Insurance Services Office (ISO) commercial general liability suite now includes optional generative AI exclusion endorsements, among them the broad CG 40 47 form, which can remove bodily injury, property damage, and personal and advertising injury arising out of generative AI. Narrower variants target advertising injury alone, or products-completed operations. Wording and adoption vary by carrier and state, but the direction is clear enough for anyone deciding what to ship.

Industry reporting through 2025 and 2026 says the same in plainer language: carriers are excluding claims linked to generative outputs, misleading AI content, and regulatory failures tied to AI implementation. Some cyber and professional lines remain relatively flexible compared with traditional liability forms. Others are tightening, through security riders, exclusions for unauthorized tools, or absolute AI exclusions across directors and officers, errors and omissions, and fiduciary programmes. The market will keep changing, but the decision about which AI work may touch customers, money, identity, or public speech has to be made now.

Earlier memos covered related ground: Agent Security Is Becoming a Shared-Responsibility Model on responsibility shared with model providers, Who Is Accountable When AI Acts on accountability inside the firm, and Evidence an AI Project Must Produce on what buyers should demand in procurement. Coverage-gap analysis asks something blunter: who holds the uninsured tail when automated work fails in public?

Six lines to fill in

The table below is meant to force one honest conversation across product, security, finance, legal, and the board. Whoever fills it in needs no underwriting expertise, but none of the six lines can be left out.

Where operators get it wrong

The commonest error is to treat all AI as one peril. A research assistant summarizing public web pages is a different exposure from an agent that emails customers, edits customer relationship management fields, or drafts regulatory filings, and sorting workflows by model brand tells you almost nothing about which is which. Coverage has to follow what the system actually does.

The next is assuming cyber cover will pick up content and professional-services losses. A prompt-injection breach that exfiltrates data may look like a cyber claim. A fluent falsehood that harms a customer may look like advertising injury, professional negligence, or media liability. If the commercial general liability form excludes generative AI and the errors and omissions form was written for human consultants, the claim can fall between the two.

Some teams wait for a purpose-built AI policy to solve the problem. Specialist AI insurance will keep developing; some startups will fill gaps, and large carriers will reprice and reword. None of that arrives in time to tell management which AI work is allowed while the market sorts itself out.

Questions a board should ask

"Do we use AI?" gets a board very little; the useful questions are operational. Which customer-facing workflows could produce a third-party claim this quarter? Which policy line is supposed to respond to each, and did generative AI exclusions attach at the last renewal? Who owns the residual risk in the uninsured middle, if anyone does? Is there evidence an underwriter could review tomorrow morning? Until that evidence exists, the workflow should stay at draft level.

The National Institute of Standards and Technology (NIST) AI Risk Management Framework helps here as a reminder that measurement, monitoring, and governance are what make a risk transfer credible. Use it as an operating reference and keep it separate from insurance advice. Underwriters and counterparties increasingly ask for what security teams already track: system inventory, actor permissions, review records, logs, and revoke paths.

How this connects to earlier memos

Earlier memos argued for rules before tools, permissions before autonomy, clear accountability before executive announcements, and browser control points before credentialed agents roam free. The coverage gap is the financial counterpart of those operating controls: however clever the agent workflow, a company stays exposed if residual loss has no policy home and no named owner.

Carriers and rating organisations will supply the exclusion wording. Management has to supply the name of whoever carries the residual loss.

What to do this week

Build a one-page AI coverage gap table for the five workflows with the largest blast radius and fill in all six lines. Where the residual owner or the exclusion surface is blank, demote the workflow from autonomous action to assisted draft until finance, legal, and the business owner can answer in the same meeting.

Do not park this until the next broker renewal. A cheap software seat does not make a consumer chatbot trial free: the expensive part may be an uninsured claim and a customer make-good, followed by the conversation in which someone tells the board "we thought cyber covered it." A long pilot list is no evidence of readiness. A team is ready when it can point to each AI action and say whether a policy covers it or the company retains it, and keep the rest switched off until a residual owner and a stop rule exist.