Your Company May Have to Pay for Your AI Agent’s Next Mistake on Its Own
October 20, 2026
Insurers are adding AI exclusions to corporate liability policies, from ISO's generative AI exclusion forms to broader carrier-written endorsements in D&O and E&O coverage, often ahead of January 1 renewals. Because AI is deployed by technology teams while insurance is bought by finance and risk teams, the person signing the renewal may not know which agent activity the policy still covers. The edition proposes three questions to answer before signing, ending with whether the company can list every agent, its owner and its data sources.
Now here is an imaginary scenario. Everything about it is run-of-the-mill, even though a much smaller scenario of this has been carried out in real life and will be mentioned shortly.
A local lender has an AI agent which responds to its customers' emails; for each email that it receives, the agent searches its database of the lender's official publications and writes a reply from the publication which most accurately answers the question; it can do nothing else, other than respond to questions and cannot make transactions or alter the lender's records.
This agent was constructed with care and passed a rigorous security review. It has all the default safety rules switched on, rules which prevent it from divulging confidential information, making unkind remarks, or following some bizarre instruction hidden in an email. If you were to ask the lender, they would say it is well governed and hand you the relevant documentation.
Last September, one part of the lender's service area was hit by a hurricane, and the lender set up an actual hardship program to assist those affected: if a customer in one of the counties in question could not afford the payment, they could miss one payment, with no late fee. Someone wrote a memo explaining the program and put it in the agent's library, so customers asking questions during the hurricane would be directed properly. It was the right decision, back then.
On December 31, the program ended. The hurricane file remained in the library. No one had been given the chore of deleting it, so it remained there. Even the "approved" label was still there, and no indication had been provided to the computer that the file had expired.
The program ended on December 31. The document never left the library.
It is mid-February when a few hundred emails are received from customers. None of them are in any of the affected hurricane counties. They all claim to be struggling financially this month and in need of assistance. The agent's library is queried and the hurricane hardship document is retrieved. When a customer inquires about financial assistance, the agent finds that the document that describes a payment deferral is the closest match. Each customer is politely notified with confidence that they can skip their March payment with no late charges incurred.
Notice that nothing was hacked. The agent is just repeating a real document that the company has agreed to. All security settings were effective. None had the responsibility of deciding if that document was still in effect or whether it applied to the customer.
Every safety check passed. None of them asked whether the document was still true, or whether it applied to this customer.
They trust the email. It is coming from their lender. They miss March.
In April, the penalties and late payments have reached their accounts, and some of these late payments are now being relayed to the credit bureaus. Complaints flood in, and the lender, as any responsible lender would have done, apologizes and repays the penalties, requests that the credit bureaus amend the reports of the errors, and generally takes the blame. The regulator phones them. It wants to know what happened. Unfortunately, some people are more affected; some have suffered greater loss, e.g., one person whose mortgage has been rejected as a result of errors in their credit report, and they have employed legal representation amongst others.
The lender also files a claim because these types of suits are precisely the kind of thing against which the lender is protected by the professional liability insurance policy, which covers the company's liability for mistakes it may have made when serving its customers.
The insurance company responded by telling the lender that the policy had a new page inserted in January when it was renewed. In insurance, this is known as an endorsement, which is a written modification to the original policy. This one stated that the policy does not cover claims "arising out of" the use of artificial intelligence.
One page added at renewal, and the whole claim now turns on two words: “arising out of.”
However, the lender's legal team fought back. They pointed out that the mistake really had occurred due to an individual who had not deleted the original document. The insurance company retorted with the fact that the email causing the problem had been generated by the AI. In the end, the case turned on the meaning of two words: 'arising out of'. It will take months of litigation between the firm and its own insurance provider, with all of the legal bills going to the firm's account.
The chief technology officer approved the agent. The chief financial officer approved the insurance renewal. They had never seen each other's documents. Nobody, before signing, questioned if the policy still covered whatever it was the company had just handed to its customers.
The CTO approved the agent. The CFO approved the renewal. Neither saw the other’s paperwork.
And if that idea sounds too outlandish, this has already happened on (comparatively) a small scale. In early 2024, the British Columbia Civil Resolution Tribunal ruled against Air Canada, as one of its chatbots on the airline's website informed a man named Jake Moffatt, who was grieving at the time of the incident and trying to book an airline ticket, that he could claim their bereavement discount after purchasing his ticket, which in reality, Air Canada's policy did not include. Air Canada argued that it should not be held accountable for what the chatbot said, but the tribunal did not see it that way and held them accountable for all information on their website, which, as it turned out, included the chatbot. As such, they were ordered to pay the man about eight hundred Canadian dollars, and the principle (that if an AI says anything to a customer, it is the responsibility of the organization in question) is what scales.
The lender is invented; from this part on is genuine.
What insurers are really doing
It is common for major companies to update their business insurance annually, and one of the busiest renewal dates is January 1st. On 8 October 2026, Global Finance ran a story on how some insurers are excluding AI risk from corporate liability coverage, and risk managers are trying to determine the extent of this exclusion during January 1st renewal discussions.
There are two factors affecting this.
The first and simplest to describe is standard language. The Insurance Services Office (ISO), which drafts the standard policy language used throughout the U.S. property and casualty insurance market, unveiled a selection of generative AI exclusion forms in July 2025, according to Insurance Business (August 25, 2026). The most comprehensive excludes coverage for claims connected with generative AI, namely injuries, property damage, or advertising claims, and a claim needs only to arise out of generative AI; they are largely designed for general liability policies, the kind for 'normal' injuries and accidents rather than the more financial error of the lender's example. But their popularity is growing; a review by Trades Coverage, a licensed insurance marketplace, showed that by July 31, 2026, 4,078 state filing records, in 49 states and Washington, D.C., had adopted generative AI exclusions. Of these, 2,369 were active, and many of them sit in policy types small and mid-sized businesses most commonly buy. These include errors and omissions coverage.
The other way is for insurers to write their own, wider exclusions, which fall the hardest on policies that cover some of the most powerful individuals. Michel Léonard, chief economist, Insurance Information Institute, told Global Finance that over 60 groups of property and casualty insurers had filed to add some kind of AI exclusion, and the broadest ones tend to occur in four areas of coverage: directors and officers insurance (D&O, for when a firm's leadership faces personal lawsuits over corporate decisions); errors and omissions (E&O), which shields firms from claims of financial loss related to their work or advice; 'employment practices' insurance, for complaints like discrimination from staff; and 'fiduciary' coverage, which protects those like a firm's retirement plan administrator. Cyber insurance, at least, has been relatively stable to date.
This example should give the average board member pause. In 2025, Berkley Insurance introduced an exclusion that takes away all D&O coverage relating to the use, deployment, or development of AI. Should a shareholder of the company sue a director arguing that the company had not effectively managed its AI, the director's insurance would not cover him. In fact, Global Finance reports Berkley has used the same exclusion in its E&O and fiduciary coverage as well.
Under Berkley’s 2025 exclusion, a shareholder claim over poor AI oversight falls outside the director’s D&O coverage (Global Finance, October 8, 2026).
This is a non-uniform situation. In a July Willis webcast, the panelists stated that explicit exclusions of AI in the D&O market were still few, as Intelligent Insurer reported on July 31, 2026. It cannot be taken for granted that your policy will look like someone else's, so someone actually has to read yours.
Global Finance also reports that shadow AI (AI tools used by staff which IT or security teams were unaware of) is now being excluded by insurers.
Why most companies only realize it when it is too late
The reason most people just do not notice it was explained by Erike Young, deputy executive director of the California Intergovernmental Risk Authority, to Global Finance. Businesses that deploy AI at scale usually do it through their technology and operations teams rather than through their risk management and insurance purchasing professionals. The updated exclusions do not always spell out anything that clearly sounds like "AI." Some of them are tucked into broader cyber exclusions, while others are new exclusions in the form of endorsements that are buried in the middle of the renewal paperwork and never appear on the declarations page, the overview at the front of the policy that most executives have actually read. According to Young, a majority of executives will hear about it only if their broker, the external company that negotiates and handles their policies, brings it up, while board members are even less likely to hear about it.
Young summed up the entire issue in a single sentence: "The exposure existed before the exclusion; the exclusion simply made it visible."
What follows is the unpleasant bit. He has yet to hear of an insurer refusing to pay a claim because of an AI exclusion, explained Manny Padilla, president of the Risk and Insurance Management Society to Global Finance. According to Léonard, the true influence of AI exclusions will only be determined in due course once renewed policies are in place. Meanwhile, rules are being rewritten on a new basis, no one has yet put a claim to the test, companies will be committing to renewals before they know what the new terms will end up costing them, and the first claim that is going to be put up for the fight is not going to feel, to those involved, like just another part of an industry trend.
A replacement market is emerging. Aman Gour, co-founder of FurtherAI, a firm building AI tools used to provide insurance underwriting automation, explained to Global Finance that firms like Armilla, Mayflower Specialty and Embroker have now begun underwriting standalone AI liability coverage, offering limits from $2 million up to $50 million. The fact, however, remains that you still need to explain your AI to an insurer, bringing the conversation back to the lender's issue once again.
Before you put pen to paper on that renewal, here are three questions that should be asked:
The first, put simply, to your broker in writing, is this: are there any exclusions or limitations to claims which involve AI in our renewal, in any of our policies, including any language that may be incorporated into cyber exclusions? If so, how would the insurer interpret the meaning of "arising out of" where the cause is partly a human error and partly the performance of an AI system? Get the answer in writing, because that is the sort of thing that just vanishes off the phone.
Secondly, what does the insurer intend by the term AI? Ask whether it covers all the AI features the software suppliers have turned on in your system, not only those that your own development team may have implemented.
Finally, and most critically, can any internal source within your company provide a full list of all the agents you currently run, their owners, and their data sources, if an insurer or a court asked for it next week?
Most companies can answer the first two questions quickly. The third is usually still a blank page.
Most companies can get answers for the first question by Friday. But few of them can even answer the third.
If the company renews on January 1, then someone will sign the renewal in the next ten weeks. But they may not know how many AI agents are running or what they are doing, and the person running them probably does not know the insurance has changed.
If that is the case, the person who signed may not know which parts of the agents’ work the policy still covers and which parts it now excludes, and the person running the agents has never read the policy at all. Everyone assumes the company is covered. When an agent gets something wrong that the policy will not pay for, nobody has been given the job of dealing with it, and nobody could describe those agents well enough to buy the coverage that would.
The example of the lending situation is purely hypothetical (it is illustrative). All other information is obtained from the sources cited below.
