Whether They Intend to Cover: A Lloyd's Market Association Report on AI and Errors and Omissions Insurance, September 2025

On September 18, 2025, the Lloyd's Market Association announced a new report on how artificial intelligence can affect the international errors and omissions (E&O) insurance market. The association's members, its release says, include all of Lloyd's managing agencies. The report, The Impact of Artificial Intelligence on the International E&O Market, was written by its International Professional Indemnity Committee.

It is written for E&O underwriters, who, it says, should now make sure that policy coverage reflects the risks and opportunities before them. Its subject is generative AI, which it describes, "for the purposes of this article," as a deep learning algorithm that analyzes large amounts of data to produce content, such as text, images, audio and video, in response to what a user puts in. It calls itself an article, and its disclaimer says it "does not constitute advice of any kind." Parts of it follow, not the whole.

Uses, Risks and Rewards

The report's account of who uses generative AI runs across lawyers, accountants and architects. Accountants, it says, are using it for tasks that include bookkeeping automation, tax return preparation, and document review, the summarizing of contracts, invoices and receipts "to identify anomalies that merit further investigation."

"When a professional relies on GenAI in the provision of their advice or service," the report says, "they are still expected to adhere to the professional standards set out by their respective regulatory body." That means, in its words, that "the individual contracted to provide the service could ultimately be liable in the event that AI produces erroneous outputs which are later relied upon."

One of its examples of liability scenarios is hallucinations. Generative AI "can produce incorrect information, sometimes without caveating the source of the misinformation," and that "can lead to a professional services company relying on inaccurate outputs." The report recounts a UK case involving a barrister who "misquoted legislation as a result of reliance on AI," and says the judge "stated that using cases hallucinated by AI without verifying them amounts to negligence." "This exemplifies how reliance on AI by lawyers could lead to an E&O claim," it says.

Confidential data is the other example. Putting private or confidential documents into a public AI model "could potentially breach duties of confidentiality or have data protection ramifications," and there "could be contractual or tortious liability if confidential information provided to a public AI model is later regurgitated."

The report also lists rewards. Depending on the task, it says, an AI model allows "analysis of a whole data set, rather than a sample of data," and "It may also be the case that AI might reduce the numbers of E&Os, as unsupervised junior staff may be less accurate than a professional relying on GenAI."

Whether to Cover

"Wordings written on a liability basis do not distinguish between claims involving GenAI and those that do not." So underwriters, it says, "should consider whether they intend to cover losses resulting from reliance on inaccurate GenAI outputs by professional services firms." If such risks are to be excluded, "careful attention should be paid to defining GenAI and the wording of an appropriately specific exclusion."

"For example, the underwriter may be prepared to cover human failure in not detecting the error, but not an issue with the insured's systems or software, which would generally fall to a cyber policy." Whether a cyber exclusion catches generative AI, it adds, "depends on the wording of the exclusion."

In the example, which policy may respond turns on whether the loss came from a person's failure to detect the error or from the firm's systems. Telling the two apart could depend on whether anyone can show who checked the output.

The report also names "untested coverage issues." One is whether generative AI "would be considered a professional service or a product." If a firm has built its own system, which then gives erroneous outputs that are relied on, the report asks whether that is "more appropriately insured under a product liability policy." A service, or a product. It adds that where an AI tool "has repeatedly caused an error," deciding whether an aggregation or series clause should apply "could potentially be a difficult endeavour."

Example Questions

Many coverage matters, it says, will also rely on a fair presentation of the risk, and on the questions underwriters asked about the use of generative AI within the firm. The report offers example questions, and says the list "is illustrative only, is not complete and, if used, will need to be adjusted to work for specific use cases."

The questions ask what the AI is used for in the company, and whether it has set out "a clear set of circumstances where they consider AI to be of use and where it is not." They ask how the tools were selected, and how far their terms and conditions were reviewed "for exposures such as ownership of the data." They ask whether an acceptable use policy has been implemented, distributed to all employees and regularly reviewed or updated as new products come to market. They ask what data goes into the models, and how the firm uses proprietary or confidential information in them.

"What training does the company provide to its staff on the grounding and framing of AI and what procedures are in place to assess the outputs for accuracy and acceptability?" Is the use of any tool "being disclosed correctly to third parties," for example where advice "has been created in part or wholly with AI?" How is the firm billing clients for work involving AI, and has that changed from what the report calls traditional advice methods? And "Has the client agreed or signed a contract confirming the use of AI and does this include limitations of liability?"

Several of these questions turn on things a firm would keep in writing: a use policy, its training, its procedures for assessing outputs, a contract with the client. How many could a firm answer from its files?