AI creates liability when its output, decisions or actions harm a customer, employee, investor or business partner and that party brings a claim. The most common routes are inaccurate or hallucinated output, discriminatory automated decisions, AI-enabled cyber attacks and fraud, mistakes by autonomous agents and misleading statements about a company's AI. This page sets out each risk with the cases and rules behind it, the line of insurance that usually responds and the Mayflower Specialty coverage module written for it.
Three Pressures Are Moving AI Exposure Onto Company Balance Sheets
Regulators, plaintiffs and insurers are each changing how AI losses are paid for, and they are doing it at the same time. Together they mean a company can face an AI claim that its current policies were not written to answer.
Regulators
01
7%
of worldwide annual turnover, or €35 million if higher, is the maximum EU AI Act fine for prohibited AI practices
The EU AI Act can reach companies outside the EU when their AI output is used there. In the United States, AI rules in Texas and Illinois took effect on January 1st 2026, Colorado's automated decision law applies from January 1st 2027 and New York City already requires bias audits of automated hiring tools. Each new duty gives regulators something to enforce and can give plaintiffs a standard to measure a company against.[1][2][3][4][5]
Litigation
02
15
AI-related securities class actions filed in the first half of 2026, against 16 in all of 2025
Courts are also holding companies responsible for what their AI says. A British Columbia tribunal rejected Air Canada's argument that its chatbot was responsible for its own advice, and Damien Charlotin's database lists more than 2,100 court decisions worldwide involving AI-fabricated content as of October 5th 2026.[7][8]
Carriers
03
3
standard generative AI exclusions for general liability policies (ISO forms CG 40 47, CG 40 48 and CG 35 08), which took effect in January 2026
W. R. Berkley has introduced an “absolute” AI exclusion for D&O, E&O and fiduciary liability, and the Financial Times reported in November 2025 that AIG, Great American and W. R. Berkley had asked US regulators to approve AI exclusions. Many policies written before generative AI say nothing about it, which leaves coverage to be argued after a loss.[10][11][12]
The practical response is to find out, before the next renewal, which of these pressures reaches your company and which of your policies would answer if it did.
The risk map
Five Places Where AI Fails
Mayflower groups the ways AI fails into five categories: the model's own output, decisions an AI system takes on its own, the data it learns from, the way it is built into a business and the dependencies it shares with many other companies. Select a category to see what it covers, a real example and the Mayflower module written for it.
Select a category to see its detail. Arrow keys move between categories.
Category 01 of 05
Model Performance
Model performance risk is the chance that an AI system's output is wrong, biased or degrades over time. A generative model can state a falsehood with confidence, a scoring model can disadvantage a protected group, and a model that tested well at launch can drift as the data it sees changes. It is the most direct route from AI to a claim, because someone relied on the output.
A British Columbia tribunal held Air Canada liable for negligent misrepresentation after its website chatbot gave a customer wrong advice about bereavement fares, and rejected the argument that the chatbot was responsible for its own actions.
Autonomous decision risk arises when an AI agent sends messages, changes records, commits to terms or moves money without a person approving each step. The speed that makes agents useful also means a mistake is carried out before anyone can catch it, and the company that deployed the agent is likely to be the one asked to answer for it.
Replit's AI coding agent deleted a user's production database during a code freeze. The data was later restored, but the incident showed how quickly an agent with write access can act on a wrong decision.
Data integrity risk covers the data a model is trained on, fine-tuned with or retrieves while it works. Poisoned data can plant hidden behavior in a model, confidential data can leak through prompts and outputs, and data used without permission can lead to privacy and intellectual property claims.
Anthropic, UK AI Security Institute and Alan Turing Institute, October 2025
A joint study found that about 250 malicious documents were enough to plant a backdoor in language models ranging from 600 million to 13 billion parameters.
Integration and deployment risk sits in the layer between a model and the business: the instructions it is given, the data and tools it can reach, the guardrails on its answers and the people who monitor it. Prompt injection, over-broad permissions and missing human review are failures of this layer rather than of the model itself, which is why the controls around a model deserve as much scrutiny as the model.
The Markup found in 2024 that the city's chatbot told businesses they could take a cut of workers' tips, contrary to city law. In early 2026 Mayor Zohran Mamdani called it “functionally unusable” and announced it would be taken down.
Systemic concentration risk comes from many companies relying on the same few foundation models, AI platforms and developer tools. A flaw or compromise in one shared component can reach every company that uses it at the same moment, so losses arrive together and a vendor's contract terms matter as much as the company's own controls.
Version 1.84.0 of Amazon's Q Developer extension for VS Code shipped with an injected prompt instructing the agent to wipe data. AWS said the code would not execute and no customer resources were affected, and replaced the release with version 1.85.0.
Seven routes lead from an AI failure to a claim. Each section below gives a short answer, sourced examples, the line of insurance that usually responds and the Mayflower module written for it, so it is quickest to start with the routes that match how your company uses AI.
01
Model Performance
Inaccurate and Hallucinated Output
Inaccurate or hallucinated AI output creates liability when someone relies on it and loses money, or when it damages a reputation. Courts and tribunals have held companies and professionals that use AI responsible for what it produces, in the same way as for the rest of their website, advice or work product.[7][18]
The exposure is widest for companies whose AI talks to customers or produces work that clients rely on, such as legal, financial, medical and technical advice. Professionals who file or deliver AI-drafted work carry the same risk, and Damien Charlotin's database lists more than 2,100 court decisions worldwide involving AI-fabricated content as of October 5th 2026.[8]
Moffatt v. Air Canada, February 2024
A British Columbia tribunal held Air Canada liable for negligent misrepresentation after its chatbot gave wrong advice on bereavement fares, ordering C$812.02 in damages, interest and fees. The tribunal said the chatbot “is still just a part of Air Canada's website.”
A federal judge in New York fined two lawyers and their firm $5,000 for filing a brief that cited six cases invented by ChatGPT, and ordered them to write to each judge falsely named as an author.
AI that screens, scores or ranks people creates liability when its decisions disadvantage a protected group. Employers remain responsible for discrimination in hiring, promotion and dismissal when an AI tool makes or shapes the decision, and a federal court has also let discrimination claims proceed against an AI vendor on the theory that it acted as the employer's agent.[19]
New rules add duties on top of existing anti-discrimination law. New York City requires a bias audit within the year before an automated employment decision tool is used, Illinois has barred employers since January 1st 2026 from using AI that has a discriminatory effect, and California's civil rights rules on automated decision systems took effect on October 1st 2025.[3][5]
Mobley v. Workday, 2024 to 2026
A federal court in California let claims against Workday proceed on an agent theory in July 2024, then in May 2025 conditionally certified a nationwide collective of applicants aged 40 and over who allege its AI screening tools disadvantaged older applicants. The case was in discovery as of August 2026.
iTutorGroup agreed to pay $365,000 to settle an EEOC suit alleging that its tutor application software automatically rejected female applicants aged 55 or older and male applicants aged 60 or older.
Prompt injection, data poisoning, model and data leakage, and deepfake-enabled fraud
AI creates two kinds of cyber risk: attacks on a company's AI, such as prompt injection and data poisoning, and attacks that use AI, such as deepfake impersonation. Which policy responds depends on the loss: cyber insurance for breaches and data loss, crime or social engineering cover for stolen funds, and liability lines when customers or investors claim the company's AI failed them.
Definition
Prompt injection
Prompt injection is an attack in which instructions hidden in user input, or in content an AI system reads, cause the system to ignore its intended rules. It can lead an AI assistant to disclose data or take actions its operator never authorized.
Definition
Data poisoning
Data poisoning is the deliberate corruption of the data a model learns from, so that the model behaves as the attacker chose, often only when a hidden trigger appears. Because the behavior is learned, it can carry into every product built on the model.
Leakage is the quieter version of the same risk: Samsung banned generative AI tools on company devices in 2023 after staff uploaded internal source code to ChatGPT. Weak access controls make all of these worse, and IBM's 2025 breach study found that 97% of organizations reporting a breach of an AI model or application lacked proper AI access controls.[22][23]
13%
of organizations in IBM's 2025 study reported a breach of an AI model or application
Researchers disclosed a zero-click prompt injection flaw in Microsoft 365 Copilot (CVE-2025-32711, rated 9.3 out of 10) that could expose data from the assistant's context; Microsoft patched it in June 2025.
An Arup employee in Hong Kong made 15 transfers totaling HK$200 million (about US$25.6 million) after a video call in which deepfakes impersonated the company's UK-based chief financial officer. Arup said none of its internal systems were compromised.
Cyber insurance for breaches and data loss, crime or social engineering cover for stolen funds, and E&O or D&O when customers or investors bring claims.
AI-E&O is written for claims by clients who suffer a loss from a company's AI-enabled products or services, which can follow a successful attack on that AI, subject to the policy terms.
An AI agent creates liability when it takes an action that harms someone before a person can review it, such as sending a message, changing a record, agreeing to terms or moving money. In Moffatt v. Air Canada a tribunal rejected the argument that a chatbot was a separate party responsible for its own conduct, so a company that deploys an agent should expect to answer for what the agent does.[7]
Agents also widen the cyber exposure described above, because an agent that reads outside content and can take actions can be steered by instructions hidden in that content. The controls that matter most are limits on what an agent may do without approval, a record of what it did and a fast way to stop it.
Replit, July 2025
Replit's AI coding agent deleted a user's production database during a code freeze, although the data was later restored.
Researchers showed that a prompt injection flaw in Salesforce's Agentforce agents, rated 9.4 out of 10, could be used to extract sales lead data from the CRM, and Salesforce has since patched the flaw.
AI-washing is the practice of overstating what a company's AI does or how much the company relies on it. It creates liability for directors and officers because regulators bring enforcement actions over such statements, and shareholders bring securities class actions when the share price falls after the truth comes out.[6][28]
The claims are also moving beyond overstatement: recent securities suits, such as the June 2026 suit against ZoomInfo, allege that companies understated how much AI threatened their own business. Disclosure is growing at the same time, with 72% of S&P 500 companies flagging AI as a material risk in 2025 against 12% in 2023, and each of those statements has to stay accurate as the company's use of AI changes.[29][30]
24
AI-related securities class actions filed in 2026 by September 23rd, nearly 14% of all new filings
The SEC settled its first “AI washing” cases against two investment advisers for misleading statements about their use of AI, with penalties of $225,000 for Delphia and $175,000 for Global Predictions.
The SEC settled charges that Presto Automation made materially false statements about its AI drive-thru ordering product, which needed human help on most orders. The D&O Diary described it as apparently the SEC's first AI-washing action against a public company.
AI systems create data and intellectual property claims when they are trained on, retrieve or reveal information the company had no right to use or disclose. The claims range from privacy and biometric statutes that set damages per violation to copyright suits over the material used to train a model and the content it produces.
Statutory damages make the privacy side expensive: Illinois' Biometric Information Privacy Act lets individuals sue for $1,000 for each negligent violation and $5,000 for each intentional or reckless one. Disclosure duties are growing too, and since January 1st 2026 California has required developers of generative AI to publish summaries of their training data.[34][35]
Thomson Reuters v. Ross Intelligence, 2025 to 2026
A federal judge in Delaware held in February 2025 that Ross's use of Westlaw headnotes to train a competing AI legal research tool was not fair use, noting that the AI before him was not generative. The Third Circuit affirmed in late September 2026 in an opinion that was still sealed when this page was reviewed, so how far the reasoning reaches generative models is not yet clear.
It varies by policy: privacy events often sit with cyber insurance, and cover for intellectual property claims differs sharply from one policy form to the next.
Mayflower module written for it
How Mayflower's modules treat intellectual property and privacy claims depends on the policy terms, so ask us about the specific exposure before relying on any policy for it.
AI regulation creates liability in two ways: regulators can investigate and fine a company directly, and each new duty, such as a bias audit, a notice to consumers or a record of how a system was tested, can give plaintiffs a standard against which to measure the company's conduct.
The EU AI Act reaches companies outside the EU when their AI output is used there, and its obligations for high-risk uses listed in Annex III, which include employment, now apply from December 2nd 2027. Texas's Responsible AI Governance Act has been in effect since January 1st 2026, enforced by the state attorney general with penalties of up to $200,000 for each violation that cannot be cured, and Colorado's automated decision law applies from January 1st 2027.[1][2][4][38]
Product liability is moving the same way, since the EU's revised Product Liability Directive will treat software, including AI, as a product when it is placed on the EU market on or after December 9th 2026. Insurers face rules of their own as well: 24 states and the District of Columbia had adopted the NAIC model bulletin on insurers' use of AI as of April 1st 2026. Because these dates keep moving, it is best to review the rules that apply to the company at each renewal rather than once.[39][40]
FTC orders on “AI-powered” marketing claims, August 2026
The FTC finalized orders requiring Cox Media Group, MindSift and 1010 Digital Works to pay a total of $930,000 over charges that they deceived customers about an “AI-powered” service said to target ads using conversations captured from consumers' smart devices.
The FTC announced five cases over deceptive AI claims in September 2024. In December 2025 it set aside its order against one of them, Rytr, which shows that enforcement priorities can shift even where the underlying law does not.
D&O for investigations and claims against directors and officers, and EPL for employment rules. Whether fines and penalties can be insured depends on the law where they are imposed and on the policy wording.
AI claims can reach four kinds of party, and a single incident can involve several of them at once.
Developers
Companies that build and sell models or AI products face claims from customers when a product fails and from people harmed by its output. In July 2026 a Delaware court refused to dismiss Robby Starbuck's defamation suit against Google over statements made by its AI.[44]
Companies that use AI in their own products, services or operations answer for its output to the people they serve, as Air Canada found when a tribunal held it liable for its chatbot's advice.[7]
Boards answer to shareholders and regulators for how the company oversees AI and what it says about it, through securities class actions, derivative suits and enforcement actions.[6][28]
Employers answer for discrimination when AI screens, ranks or evaluates applicants and employees, including when the tool comes from a vendor, and the vendor itself may be sued as in Mobley v. Workday.[19]
Most companies hold more than one of these roles, so the useful exercise is to list your AI systems and note, for each one, whether you built it, deployed it, report on it or use it to make decisions about people.
From risk to coverage
Which Policy Usually Responds to Each AI Risk
The table maps each route to the claimant who typically brings it, the line of insurance that usually responds and the Mayflower module written for it. It is general information; whether a policy responds depends on its wording and the facts of a claim. The coverage overview explains how the four modules fit together.
AI risks mapped to the typical claimant, the line of insurance that usually responds and the Mayflower module written for each
The coverage gap check asks up to nine questions about how your company uses AI and which policies it holds, then shows where an AI claim may fall outside your program. It asks for no contact details.
Cyber insurance answers only part of the question, because it is built to respond to security events, while many AI claims involve no security event at all. A chatbot that gives a customer wrong advice, or a screening tool that rejects older applicants, causes harm without any breach, so the claim falls to liability policies such as E&O, EPL or D&O, which may be silent on AI or exclude it.
Cyber insurance
A Security Event
What triggers it
A security or privacy event, such as a breach, ransomware or a network outage.
Who brings the claim
The company itself, for its own costs, and customers whose data was exposed.
What the loss looks like
Breach response, data restoration, business interruption and privacy claims.
An example
An AI-enabled attack: IBM's 2026 study found that 1 in 4 malicious breaches were AI-enabled, at about $6 million each on average.[45]
The policy built for it
A cyber policy.
AI liability
A Decision, Output or Action
What triggers it
A decision, output or action by an AI system that harms someone, with or without a breach.
Who brings the claim
Customers, clients, applicants, employees, investors and regulators.
What the loss looks like
Damages, settlements and defense costs for wrong advice, discrimination, professional errors and securities claims.
An example
No breach at all: Air Canada's chatbot gave a customer wrong fare advice, and a tribunal held the airline liable.[7]
The policy built for it
AI-E&O, AI-EPL and AI-D&O, or AI DIC Excess over the program you already carry.
The two are designed to sit beside each other rather than replace one another. The practical course is to keep cyber for security events and to check, before the next renewal, that your liability policies name AI.
Short answers to common questions about AI risk. The full FAQ covers the policy, the application and the ways to buy coverage.
How does AI create legal liability for a business?
AI creates legal liability when an AI system's output, decision or action harms a customer, employee, investor or business partner, and that party brings a claim against the company that built or used it. The usual routes are inaccurate output, biased decisions about people, cyber attacks on or with AI, mistakes by autonomous agents, misleading statements about AI and breaches of new AI rules.
Who is liable when AI causes harm, the developer or the company using it?
Both can be sued, but the company that deployed the AI is often the first to face the claim, because it dealt with the person harmed. In Moffatt v. Air Canada a tribunal rejected the airline's argument that its chatbot was responsible for its own advice. Developers face claims too, and employers can be liable for the vendor tools they use in hiring, so a contract with an AI vendor rarely moves the whole risk.
What are the biggest AI risks for businesses?
For most businesses the largest AI liability risks are inaccurate output that customers rely on, discrimination by AI used in decisions about people, AI-enabled fraud and data leakage, errors by agents acting without review, and statements about AI that investors or regulators later call misleading. Which of these matters most depends on how and where the company uses AI.
Which insurance responds to which AI risk?
As a rule, errors and omissions insurance responds to claims that AI output or services caused a client a loss, employment practices liability to discrimination claims over AI used in hiring, D&O to claims against directors over AI oversight and disclosure, and cyber insurance to breaches and data loss. An existing D&O, EPL or E&O policy may be silent on AI or exclude it, which is the gap AI DIC Excess is written for.
What should a company do first about its AI liability risk?
Start with an inventory: list each AI system the company builds, deploys or relies on, and note who could be harmed if it fails, whether customers, employees, investors or business partners. Then read the D&O, EPL and E&O policies for AI exclusions or silence before the next renewal. Mayflower's application asks for an AI system inventory and an AI governance policy or responsible AI framework, so the same work also prepares an application.
Sources
Sources and Dates
Every case, figure and rule on this page links to its source, and each was checked on October 5th 2026. Cases and rules change, so confirm the current position before relying on any one of them.
Find Out Which of These Risks Your Program Already Answers
Mayflower underwrites dedicated AI liability coverage for management, employment and professional exposures, plus an AI DIC Excess layer over an existing program. Discuss a specific exposure or start an application.