Most business insurance was written before AI agents existed. Here is where your policies stop, what to ask your broker, and what makes cover possible.
Each policy you hold was written for a particular kind of loss. An agent that decides, promises, sends or pays on your behalf creates a different kind. This is where the two part company.
Written forA professional service that causes a client financial loss through a negligent act by a person.
Where it stopsWhen the agent made the decision. Insurers are adding exclusions for loss arising out of artificial intelligence, and where the policy treats an insured act as a person's act, an agent's error may not qualify.
Written forUnauthorised access, data breaches and the events around them.
Where it stopsA system working exactly as intended. A hallucinated answer is not a breach, and a bad decision is not an exploit.
Written forBodily injury and property damage.
Where it stopsEconomic loss from an agent's output, the most common AI harm. ISO generative AI exclusions now exist for this line: CG 40 48 01 26 for personal and advertising injury, CG 35 08 01 26 for products and completed operations.
Directors and officers cover is the fourth, and the one people forget. If shareholders or a regulator argue the board failed to supervise the deployment, the claim lands there. Why this matters, with the cases behind it.
We sell nothing. We show you where the cover stops.
Most standard E&O, cyber and general liability policies do not cover AI agent mistakes. They were written before autonomous systems existed, and insurers have been adding explicit AI exclusions at renewal through 2025 and 2026. In the EU the exposure is widening: the AI Act's transparency duties have applied since 2 August 2026, and the revised Product Liability Directive (Directive (EU) 2024/2853) brings software inside strict liability, with a transposition deadline of 9 December 2026. Moffatt v. Air Canada (2024) confirmed that a business answers for what its agent tells a customer. Dedicated AI cover exists, and it asks for structured governance evidence first.
The AI Omnibus, in force since 27 July 2026, moved the AI Act's high risk obligations to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I. Article 5 prohibitions, Article 50 transparency, the Product Liability Directive and the exclusions appearing at renewal did not move. One narrow relief: systems on the market before 2 August 2026 have until 2 December 2026 to meet the Article 50(2) duty to mark AI generated content. The Omnibus brief at agentliability.eu.
In most cases, the business that deployed it.
Under ordinary agency principles the deployer authorised the agent to act and answers for what it did. In the EU the deployer also carries duties under the AI Act (Regulation (EU) 2024/1689), and the revised Product Liability Directive means a claimant may not need to prove negligence. In Moffatt v. Air Canada (2024) a tribunal held that a business cannot disclaim what its chatbot tells a customer.
It tends to respond to five kinds of loss, and the policy form decides which.
Bodily injury or property damage caused by an AI system. A customer's financial loss from an incorrect output. Regulatory defence costs. Data incident and crisis costs. Intellectual property claims over generated content. Exclusions vary widely, so the wording matters more than the name of the product. The first AI agent policy backed by the AIUC-1 standard was secured by ElevenLabs in February 2026.
No carrier in this market publishes a rate card.
Any figure quoted online, including on this site previously, is an estimate rather than a filed price. The only way to a real number is a broker submission naming your sector, revenue, agent type and volume. What specialist programmes consistently ask about is documented governance: written deployment policies and records of human oversight. What is actually published, and what is not.
A small number of providers, much of it through Lloyd's of London or written in the United States.
Munich Re's aiSure is a performance guarantee for AI systems that settles on measurable performance data, aimed at companies that build AI. Armilla is a Lloyd's coverholder writing affirmative AI liability. HSB, part of Munich Re, introduced AI liability cover for small and mid sized businesses in the United States on 18 March 2026, added to partner carriers' business policies rather than sold direct. Agensure works in Europe on adversarial testing and a risk score, and is not an insurer. We have not found a European carrier selling off the shelf AI agent cover to SMEs at scale.
The carrier comparison at agentinsured.eu follows the market in detail.
A written account of what the agent does, how it is governed, who oversees it and what happens when it fails.
AIUC-1, published by the Artificial Intelligence Underwriting Company (AIUC), is a standard for AI agents of 51 requirements and 130 controls across six pillars; the first policy backed by it was placed through Lloyd's in February 2026. Agent Certified, published by Future Proof Intelligence, scores one agent across seven dimensions and produces an evidence file written to be read by underwriters. It is voluntary, and no insurer is committed to a premium change for holding it.
Sixteen questions about your business, your agents, the policies you hold and the records you keep. Your answers stay in your browser; nothing is sent to a server.
You get a reading of each policy you hold, the three risks that matter most for your profile, the products worth asking a broker about, and five next steps.
Run the auditSend them before your next renewal and do not accept a verbal answer. With the replies in hand you know where your gaps are, and can decide whether to accept the exposure, carry it yourself, or work toward dedicated cover.
Errors and omissions, cyber, general liability and D&O, each answered separately, clause by clause.
By number and page. A policy that quietly covered you before can exclude AI at renewal without anyone saying so.
Ask about the ISO endorsements by name, including CG 40 48 01 26 and CG 35 08 01 26, and for the endorsement itself.
If the insured act is defined as a person's, an agent's error may fall outside it.
An exclusion with no alternative is the clearest sign of a gap you are carrying yourself.
Where AI riders exist on cyber, sub limits are typical. Know the number before you rely on it.
Three questions to answer for yourself first What to tell your broker
Each piece takes one question we keep hearing from people running AI in production, and answers it in plain English with the sources shown.
There is no shortcut, and the work is ordinary. Underwriters writing this line want to see that a business knows what it runs and can show it. We explain the path. We do not sell insurance, and we do not sit in the room between you, your broker and the carrier.