Insuring AI agents,
followed closely.
Who writes cover in Europe, what it responds to, and what underwriters ask to see first. An independent desk: we sell no insurance.
- USD 25 millionArmilla's published limit, up to that amount per organisation, underwritten by certain underwriters at Lloyd's.
- EUR/USD/CAD 15 millionInitial capacity for Munich Re aiSure, underwritten and marketed by Mosaic Insurance.
- 9 December 2026Transposition deadline for the revised Product Liability Directive, which treats software as a product.
Five kinds of loss. What answers each today.
No European-domiciled carrier sells an off-the-shelf AI agent liability policy. Four sources of dedicated cover reach a European buyer: Munich Re aiSure through Mosaic, Armilla as a Coverholder at Lloyd's, Counterpart on professional liability lines, and bespoke Lloyd's programmes placed through a broker. Read by the kind of loss, the picture is narrower still.
| Kind of loss | What responds today | Where it stops |
|---|---|---|
| A wrong output causes financial loss | Munich Re aiSure, written for AI developers and vendors, settles performance shortfalls on measurable performance data. Vouch names hallucinations in its affirmative AI cover, in the United States. | Coalition's AI endorsement does not respond to loss from a wrong output. HSB's small business cover excludes model errors and biased outputs. |
| Data leakage and privacy | Cyber wordings. Coalition extends security failure and data breach to an AI security event, on US Surplus and Canada cyber policies. | Outside a security event, the answer sits in your own cyber and professional indemnity wordings, read line by line. |
| Intellectual property | Vouch names certain intellectual property disputes. HSB covers personal and advertising injury arising from the insured's use of AI. Both are United States products. | ISO's CG 40 48 01 26, where attached, removes personal and advertising injury arising out of generative AI from a general liability policy. |
| Regulatory fines and defence | Armilla covers AI regulatory violations, including defence costs and insurable fines under the EU AI Act and the Colorado AI Act. QBE North America publishes an AI regulatory compliance endorsement. | Whether a fine can be insured at all is a question of member state law. Article 99 sets ceilings of up to EUR 35 million or 7 percent of worldwide turnover. |
| An agent acts, or is deceived into acting | Coalition extends funds transfer fraud to a fraudulent instruction sent through deepfakes or other AI. HSB covers bodily injury and property damage from the insured's use of AI. | An agent acting within its authority and getting it wrong is placed bespoke, through a Lloyd's broker, where it is placed at all. |
Every line is read at the carrier's own domain. The detail for each programme, with its source and the date it was checked, is in the carrier matrix.
The carrier matrix Where your E&O, cyber and liability policies stop
Every carrier fact is read at the carrier's own source.
What underwriters ask to see first.
Underwriters price to what an operator can show. The questions fall into three groups, and each maps to dimensions of Agent Certified, the published evidence standard for one AI agent in production.
Technical risk
Model reliability, evaluation coverage, failure modes, and the provenance of the data the agent reasons over. The most scrutinised group at submission. Trust and Safety, Context Integrity, AI Integration.
Organisational risk
Board accountability, oversight staffing, named roles, and the autonomy envelope the operator has set for the agent. The group regulators look at first. Governance, Autonomy Envelope.
Operational risk
Who can invoke the agent, under what authority, with what approval and rollback. The group that decides what triggers a notice of loss. Distribution Control, Product Maturity.
No insurer has adopted Agent Certified as a named underwriting reference, and none is committed to a premium change for holding it. It is written so that an underwriter can read it.
The evidence standard, at agentcertified.eu What underwriters ask before writing
The dates cover is written against.
The AI Omnibus entered into force on 27 July 2026 and moved the high-risk dates. Underwriting did not wait for it. Specialist capacity kept expanding through the whole period in which the delay was uncertain, because specialist underwriters price to operator practice.
- AI Omnibus in forceThe high-risk dates move, six days before the original deadline.
- Transparency and governanceArticle 50 applies, and the AI Office and national authorities hold enforcement.
- Product Liability DirectiveTransposition deadline for Directive (EU) 2024/2853. Software is a product.
- Annex III high-riskObligations for stand-alone high-risk systems apply.
- Annex I high-riskSystems embedded in regulated products.
Recent writing from the desk.
Long form notes on how AI agent liability is being written, what the wordings actually say, and where a European buyer stands. Each carrier fact is verified at the carrier's own domain.
- Two clocks after an AI incident: the regulator's report and the insurer's notice
- A ten year tail on a one year policy: the Product Liability Directive and claims-made AI cover
- An AI-forged instruction moved the money. Which policy responds?
- Your insurer received the same frontier AI warning you did
- What the panel behind an AI policy actually tells you
- Defence costs are the number that decides what your AI limit is worth
Tell us what you run. Hear when it matters.
Joining records your interest and nothing more. It is not an application, a quote or an offer of insurance, and Agent Insured does not pass your details to insurers or brokers. When the market moves for the kind of agent you describe, we may write to you.
For risk leads, brokers and carriers. Published by Future Proof Intelligence.