If your company runs a customer-facing AI system and your renewal broker has not shown you the AI exclusions in your liability program, assume the coverage you think you have is gone or going. Since January 2026 carriers can attach an ISO endorsement that removes generative AI losses from general liability, and the broadest carrier wordings strip AI from D&O and E&O too. The affirmative AI market that fills the hole is real, small and expensive. For an enterprise that deploys AI, start with Armilla's Lloyd's-backed AI liability policy, take Testudo as the second quote, and buy Munich Re's aiSure only if you sell an AI product whose accuracy you warrant in writing.
All limits and product details below were checked on each provider's own page or in trade reporting on November 10, 2026. None of these carriers publishes a rate card, so every price in this guide is indicative.
| Product | Built for | What pays out | Published limit | Price (Nov 10, 2026) | Do not buy it if |
|---|---|---|---|---|---|
| Armilla AI Liability (Lloyd's) | Enterprises deploying gen AI and agents | Third-party claims from model errors, hallucinations, agent mistakes, output harms, AI regulatory fines | Up to $25M aggregate, $5M per model | Not published; quote via broker | Your exposure is first-party loss of your own revenue |
| Chaucer + Armilla Vanguard AI | Buyers who want cyber, tech E&O and AI in one structure | Chaucer for cyber events, Armilla for AI losses with no cyber event | $25M+ AI aggregate plus $10M cyber | Not published | You already have a cyber tower you like and won't move |
| Munich Re aiSure | AI vendors backing performance warranties; some deployers | Measured underperformance against a threshold written into the policy | None on Munich Re's page; up to €/$15M via Mosaic | Not published | You can't define accuracy as a number you track |
| Testudo (Lloyd's) | US enterprises adopting gen AI | Third-party claims from AI-generated output, defence and damages | Up to $9.25M per insured | Not published | You need more than ~$9M from one market |
| AIUC (AIUC-1 + cover) | Agent vendors selling to enterprises | Terms tied to AIUC-1 audit results | Not published on its site | Not published | You deploy third-party AI and sell nothing |
The loser is the option most companies are on today: relying on silent cover in a CGL, E&O or D&O form that was written before anyone thought about AI. That cover is being withdrawn by endorsement, one renewal at a time.
Which AI Exclusions Are Entering Standard Policies?
Three ISO endorsements and a growing set of carrier wordings now remove AI losses from policies that used to respond by default. Verisk's ISO made the general liability forms effective from January 2026, and they are optional, so whether you have one depends on your carrier and state.
Per Insurance Journal's August 17, 2026 report, the three forms are:
- CG 40 47 removes bodily injury, property damage, and personal and advertising injury arising from generative AI under commercial general liability.
- CG 40 48 removes only personal and advertising injury, leaving bodily injury and property damage in place.
- CG 35 08 removes bodily injury and property damage tied to generative AI in products and completed operations.
Adoption is fast. Insurance Business reported on September 16, 2026 that W.R. Berkley, Chubb, Travelers, Berkshire Hathaway and AIG had each filed to adopt the ISO forms or proprietary versions by April, and that a review of state filings found 2,369 generative AI exclusion records in force across 49 states by mid-2026.
Management and professional lines are moving the same way, with broader language. W.R. Berkley's "absolute" AI exclusion, aimed at D&O, E&O and fiduciary liability, bars claims tied to "any actual or alleged use, deployment, or development of Artificial Intelligence," according to Hunton Andrews Kurth. Its definition of AI starts with "any machine-based system that, for explicit or implicit objectives, infers, from the input it receives," which could reach almost any software you run. In November 2025 the Financial Times reported that Great American, Chubb and W.R. Berkley had asked U.S. regulators to approve AI exclusions; AIG, also named in the FT story, told TechCrunch it had "no plans to implement them at this time."
The insurers' worry is correlation. TechCrunch, citing an Aon executive, put it this way: "What they can't handle is an agentic AI mishap that triggers 10,000 losses at once." A Lathrop GPM partner told Insurance Journal she expects "rapid adoption of some version of an AI exclusion in most lines of coverage," with E&O and D&O first.
Fenwick's June 2026 analysis also warns that AI risk can be moved out of a policy through definitions and form changes even when no clause carries the word "AI," and that exclusions across cyber, tech E&O, D&O and EPLI can stack so that no policy responds to a claim at all.
Cyber is the one line leaning the other way in places. Coalition added a deepfake response endorsement to its cyber policies in December 2025. That helps with responding to a deepfake (forensics, takedowns, crisis communications) and the reputational harm. It does nothing for a chatbot that gives a customer wrong advice, because no security event occurred.
What Armilla, Munich Re and Testudo Actually Insure
The three affirmative products split on one question: does the policy pay for a third party's claim against you, or for your own measured underperformance? Armilla and Testudo are liability products. aiSure is a performance guarantee.
Armilla: The Broadest Deployer Policy
Armilla is a Lloyd's coverholder and MGA, and its AI liability policy is underwritten by certain underwriters at Lloyd's. In January 2026 it raised limits to $25 million per organisation. Its homepage lists six covered areas: AI model error (third-party financial loss from errors, inaccurate outputs and hallucinations), model output (defamation, breach of confidentiality, trade secret disclosure), agent mistakes (wrong decisions, failures to escalate), non-breach privacy and data leakage, AI property damage, and AI regulatory violations including defence costs and fines under the Colorado AI Act and EU AI Act. The limit is up to $5 million per model inside the $25 million aggregate, and the page notes that "coverage depends on the issued policy wording."
The Chaucer partnership, Vanguard AI, launched February 10, 2026, pairs Chaucer's cyber and tech E&O cover with Armilla's AI policy and sets allocation rules in advance for mixed incidents. That matters more than the limit. The ugliest AI claims will be the ones where your cyber carrier says it was a model error and your AI carrier says it was a security event.
Skip it if your main fear is first-party loss, your own revenue lost because a model drifted. Armilla's list is third-party claims and regulatory defence. And at $5 million per model, a company running one high-exposure agent gets a fifth of the headline limit for that agent.
Munich Re aiSure: For Vendors Who Warrant Accuracy
aiSure pays when an AI system misses a performance threshold written into the policy. Munich Re's aiSure page offers a provider version that backs a vendor's performance warranty to its customers and a deployer version covering lost revenue, business interruption and legal damages from AI errors. Listed risks include underperformance, hallucinations, bias, privacy leakage and IP infringement from training data. The page frames the risk as performance failures not caused by malicious attacks, and it publishes no limits, deductibles or exclusions.
Distribution runs through partners. Mosaic Insurance launched aiSure-backed cover on February 27, 2026 with up to €/$/CAD 15 million in initial limits for AI developers and vendors, leaving outages, data breaches and negligent deployment to cyber and tech E&O.
Skip it if you are a deployer whose AI risk is a chatbot saying something wrong to a customer. aiSure needs an accuracy metric you can measure against ground truth (a lending model's default predictions, an image classifier's labels). If you cannot write that number into a contract, you cannot write it into this policy. For a typical enterprise rolling out gen AI assistants, this is the wrong product.
Testudo: The Second Quote
Testudo is a Lloyd's coverholder writing standalone generative AI third-party liability for US enterprises. In March 2026 Atrium and QBE joined Apollo on its panel, lifting capacity to $9.25 million per insured, for claims arising from AI-generated outputs that a CGL may not cover.
Don't rely on it alone if you need tens of millions of limit. At $9.25 million, Testudo is a primary layer or an excess layer above Armilla, not a tower.
AIUC: Certification First
AIUC calls itself a standards and insurance company and publishes the AIUC-1 standard for agents, which its site says is grounded in technical testing and updated quarterly. As checked November 10, 2026, the site lists no insurance limits, carriers or covered perils. It belongs on your list only if you sell agents to enterprises and their procurement teams are starting to ask for AIUC-1.
What Does AI Liability Insurance Cost?
No provider publishes a rate, so treat any number as indicative. Insurance Insider reported on September 1, 2026 that some players aim for roughly $100,000 per $1 million of cover for large insureds and about a quarter of that for smaller companies, that large buyers should expect self-insured retentions of $1 million to $5 million, and that most AI liability underwriters offer $5 million to $25 million of limit.
Here is the same workload priced against that range: a large enterprise running two customer-facing gen AI agents that wants $10 million of AI liability limit. At the large-insured rate that is about $1 million a year in premium, above a retention of $1 million to $5 million. At the quarter rate the report attributes to providers working with smaller companies, it would be about $250,000, though a large enterprise should not count on that rate. That is 10% of limit per year at the top of the range and 2.5% at the bottom, set by a market with very little loss history, so expect it to move at every renewal.
The same report says submissions have come from Anthropic, JP Morgan, Walmart and Meta, often seeking limits in the tens or hundreds of millions, which the market cannot yet fill.
What Underwriters Ask For Before They Bind
Each of the three affirmative carriers underwrites the model, not just your company. Armilla says each policy includes independent AI system certification and risk reporting, and that it draws on more than 500 AI evaluations. Munich Re says AI solutions must pass its technical due diligence to obtain coverage. QBE, on joining Testudo's panel, pointed to the AI risk scoring and reporting tools it offers clients.
In practice, have these ready before the broker goes to market:
- A model inventory covering every AI system in production, including vendor AI embedded in tools you buy and what employees use on their own. Fenwick lists AI-use inventories among the steps that will matter.
- Evaluation evidence per system: accuracy against a held-out set, hallucination and refusal rates on your own prompts, red-team results for agents with tool access.
- Monitoring and escalation: drift alerts, human review thresholds, and a kill switch for agents. Our agent monitoring guide covers the tooling, and Langfuse or a governance platform such as Credo AI can produce the logs an underwriter wants to see.
- The contracts: your AI vendor terms, including caps and indemnities, because the underwriter will ask who else pays.
Answer the application carefully. Hunton notes that AI use changes so fast that answers can be out of date before renewal, and a wrong answer on an application is an easy basis for a coverage dispute.
Where Vendor Indemnities Stop and Insurance Starts
Vendor contracts cover IP claims and almost nothing else, so they do not replace AI liability insurance. Anthropic's Commercial Terms (effective June 17, 2025, checked November 10, 2026) are a typical example. Anthropic defends you against third-party claims that your paid use infringes intellectual property, with carve-outs for your own inputs, your modifications, and combinations with technology Anthropic did not provide. Its liability is capped at fees paid in the previous 12 months, though the cap does not apply to indemnification. The same terms disclaim that outputs are accurate, and require you to tell your users not to rely on factual assertions without checking them.
So a hallucinated answer that costs a customer money is your loss under the vendor contract, and increasingly it is excluded under your liability program too. That is the gap Armilla's model error cover and Testudo's output liability are written for. For the IP side, our copyright indemnity comparison shows how much the cover varies by route of purchase.
How to Decide: The Criteria That Predict Regret
The choice turns on three facts about your exposure, and most buyers get the first one wrong.
- Who gets hurt when the model is wrong. If it is a customer or the public, you need third-party liability (Armilla, Testudo). If it is your own P&L, or a customer you have given a contractual accuracy promise, you need a performance product (aiSure).
- How much limit one incident could burn. One agent touching thousands of customers can produce correlated claims, the scenario insurers fear most. Size the limit to the worst batch event, then check Armilla's per-model sub-limit against it.
- Whether your cyber and AI covers talk to each other. Two carriers arguing over one claim is the most likely way a paid premium ends in a denied claim. A coordinated structure like Vanguard AI, or matching wording negotiated by your broker, beats a higher limit on paper.
What changes the answer: if your CGL renewal comes back without CG 40 47 and your E&O carries no AI exclusion, a standalone policy may be worth deferring a year, provided you get that confirmation in writing. If you are an AI vendor, aiSure moves to the top of the list.
Your Renewal Checklist for AI Exclusions
This Week:
- Ask your broker in writing whether CG 40 47, CG 40 48, CG 35 08 or any proprietary AI exclusion is on your current CGL, or proposed for renewal.
- Pull your D&O, E&O, fiduciary and cyber forms and search the endorsements for "artificial intelligence," "machine learning," "algorithm" and "automated." Read the definitions as well as the exclusions.
This Month:
- Build the model inventory and evaluation pack above. You need it for the application whether or not you buy new cover.
- List every AI vendor contract with its liability cap and indemnity scope, and mark which losses no contract covers.
Before Your Next Renewal:
- Get quotes from Armilla (alone and as Vanguard AI with Chaucer) and Testudo, and ask aiSure only if you sell an AI product with a measurable warranty.
- Ask each carrier to state in writing how it allocates a claim that involves both a security event and a model error.
The Bottom Line
Cyber insurance went through the same cycle a decade ago: risks covered silently in property and liability forms, then written out by exclusion, then sold back as a standalone line with its own questionnaire and price. AI is moving through that cycle much faster, with ISO forms published and filings in 49 states within a year. The standalone products are thin and costly, but a model inventory and an evaluation pack cost you little and improve every quote.
Get the exclusion list from your broker this week, before the renewal paperwork arrives.
Continue Reading
- Your Claude Copyright Cover Depends on the Route You Buy
- AI Cyber Insurance Riders: Why 40% of Claims Get Denied
- AI Agent Accountability Act Targets the Company Running the Agent
- FTC Chair's Rogue-Agent Theory Blames Whoever Gave the Order
- AI Vendor Security Review: 6 Questions That Change the Answer
- When the Colo's Chillers Fail, the GPU Loss Is Yours