Editor's note.
Every product description, product name and carrier name in this article was read on armilla.ai on 31 August 2026, with the product page fetched three times using different questions to guard against a single misread. Nothing about any other provider's terms is asserted here. Section 4 records a correction to this desk's own sourcing note about where a widely quoted figure appears, and the general structural description of liability and warranty products is market description rather than a quotation of any wording.
- One provider, two products, two panels. Armilla publishes Chaucer Group, Axis Capital and Convex behind its liability product, and Chaucer Group, Greenlight Re and Swiss Re behind its warranty product.
- The two are not tiers of the same thing. A liability policy responds to a third party's claim against you. A warranty compensates you when the system misses an agreed measure. Different claimant, different trigger, different proof.
- A reinsurer standing behind a performance warranty is the more informative of the two facts, because it implies somebody has been willing to price measurable AI underperformance rather than treat it as unquantifiable.
- The product page carries no monetary limit anywhere. The figure everyone quotes lives in the resources index, in the title of an announcement item. That is a correction to our own sourcing note, not a change in the market.
- The question a buyer should take from this is not which provider, but which shape. Establish whether your loss arrives as somebody else's claim or as your own shortfall, then buy against that.
Section 1. Two products, and they are not variants
Read at source on 31 August 2026, the provider publishes two distinct offerings under two names.
Armilla Insured is described as affirmative liability cover. The page states that it responds to errors, omissions, and unforeseen performance issues in AI-driven products, and that it covers defense costs, settlements, and third-party claims from AI underperformance. That is the standard architecture of a liability contract, applied to AI: the trigger is a claim made against the insured, and the money moves toward a third party or toward the lawyers arguing about a third party.
Armilla Guaranteed is described as a warranty. The page states that if your AI solution fails to meet contractual KPIs, giving accuracy and bias as its examples, or defined business outcomes, the warranty provides compensation. That is not a liability contract at all in structure. The trigger is a measurement, not a claim, and the money moves toward the insured.
The distinction is worth labouring because the two are routinely discussed as though the warranty were a lighter version of the policy. It is not lighter. It is a different instrument answering a different question, and an organisation can easily need both or neither. The generic architecture of trigger, notification and payment across AI products is set out at how AI insurance claims work, and the first party against third party division that underlies all of this is at first party and third party AI liability cover in Europe.
Section 2. What a panel is, and why the names are information
A panel is the set of carriers and reinsurers sharing a risk behind a product. In most lines of business the composition is uninteresting to a buyer, because the risk is well understood and the names are interchangeable. In a category three years old, the names are one of the few public signals available about who has been persuaded that this risk can be written at all.
Two things follow practically. The first is capacity: more balance sheets behind a product usually means more available limit and less dependence on any single appetite surviving the next renewal. The second is stability: a panel that changes composition between renewals is worth asking about. Change is not a warning sign in itself, and panels rotate for reasons that have nothing to do with the risk. But somebody re-examined the exposure and reached a conclusion, and a buyer is entitled to ask what it was.
The page displays a financial strength grade next to each name. It does not identify which rating agency issued them, so this desk reports the grades as displayed and attributes them to no agency: Chaucer Group shown as A plus, Axis Capital as A and Convex as A on the liability side; Chaucer Group as A plus, Greenlight Re as A minus and Swiss Re as A plus on the warranty side. If a rating matters to your procurement policy, take it from the agency rather than from a product page, which is good practice regardless of provider.
Our standing reading of this provider's position in the European market, including its status as a coverholder, is at the Armilla coverholder analysis, and the broader map of who is writing what is at the AI liability insurance market map.
Section 3. The reinsurer behind the warranty is the interesting fact
Of the two panels, the warranty panel is the one that should hold a European reader's attention, and the reason is technical rather than reputational.
Underwriting a liability policy over AI is difficult but familiar in shape. Somebody suffers a loss, brings a claim, and the questions are the ones liability underwriters have always asked: was there a duty, was it breached, did the breach cause the loss, and how much. AI makes each of those harder to answer, especially causation, which is why defence spend on these claims behaves the way it does, a point we set out at defence costs inside or outside the limit. But the instrument is not novel.
Underwriting a performance warranty over AI is a different proposition. It requires an agreed measure, an agreed baseline, an agreed method of reading the measure, and an agreed view of what counts as the system rather than the environment around it. Accuracy against which test set, evaluated when, by whom, and with what treatment of the cases where the model was right and the process around it was wrong. Naming bias as an example measure raises the difficulty further, since bias is a family of measures that can move in opposite directions on the same system depending on the definition chosen.
For a reinsurer to stand behind that structure implies somebody has done the work of making AI underperformance measurable enough to price. That is a more consequential development for this market than any single coverage limit, because measurability is the constraint the whole category has been stuck on. Our reading of the European reinsurer position more broadly is at European reinsurers and the AI liability market.
One caution against over-reading. A name on a panel is a name on a panel. It is not a statement about how much capacity that party has committed, on what terms, for how long, or with what attachment point, and none of that is published. What the panel tells you is who is in. It does not tell you how far in.
Section 4. The figure that is not where we said it was
A correction, recorded in the body of an article rather than in a footnote, because the habit of hiding corrections is how a corpus rots.
A figure of up to USD 25 million for the standalone AI liability policy is widely quoted, including on this stack, and this desk's own record attributes it in part to the product page at armilla.ai/ai-insurance. On 31 August 2026 that page was read three times with different questions, including an explicit search for the phrase and for any currency figure. No monetary limit, capacity figure or coverage amount appears on it. The only reference to limits is generic: the FAQ states that the product covers financial losses, including defense costs, up to policy limits.
The figure itself is not a fabrication and this is not a report that it has been withdrawn. It appears in the provider's resources index, in the title of an announcement item about the raising of its coverage to that amount. So the fact is sourced. Our note about where it is sourced was wrong, and the difference matters, because a reader following our citation to the product page would have found nothing and reasonably concluded that we or the provider had changed our story. What we have read is an index entry rather than the announcement behind it, so this article states where the figure lives and does not characterise the structure supporting it.
The general rule that comes out of this is worth more than the correction. A published limit is a marketing statement about a programme's maximum. A limit is a term of a specific contract offered to a specific applicant after underwriting, and the only version that governs anything is the one on your quote and your schedule. Treat any figure from any provider's website as an indication of the shape of the market and nothing else. What actually determines whether a limit is adequate is the interaction between the limit, the sublimits, the aggregate and the defence cost structure, which we set out at sublimits and aggregate caps explained.
Section 5. Regulatory violations, and the word insurable
Cover for AI regulatory violations, including defence costs and insurable fines under regimes such as the EU AI Act, is published by this provider and is an increasingly common feature of the category. The word carrying all the weight in that sentence is insurable.
Defence and representation costs are generally insurable in European markets. The fine itself is insurable only where the applicable national law permits a penalty to be insured, and that varies across member states, which means an identical wording produces different outcomes in different jurisdictions with no change to the policy. No insurer can contract around a national rule that a regulatory penalty must be borne by the party penalised, and no buyer should read a coverage grant as a promise that it will be.
The practical version of this question, with the Article 99 ceilings that sit behind it, is at does AI insurance cover EU AI Act fines. The timing point is that the Annex III obligations whose breach would generate the largest exposures now apply from 2 December 2027 under the AI Omnibus, while the transparency obligations already live since 2 August 2026 are enforceable now.
Section 6. Which shape do you actually need
Three questions settle it faster than a product comparison does.
Who suffers the loss when the system underperforms? If the answer is a customer, a patient, an applicant or the public, your exposure is third party and liability cover is the instrument. If the answer is you, because a promised accuracy level is not met or a contracted outcome is not delivered, the exposure is first party and a warranty is closer to the shape of the problem.
Is the failure measurable in advance? A warranty needs a metric agreed before anything goes wrong. If you cannot state today what number the system must hit and how it will be read, a warranty will be difficult to place and difficult to claim on. A liability policy needs no such agreement, because the definition of the loss arrives with the claim.
Where do your contracts already put the risk? If you have promised service levels or accuracy commitments to your own customers, you have already created a first party exposure that a liability policy will not answer. That is the most common gap we see, and it is created by the sales contract rather than by the technology.
Many organisations have both exposures, which is why the two products exist separately. The preparation that makes either placeable is the same underwriting submission, set out at preparing an AI agent underwriting submission, and the evidence an underwriter asks for is converging with the evidence an assessment produces, which we cover at certification, eligibility and premium.
Section 7. Six questions for the broker
- Which of these two shapes are you quoting me, liability or warranty, and if the answer is both, are they one contract or two?
- Who is on the panel for each, in what shares, and has the composition changed since the last renewal?
- What is the limit on my quote, as distinct from any figure published anywhere, and is it per claim or in the aggregate?
- Are defence costs inside the limit or in addition to it, and does that answer differ between the two products?
- For the warranty, what exactly is the measure, who measures it, on what data, and what happens if we disagree about the reading?
- For the regulatory violations cover, in which jurisdictions do you consider fines insurable, and what happens where they are not?
The fifth question is the one that separates a warranty that will pay from one that will argue. A metric with no agreed method of measurement is a dispute with a number attached to it.
Section 8. The point in one sentence
The names behind a product tell you who has agreed to carry the risk and, when they differ between two products from the same provider, they tell you those products are answering different questions, which is the fact a buyer needs before any comparison of limits or premiums means anything.
Questions
What is the difference between AI liability insurance and an AI performance warranty?
The claimant is different, and everything else follows from that. A liability policy responds when somebody else suffers a loss and comes after you for it, so the trigger is a third party claim and the proof is about fault, causation and damage. A performance warranty responds when the system itself fails to meet an agreed measure, so the trigger is a metric being missed and the proof is measurement. Armilla publishes both shapes: it describes its liability product as responding to errors, omissions, and unforeseen performance issues in AI-driven products, and covering defense costs, settlements, and third-party claims; and its warranty product as providing compensation where an AI solution fails to meet contractual KPIs or defined business outcomes.
What is an insurance panel and why does its composition matter?
A panel is the set of carriers and reinsurers sharing the risk behind a product. Its composition matters for two reasons. Practically, more balance sheets usually means more available capacity and less dependence on any one appetite at renewal. Diagnostically, the names tell you who has been persuaded to underwrite this specific risk, which is a live question in a category this young. A panel that changes between renewals is worth asking about, not because change is bad, but because it usually means somebody re-examined the exposure and reached a new conclusion.
What does it mean that a reinsurer stands behind an AI performance warranty?
It implies somebody has been willing to price measurable AI underperformance as a risk rather than treat it as unquantifiable. A performance warranty pays on a metric being missed, which requires an agreed measure, an agreed baseline and an agreed method for reading them. Standing behind that structure is a different underwriting judgement from standing behind a liability policy, where the loss is defined by somebody else's claim. Armilla names Chaucer Group, Greenlight Re and Swiss Re behind its warranty product, and Chaucer Group, Axis Capital and Convex behind its liability product.
Should I quote a coverage limit I read on a provider's website?
No, and this desk had to correct its own record to say so. A widely cited figure of up to USD 25 million for Armilla's standalone AI liability policy does not appear on the product page at armilla.ai/ai-insurance, read three times on 31 August 2026. It appears in the title of an item in the resources index. The figure is not wrong; our note about where it lives was. The general rule that follows is that a limit is a term of a specific contract offered to a specific applicant, and the only version that governs is the one on your quote and your schedule.
Does an AI liability policy cover EU AI Act fines?
Partly, and the distinction is legal rather than commercial. Defence and representation costs are generally insurable. The fine itself is insurable only where the law of the relevant jurisdiction permits a penalty to be insured, and that varies across Europe. Armilla publishes cover for AI regulatory violations including defence costs and insurable fines, and the operative word in every such wording is insurable. Nobody can contract around a national rule that a regulatory penalty must be borne by the party penalised.
Which do I need, the liability policy or the warranty?
It depends on who bears the loss when the system underperforms. If your customers or the public can suffer harm from what the system does, the exposure is third party and liability cover is the answer. If your loss is commercial and internal, an agreed accuracy level not being met, a promised outcome not delivered, a contractual service level breached, then the exposure is first party and a warranty is closer to the shape of the problem. Many organisations have both exposures, which is exactly why the two products exist separately rather than as tiers of one.