Munich Re aiSure. How parametric AI performance insurance works.

Why Munich Re matters to this market

Munich Re is the world's largest reinsurer by premium volume. Its decision to write AI performance risk, starting in 2018, was a signal that the category was insurable at all. Reinsurers sit behind primary insurers, absorbing the upper layers of large losses. Without reinsurance capacity, primary insurers cannot offer meaningful policy limits. With it, the market can develop products with limits that are commercially relevant to enterprise buyers.

The significance of Munich Re's entry into AI coverage is therefore structural, not just commercial. When Munich Re underwrites AI performance risk, it creates the conditions under which Lloyd's syndicates, specialist MGAs such as Armilla and AIUC, and mainstream carriers such as Counterpart can offer primary products backed by credible capacity. For European operators, this matters because it determines whether adequate coverage limits will be available in the European market before the EU AI Act and Product Liability Directive deadlines arrive in the second half of 2026.

What parametric insurance is

Traditional indemnity insurance responds to actual losses. The insured suffers a loss, files a claim, the insurer adjusts the claim to determine what happened and what the loss was worth, and pays accordingly. This process works well for established risk categories where loss is measurable, causation is clear, and precedent is available. For AI, it faces three structural problems.

The first problem is attribution. When an AI system produces a bad output, determining whether the loss resulted from the AI's failure rather than from user error, from the deployer's configuration, or from an upstream data quality problem requires exactly the kind of technical investigation that few claims teams are equipped to conduct. Attribution disputes extend the claims process, increase costs, and often produce settlements rather than principled outcomes.

The second problem is measurement. Many AI failures produce losses that are real but difficult to quantify in the currency of an indemnity claim. A model that systematically undervalues credit applications from a particular demographic produces a regulatory violation and a reputational harm before it produces a quantifiable financial loss that a claims adjuster can evaluate.

The third problem is moral hazard. Indemnity insurance covers actual loss, which creates an incentive for the insured to allow losses to occur in order to recover them. For AI systems, where the insurer has no real-time visibility into performance, this moral hazard is difficult to manage.

Parametric insurance addresses all three problems by separating the insurance trigger from the loss. The trigger is a measurable event, agreed before the policy is bound: accuracy below a specified threshold, hallucination rate above a specified level, uptime below a defined minimum. If the trigger occurs, the insurer pays a pre-specified amount regardless of what the actual loss was. There is no claims adjustment process, no attribution dispute, and no incentive to allow losses: the payout is fixed whether the actual loss is larger or smaller than the parametric amount.

Munich Re aiSure: what it covers

aiSure covers four primary risk categories. The first is algorithmic bias, where the AI system produces decisions or outputs that discriminate on protected characteristics. The second is privacy failures, where the system exposes personal data or confidential information through its outputs. The third is intellectual property infringement, where the system reproduces or generates content that infringes third-party copyright or other IP rights. The fourth is performance shortfall, where the system fails to perform at the level specified in the policy terms, whether measured by accuracy, hallucination rate, uptime, or a custom metric agreed during underwriting.

Coverage limits via the Mosaic Insurance partnership, which provides the direct-to-market distribution for aiSure, are in the USD 15 million range for initial placements. Munich Re has indicated capacity for larger programmes for enterprise clients with established relationships and strong governance documentation.

The coverage is global, meaning European operators can access it without a European-domiciled carrier. This matters in the current market, where no European-native AI insurer exists. Until a European equivalent of AIUC or Armilla emerges, European operators with serious AI liability exposure will need to access coverage through US or UK carriers, often via the Lloyd's market.

How the underwriting process works

Munich Re's underwriting process for aiSure is more technically intensive than underwriting for a traditional property or liability product. The underwriter needs to understand the AI system being covered: how it was trained, what data it was trained on, what its performance benchmarks are, how it is monitored in production, and what governance arrangements exist around it.

In practice, an applicant for aiSure coverage should expect to provide: a technical description of the system, including architecture, training data sources, and test methodology; performance benchmark results, including accuracy, precision-recall, and any hallucination or bias test results; a description of the operational monitoring programme, including what is measured, how often, and by whom; and a description of the governance framework, including the risk management process, oversight arrangements, and incident response procedures.

This documentation requirement is directly aligned with the technical documentation that EU AI Act Articles 11 and Annex IV require for high-risk AI systems, and with the risk management documentation that ISO/IEC 42001 requires for any AI deployment within its scope. An organisation that has invested in compliance documentation for regulatory purposes finds that the same documentation substantially accelerates the aiSure underwriting process.

Third-party validation further shortens the underwriting timeline. Where an applicant can present an independent assessment of the AI system, such as an assessment under the Agent Certified framework, or a technical audit conducted by a recognised security or AI testing firm, the underwriter can rely on that assessment rather than reconstructing the evidence from scratch. The quality of third-party validation affects both the coverage terms offered and the premium.

What parametric coverage leaves exposed

Parametric insurance is a specific product designed for specific risk. It does not cover everything. European operators relying on aiSure or any parametric product as their sole AI insurance should be aware of four coverage gaps.

The first gap is basis risk. Parametric insurance pays on the trigger, not on the actual loss. If the actual loss is larger than the parametric payout, the insured absorbs the difference. If the actual loss is smaller, the insured receives more than they lost. Over a portfolio of incidents, these differences tend to average out, but for a single large incident, basis risk can be material. Buyers should size the parametric limit against a realistic worst-case loss scenario, not a typical one.

The second gap is below-threshold losses. Parametric coverage does not respond unless the trigger threshold is breached. An AI system that produces systematic errors that each fall below the trigger individually but that accumulate into significant aggregate loss over time will not trigger parametric coverage, even though the aggregate loss may be substantial.

The third gap is regulatory penalty exposure. The EU AI Act's Article 99 penalty regime can produce fines up to EUR 35 million or 7 per cent of worldwide annual turnover for the most serious violations, and up to EUR 15 million or 3 per cent for deployer-level breaches. These fines are not insurable under most European insurance regimes, which prohibit coverage for regulatory penalties on public policy grounds. A parametric product that triggers on AI performance failure does not address regulatory penalty exposure.

The fourth gap is product liability claims under Directive 2024/2853. The revised Product Liability Directive, which EU member states must transpose by 9 December 2026, creates strict civil liability for damage caused by defective AI software. A product liability claim does not need a performance threshold to be triggered: it needs a defect and a loss. For the full scope of what the directive covers and what it means for AI providers and deployers, see Agent Liability EU's briefing on the revised Product Liability Directive.

The broader market context

Munich Re is not the only carrier writing AI risk in 2026, but it is the most significant from a market infrastructure perspective. AIUC, founded in 2024 and funded by Nat Friedman and others including Anthropic's Ben Mann, has published its own underwriting standard (AIUC-1) and backed ElevenLabs with the first AIUC-1-backed policy in February 2026. Armilla, a Canadian MGA operating as a Lloyd's coverholder, offers coverage up to USD 25 million and has partnered with Trustible for AI governance evaluation. Counterpart launched its Affirmative AI Coverage product in November 2025, covering hallucinations, misclassification, bias, and deepfake fraud.

What none of these carriers provides is European-native coverage built around the EU regulatory framework. Coverage from US and Canadian carriers is available to European buyers, but the policy language, the underwriting criteria, and the claims process are not designed around Articles 26 and 99 of the EU AI Act or the specific liability regime of Directive 2024/2853. The gap between available coverage and needed coverage is the same gap that makes this publication, and the Agent Certified and Agent Insured networks, structurally relevant to the European market.

The Agentic Liability Monitor, published via the monitor section of this site, tracks market developments including new product launches, underwriting criteria updates, and significant claims or regulatory actions. Operators building an AI risk programme should treat it as a regular reference alongside the coverage framework documentation.

What to do if you are evaluating AI coverage now

The practical steps for a European operator evaluating AI coverage before the August and December 2026 regulatory deadlines are the following. First, map the AI systems in use against their risk profile, using the coverage categories that existing products address and the gaps they leave. Second, produce the technical documentation that parametric underwriters require, because that documentation is also required for regulatory compliance and will be needed regardless of whether coverage is obtained. Third, consider third-party validation to compress the underwriting process and potentially improve coverage terms. Fourth, assess whether parametric coverage alone addresses the organisation's exposure, or whether additional instruments, including a specialist liability product for product liability claims, are needed alongside it.

For organisations at an early stage of AI risk management, the Agent Insured waitlist provides access to a structured intake process that assesses coverage needs, identifies available products, and prepares the documentation file for underwriting engagement.