Munich Re Insures AI: Who Pays When the Agent Gets It Wrong

A policy that guarantees the model's performance

Munich Re describes aiSure as a suite of covers for artificial intelligence systems, designed for those who sell models and for those who put them into production. The group's official page^[1], consulted in September 2026, sets two numbers side by side: 98% of CEOs expect immediate benefits from adopting AI, while 60% remain hesitant in the face of an uncertain financial return.

Between those two percentages lives the commercial problem facing anyone who sells agents.

The buyer signs a contract based on a promise of accuracy. The seller knows their own model; the buyer sees a closed box. A performance guarantee shifts the burden of proof onto the seller, and that burden carries a price.

aiSure works on exactly that price. The same page explains why the vendor looks for help: the promise of accuracy becomes a liability booked on its balance sheet.

The original idea: the guarantee leaves the vendor's balance sheet

The module built for vendors goes by the name «Contractual Liabilities». The vendor guarantees to the customer that the AI will do what was promised, and the insurer stands behind that guarantee, compensating the customer's losses when the model gets it wrong.

Munich Re's wording refers to performance guarantees that allow the vendor to indemnify customers for financial losses or legal liabilities directly linked to AI errors.

The leap sits right here: trust comes from a third party with a reinsurer's balance sheet, rather than from the seller's slides. A procurement director is then assessing a promise backed by capital, and the negotiation changes in nature. The risk stays intact; what changes is the hand holding it.

For anyone running a technology SME the message is blunt and useful: competitive leverage lies in guaranteeing the outcome, well before polishing the demo.

Three distinct risks, three different recipients

The suite separates three situations, and the distinction matters more than the commercial label.

• Contractual Liabilities: for those who sell AI, it covers the performance guarantee given to the customer.
• AI Liabilities: for vendors and companies alike, it covers discrimination, intellectual property infringement, hallucinations and regulatory penalties.
• Financial Losses: for those automating critical processes, it covers lost revenue, business interruption and legal damages.

The three items tell different economic stories. An error that triggers a fine belongs to the world of compliance; an error that halts a logistics line belongs to the income statement. Treating the two as equivalent leads to buying the wrong cover.

Munich Re adds one operational detail for large companies: the cover spans multiple models and multiple loss scenarios. Anyone running twenty agents in production is therefore reasoning about a portfolio of risks, rather than about a single application.

The technical assessment comes before the signature

Here is the friction that makes the case instructive. To obtain cover, the solution goes through a technical examination by the insurer, as the product documentation indicates: whoever guarantees an outcome insists on measuring it.

The practical consequence is severe. An agent with no defined metric, no baseline and no agreed measurement method falls outside the insurable perimeter.

That filter cuts the market in two. On one side, the vendors able to document accuracy, test datasets, degradation over time and alert thresholds. On the other, those who describe their AI with adjectives: the first group obtains a policy, the second stays stuck at the pilot.

The scientific ground beneath that measurement remains open and heavily debated: the Harvard Data Science Review^[2], published by MIT Press, hosts the peer-reviewed debate on how data science systems are evaluated. A CTO who follows that debate understands why an insurer asks for numbers and protocols.

The perimeter Directive 2024/2853 leaves open

The new European directive on defective products, adopted in 2024, brings software and AI systems inside the notion of a product. Its Article 6 lists the recoverable damages: death and personal injury, damage to property, destruction or corruption of data used for private purposes.

Pure economic loss between businesses stays out of that list.

An agent that gets a demand forecast wrong and burns three million in margin produces precisely that kind of damage. The affected customer looks at the contract, not at the directive. This is where aiSure slots in: it insures economic loss between businesses, the territory the European legislator hands over to private agreement.

For a board the reading is direct: protecting revenue runs through clauses and policies, while the directive protects people and property.

The agent's log becomes the evidence

Every guarantee lives on evidence. To show that the AI missed the promise you need the record of its actions: model version, prompt, input data, response, confidence threshold, human approval step.

A partial log turns a claim into a clash of opinions. A complete log turns it into a calculation.

The European text pushes in the same direction, because it strengthens the producer's obligations to disclose evidence and allows the court to presume the defect when that evidence stays in the drawer. Whoever keeps traceability defends their position twice: before the insurer and before the court.

The operational advice for a head of product carries a low cost and a high value: design the log in the same sprint in which you design the agent. Adding observability after the first incident costs a great deal more.

Sustaind and Mosaic: the product in circulation

Two named references show that the cover has left the laboratory. Sustaind offers automated monitoring of legal obligations and their updates in more than 120 countries, backed by a guarantee from Great Lakes Insurance SE, a Munich Re group company.

The second reference is Mosaic Insurance, which has entered a partnership with Munich Re to bring aiSure to its own clients. The stated rationale carries weight: AI performance risk often sits outside traditional cyber policies.

Here comes the part I keep carefully separate. The group publishes a downloadable case study and a description of the agreement, while figures on claims paid, premiums collected or recovery rates stay outside public communication.

An announced result is worth less than a result verified by an independent third party. The reader deserves this distinction: the product exists and has partners with names, while its indemnity track record is still waiting for public numbers.

What you can take away

A CFO or a controller buying an agent now has a list of concrete questions to put on the vendor's table.

• Which performance metric do you guarantee, with which baseline and which measurement method?
• Does the guarantee rest on a policy? Which insurer, which limit, which exclusions?
• Who pays for the economic loss in my process when your model gets it wrong?
• Which action log do you hand over, in which format and for how many months?
• How often do you re-measure accuracy after a model update?

The answers separate mature vendors from the rest in a single meeting. A seller who has passed an insurer's technical examination answers with documents; the others answer with enthusiasm.

The open question concerns every organisation putting an agent inside a process that is worth money today: does the contract you signed say who pays for the damage, or does it leave that field blank?