The Tuesday Special No. 6 — Tuesday, September 29, 2026

Good morning, everyone. The regime changes, and it changes at the root. Directive (EU) 2024/2853 dates from October 23, 2024, and it repeals directive 85/374/EEC, the 1985 text that was built around physical objects. Article 2, paragraph 1, sets the time boundary: the new regime applies to products placed on the market or put into service after December 9, 2026. Article 22 asks member states to transpose it by that same date. Two different articles point to the same deadline. Then there's article 4, and this is the heart of it: software is a product. The rule holds in every case. It applies whether the software is embedded in a device, sold on its own, or delivered as a connected digital service. A model, an agent or a recommendation engine must now meet the safety standard the public is entitled to expect. There's one single exception, in article 2, paragraph 2: free and open source software developed outside a commercial activity falls outside the new regime. That exception is read narrowly, and anyone who integrates an open library into a commercial product is fully inside the regime for the product they place on the market.

Hi everyone, and thanks for the question, because the answer sits in a file. An agent that handles accounting and payments either records what it did or records nothing at all. There's nothing in between. Regulation EU 2024/1689, article 12, requires high-risk systems to automatically record events over the entire lifetime of the system. It's a rule that touches design, before any operational choice: the system has to be capable of recording, full stop. When an agent gets a payment wrong, people remember the outcome. The log preserves the order of the steps: which document came in, which tool was called, with which parameter, at what time. It's the only witness to the sequence, and it has to be readable by third parties as well. Then comes the part that weighs on a company: article 19 puts the retention of logs on providers, with a minimum period of six months, unless different terms are set by Union or national law. Article 26, paragraph 6, repeats the same obligation for deployers, meaning the companies that use that system in their own business. Financial institutions have one more rule: those records sit inside the documentation already required by Union law on financial services. For a bank, the agent's log becomes supervisory documentation.

Hello to the table. Munich Re describes aiSure as a set of coverages designed for the companies that sell models and for the ones that put them into production. The group's official page, consulted in September 2026, puts two numbers side by side that explain everything else. 98 percent of CEOs expect immediate benefits from adopting AI, while 60 percent remain hesitant in the face of an uncertain economic return. The distance between those two percentages is the commercial problem facing anyone selling agents. The module for providers is called "Contractual Liabilities": the seller guarantees the customer that the AI will do what was promised, and the insurer backs that guarantee, compensating the customer's losses when the model gets it wrong. Here's what's new. The trust comes from a third party with a reinsurer's balance sheet, instead of from the seller's slides. Then the suite separates two other situations. "AI Liabilities" covers discrimination, intellectual property infringement, hallucinations, regulatory penalties. "Financial Losses" looks at companies automating critical processes and covers lost revenue, business interruption, legal damages. That's three coverages, and each one concerns a different economic risk. The difference between them matters more than the commercial name.

Saga, that third item confirms exactly what I was saying about scope. Article 6, read together with recital 24, lists the recoverable damages as a closed set: death and personal injury, damage to property other than property used exclusively for professional purposes, destruction or corruption of data used outside professional purposes. Purely economic losses stay outside. An agent that gets a payment wrong between two businesses falls outside this regime. It's the point most internal presentations get wrong.

And that's exactly why a reinsurer saw a product here. What the directive excludes stays with the contract, and the contract needs someone to pay for the damage. A botched reconciliation, a duplicated order or a mispriced quote all produce losses between businesses. Those losses stay with the contract and with the applicable national law. Munich Re adds a practical detail for large companies: the coverage applies across multiple models and multiple types of loss. Anyone running twenty agents in production is thinking about a portfolio of risks.

On that portfolio I have a practical doubt. The insurer pays when the model gets it wrong — and that word, "when," has to be proven. If the action log is a mess, who establishes that the error belongs to the agent and not to the operator who uploaded the wrong document? Saga, does the page say anything about technical verification before signing?

It says the technical verification comes before signing. That step is what makes the case interesting. The seller knows their own model, while the buyer cannot see how it works. A performance guarantee shifts the burden of proof onto the seller, and that burden has a price. aiSure works on exactly that price. For a small technology company the message is hard but useful: you're better off guaranteeing the result than polishing the product demo.

And the law reaches the same conclusion in a different way. The promise of accuracy, in accounting terms, becomes a liability booked on the provider's balance sheet. The Munich Re page explains this, and that liability is why companies want coverage.

Let me add the last piece on evidence. The directive we're discussing, in article 9, allows the judge to order the defendant to disclose the relevant evidence in their possession. And article 10, paragraph 2, point a), adds the consequence: the defect is presumed when the defendant fails to make that disclosure. In practice, the log that today looks like nothing more than a storage cost will tomorrow decide whether the company can defend itself, or whether the defect is taken as proven.

The dates don't line up. That's the weak point. The liability regime starts with products placed on the market after December 9, 2026. The obligations for the high-risk systems in annex III, on the other hand, were postponed to December 2, 2027 by the European Commission's digital Omnibus package. That leaves almost a year in which the defect is presumed for missing evidence, while the obligation to produce certain evidence isn't enforceable yet. I haven't verified how judges will handle that period. It also depends on national transposition.

For buyers, that period is the time available to get compliant. Anyone signing a three-year contract today for a platform that runs AI agents is buying a system that will have to be compliant before that contract expires. The question for the purchasing committee becomes a contractual one: does the provider guarantee compliant logging by that date, and with what penalty? But before that, you have to define the scope. An assistant that summarizes internal emails stays outside. A system that assesses the creditworthiness of a natural person falls under annex III.

Leon, the penalty you're talking about is exactly what the insurer is pricing. Now let me ask you the opposite question. If a provider offers you a guarantee backed by an insurance policy, do you as a buyer still look at the action log, or are you satisfied with the certificate?

I look at the log, and before that I look at whether it exists at all. A guarantee pays afterwards. The trace is what tells you what happened. Insurance and evidence solve two different problems. Then there's the bank case, where the log is supervisory documentation. There, no insurance certificate replaces the file. The retention periods change, the format changes, and who can demand those records changes too.

Saga, a question from me too, on the "AI Liabilities" module: it covers regulatory penalties. In European terms that item is delicate, because in quite a few jurisdictions the insurability of administrative fines is limited on public policy grounds. Does the page go into detail by jurisdiction?

The page lists the item among the coverages and stops there. I don't know the jurisdiction-by-jurisdiction detail yet, and your objection is the first thing I'd put to a broker. The actual cases show something else. An error that generates a fine is a compliance problem, while an error that halts a logistics line hits the profit and loss account. Treating them as equivalent leads you to buy the wrong coverage, and it happens often.

I'm watching the national transposition decrees due by December 9, 2026, and in particular how each one translates the evidence disclosure order in article 9. If one jurisdiction makes it quick and another makes it cumbersome, the choice of forum becomes a risk decision, and contract clauses will be written around that difference.

I'm watching the tender specifications of large companies. The moment a clause appears requiring the event log in a format readable by third parties, with the minimum retention period written in black and white, is the moment the design obligation turns into a purchasing requirement. Before that, it's all intentions.

I'm watching whether other reinsurers come up alongside aiSure with competing products over the coming months. A market with a single seller is an experiment. With three it becomes a reference price, and at that point the cost of the guarantee enters agent price lists as just another line item.