Who Is Liable When AI Makes a Mistake? New Product Liability from December 2026

The phone assistant promises a customer a delivery date that production cannot keep. The chatbot quotes a price that does not exist on the price list. The AI-assisted vision system lets defective parts through. In all three cases the same question hangs in the room: who pays?

The short answer today: usually you, not the provider. And the long answer changes on 9 December 2026: that is when Germany's new Product Liability Act, which transposes the European Product Liability Directive, is due to enter into force; the Bundestag passed it on 8 October 2026 and the Bundesrat let it through on 16 October 2026. It expressly counts software, and with it AI systems, as products - with strict manufacturer liability. What it actually means for a mid-sized company, however, is not the headline "manufacturers are now liable", but two points most summaries skip: operational damage remains uncovered, and anyone who uses AI under their own name or modifies it can become the manufacturer themselves.

Key takeaway: A mid-sized company using AI is liable to its own customers for the system's mistakes - the model providers' terms largely exclude liability for outputs and consequential damage. From 9 December 2026 Germany's new Product Liability Act applies, passed by the Bundestag on 8 October 2026 and cleared by the Bundesrat on 16 October 2026 to transpose Directive (EU) 2024/2853: software and AI systems become products whose manufacturers are strictly liable. But the protection covers personal injury and private property, not the damaged machine in your own plant. The real novelty sits elsewhere: anyone who offers a product under their own name or brand, has it made or substantially modifies it counts as the manufacturer - and is strictly liable with no cap.

Who is liable today when AI makes a mistake?

Toward your customer, you are. The customer has a contract with you, not with the model provider; whether an answer came from a human or an agent changes nothing in the contract. Toward the AI provider, only what the contract gives applies - and the terms of the major providers regularly exclude liability for the correctness of outputs and for consequential damage, with low liability caps.

On top sits the layer that has nothing to do with the model: if the error harms third parties - an AI agent autonomously placing a wrong order, a statement being spread - the company running the process remains liable in tort. And internally, managing-director liability under section 43 of the German GmbH Act threatens when the deployment lacked a documented organisation behind it. What exactly a prudent manager must be able to evidence is set out in the article on AI adoption for managing directors.

What does the new Product Liability Act change from 9 December 2026?

Directive (EU) 2024/2853 replaces the 1985 Product Liability Directive and must be transposed into national law by 9 December 2026. Germany transposes it with the Act to Modernise Product Liability Law: the Bundestag passed it on 8 October 2026 in the version recommended by its legal affairs committee (printed papers 21/4297 and 21/8429), and on 16 October 2026 the Bundesrat decided not to call the mediation committee. The act has thus been adopted; it still has to be signed and promulgated in the Federal Law Gazette. The new Product Liability Act is due to enter into force on 9 December 2026. For products placed on the market or put into service up to and including 8 December 2026, section 23 keeps the previous 1989 act in force (status: 18 October 2026).

Five changes matter for mid-sized businesses:

1. Software is expressly a product (Article 4 of the directive, section 2 of the new act), whether it is installed or used as a cloud service; recital 13 of the directive expressly names AI systems as an example of software. The manufacturer is strictly liable - it is enough that the product was defective, regardless of fault.
2. Evidence relief for claimants: the court can order the manufacturer to disclose evidence (Article 9, section 19), defectiveness is presumed in defined circumstances, and where technical or scientific complexity makes proof excessively difficult, a likelihood standard suffices (Article 10, section 20). With an AI system whose internals nobody fully documents, that is a tangible shift.
3. Learning and updates count: when assessing whether a product is defective, the effect of its ability to keep learning after being placed on the market must expressly be taken into account (section 7 number 3). As long as the manufacturer can supply updates, it is also liable for defects caused only by an update or by a missing security update (section 8 paragraph 2, section 9 paragraph 2).
4. Old limits fall: the 500-euro deductible for property damage and the 85-million-euro cap on personal injury are gone; the new act contains neither.
5. Free software stays outside - but only where it is not developed or supplied in the course of a commercial activity (Article 2(2) of the directive, section 2(1) no. 3 of the act). Anyone embedding open source into a product does not buy the exemption along with it.

Why does the new liability barely protect your business?

Because the directive protects natural persons, not companies. Covered are death, personal injury and damage to goods not used exclusively for professional purposes; for data, cover applies only to the destruction or corruption of data not used professionally. A faulty AI module that damages your production line or corrupts your inventory system falls outside its scope - while the same module injuring an employee is covered.

The honest takeaway for procurement: for operational damage, the contract with the provider remains the only lever. Liability clauses, usage limits and recourse rules belong in the contract, not in the hope that a directive will settle it. That holds doubly because the model providers' terms exclude exactly these damages.

When do you become the manufacturer yourself?

That is the directive's actual message, and it affects more mid-sized companies than it appears. Under Article 4 number 10 of the directive and section 3 of the new act, a manufacturer is anyone who develops or produces a product, who has it designed or made - and anyone who puts their name, trademark or other distinguishing mark on it. Under Article 8 paragraph 2, anyone who substantially modifies a product outside the original manufacturer's control and re-supplies or commissions it also counts as the manufacturer; section 5 of the German act ties this to the change being made without the manufacturer's consent.

When a change is substantial is first determined by product safety law, as section 5 paragraph 2 sets out. Where that law contains no yardstick, a change is substantial if it alters the product's performance, purpose or type in a way the manufacturer's risk assessment did not provide for, and if it changes the nature of the hazard, creates a new hazard or raises the risk level. Connecting a model to your own documents in the way the manufacturer provides for is unlikely to reach that threshold as a rule; rebuilding it for a purpose the manufacturer did not intend, creating new hazards, can cross it. No court has drawn the line yet, so every adaptation belongs on record.

Three constellations from the Mittelstand hit this directly:

• White label and own branding: you have an AI agent built and offer it to your customers under your own name - as part of your software or as "company X's AI service". Whenever your name is on it, you are liable like the producer, even if a third party supplied the technology.
• AI as a component: you build an AI module into your own product - vision inspection inside the machine, control logic inside the device. For defects, you as product manufacturer and potentially the component manufacturer are liable side by side (Article 8).
• Substantial modification: anyone who adapts an existing model so far that the change lies outside the original manufacturer's control legally produces a new product - including manufacturer liability.

The directive also matters from the other direction: where the manufacturer sits outside the Union, its authorised representative, the importer and potentially the supplier or fulfilment provider are liable too (Article 8 paragraph 1, sections 10 to 12 of the new act). For AI software from the US there will be, from December 2026, a liable addressee inside the EU for the first time.

Who carries what in a real claim?

Two points sit beside the table that no table carries: insurance and documentation. Whether business or cyber liability covers AI-related damage is decided solely by the policy wording - it belongs on the checklist before a claim, not after. And whoever records the system selection, its usage limits and the approvals in writing stands better in every liability question; the same documentation is what the company AI policy requires anyway.

What became of the planned AI Liability Directive?

It never arrived. In 2022 the Commission proposed a dedicated regime for AI damage - with presumption rules and relief for claimants beyond product liability. It announced the withdrawal in its work programme of February 2025, and the formal withdrawal appeared in the Official Journal on 6 October 2025; the Commission argued that the core concerns - software and AI defects - are now covered by the reformed product liability.

For users the sober conclusion is: there is no dedicated AI liability regime and none is coming. What remains is the new product liability for private-side damage on one hand and ordinary contract and tort law on the other. The middle ground - the operational damage that matters in practice - is still filled by the contract. Anyone tracking AI Act duties should note that fines and roles are a second topic, separate from liability; the assignment of provider and operator roles is covered in the EU AI Act overview.

What should you check before 9 December?

Five points, all feasible without a lawyer - and all worth putting in writing:

1. The liability clauses of running AI contracts: what does the provider exclude, what does it assume, and is there a cap? Whoever sees a cap knows the real protection.
2. The name and brand question: does your name or brand sit on a product with an AI component - including one sold as "powered by" a third party? From December 2026 you count as the manufacturer.
3. The list of substantial modifications: which third-party systems have you adapted, and does the change still sit under the original provider's control? Without a written record you cannot draw that line in a dispute.
4. The placing on the market of your own products: the new liability only bites for products placed on the market or put into service from 9 December 2026; older ones stay under the previous law according to section 23. Anyone shipping an AI feature inside an existing product should know from when they count as the manufacturer of its new state.
5. The insurance policies: ask explicitly whether software- and AI-related damage is covered. A written confirmation beats a verbal assurance.

Does this also apply in Austria and Switzerland?

In Austria, yes: as an EU member it must transpose by 9 December 2026 too - manufacturer definition, strict liability and evidence relief are the same. Switzerland is not an EU member and has its own product liability act, which does not automatically adopt the directive. Two cases still matter: Swiss providers shipping into the EU are caught via importer liability; and a German company selling into Switzerland keeps clarifying liability there by contract.

Frequently asked questions

---

For a concrete project - for example an agent that autonomously places orders or answers enquiries - usage limits and documentation duties are settled most cleanly before launch; what a first process with a legal framework and technical documentation looks like is on the AI automation for companies page. This is not legal advice; for interpretation in a specific case, a specialist lawyer remains the right address.

This article was first published on 19 September 2026. We updated it on 18 October 2026: the Bundestag passed the new Product Liability Act on 8 October 2026 and the Bundesrat cleared it on 16 October 2026; the section numbers refer to that adopted text.

Sources

• EUR-Lex: Directive (EU) 2024/2853 on liability for defective products - transposition deadline 9 December 2026, product definition including software (Art. 4, recital 13), liable economic operators and substantial modification (Art. 8), evidence relief (Art. 9 and 10)
• Bundesrat: Act passed by the German Bundestag to modernise product liability law and amend other provisions, printed paper 584/26 of 9 October 2026 - wording of the new Product Liability Act (sections 1 to 23), entry into force 9 December 2026
• Bundesrat: Procedure 584/26 - plenary of 16 October 2026 (1069th session, item 46): no request to convene the mediation committee
• German Bundestag: Government bill, printed paper 21/4297 of 25 February 2026 - explanatory memorandum, including on software and AI systems
• German Bundestag: Bundestag beschließt Modernisierung des Produkthaftungsrechts - vote of 8 October 2026 based on recommendation 21/8429
• EAPIL (9 October 2025): European Commission Withdraws Proposals on Assignments of Claims and AI Liability - formal withdrawal of the proposal for an AI Liability Directive (COM(2022) 496) in the Official Journal of 6 October 2025