Anyone who develops software, or builds AI solutions for clients, has so far thought about liability mainly in terms of the contract: warranties, service levels, limitations of liability. From 9 December 2026 that is no longer enough. Directive (EU) 2024/2853 on liability for defective products, which replaces Directive 85/374/EEC after almost forty years, expressly classifies software, AI systems included, as a "product". This means that whoever produces it is liable for damage caused by a defect without the injured person having to prove fault, and no contractual clause can exclude it.
What changes and from when
The directive (published in the EU Official Journal on 18 November 2024) must be transposed by 9 December 2026 and applies to products placed on the market or put into service after that date; earlier products remain under the current regime. In Italy the delegation is in the European Delegation Law (Law no. 36 of 17 March 2026). On 4 August 2026 the Council of Ministers approved the draft decree in preliminary reading, sent to Parliament as Government Act no. 434: it rewrites Title II of Part IV of the Consumer Code (Articles 114-127, plus a new Article 126-bis on publication of judgments). At the end of October 2026 the decree has not yet been finally adopted: its core content, however, is set by the directive.
The changes that matter for those working with software and AI:
| Topic | Current regime | From 9 December 2026 |
|---|---|---|
| Software and AI | Not expressly included | Products, however supplied (installed, cloud, SaaS) |
| Damage to property | €387 threshold (Italy) | No threshold, no cap |
| Data | Not covered | Destruction or corruption of data not used for professional purposes |
| Psychological harm | Not expressly provided for | Covered if medically recognised |
| Proof | On the injured person | Disclosure of evidence and presumptions of defect and causation |
One rule stays the same: only natural persons can claim. Property used exclusively for work and purely economic losses remain outside.
Who is liable: developer, integrator, modifier
Article 8 sets up a chain of liable parties. In the front line is the manufacturer, i.e. whoever develops the software or AI system and places it on the market under its own name; alongside it, the manufacturer of a defective component integrated into the product. If the manufacturer is outside the EU, the importer or the authorised representative is liable and, failing them, the fulfilment service provider; distributors and online platforms are liable on a residual basis, if they do not identify the liable party within one month.
Two cases closely concern SMEs that build AI solutions:
- Integrating is producing. Whoever assembles a product using a third-party model, libraries or components is liable for the finished product, including the open source it integrates: the exclusion in Article 2(2) only covers free software supplied outside a commercial activity.
- Substantially modifying is producing. Whoever modifies a product outside the original manufacturer's control, in a way not foreseen in its risk assessment, and puts it back on the market becomes a manufacturer in turn. It is the same reasoning the AI Act applies to those who customise someone else's model or system.
When software is "defective"
Defectiveness is about safety, not usefulness: a product is defective if it does not provide the safety a person is entitled to expect or that Union or national law requires. Article 7 lists the circumstances to consider, and some are designed precisely for software:
- the product's ability to continue to learn or acquire new features after being placed on the market;
- the foreseeable effects of interconnection with other products;
- relevant cybersecurity requirements, such as those of the Cyber Resilience Act;
- recalls and other interventions by authorities.
The most delicate point is time. For software, the manufacturer cannot argue that the defect did not exist at the time of sale if the defect stems from an update, a related service or the failure to supply a security update within its control (Article 11(2)). The directive does not require updates, but whoever updates, or stops doing so, bears the consequences. A substantial modification also restarts the ten-year period within which claims can be brought (Article 17); the limitation period remains three years from awareness of the damage, the defect and the liable party (Article 16).
The trial: evidence to disclose and presumptions
The real shift is procedural. If the injured person presents enough to make the claim plausible, the court can order the defendant to disclose relevant evidence (Article 9), with protection for trade secrets. If the defendant fails to do so, defectiveness is presumed. It is also presumed when the product does not meet mandatory safety requirements for that risk, or when the malfunction is obvious during reasonably foreseeable use (Article 10). Causation is presumed where the damage is typically consistent with the defect; and where technical complexity, typical of AI systems, makes proof excessively difficult, the court may presume defect, causation or both if the injured person shows they are likely. These presumptions are rebuttable: they are overcome with evidence to the contrary, that is, with documentation.
It is the same thread that runs through Italy's Legislative Decree 160/2026, which at national level governs damage caused in the use of AI systems. The two routes are cumulative: the directive's route requires proving neither fault nor an AI Act violation. The proposed EU directive on non-contractual liability for AI, by contrast, was withdrawn by the Commission in 2025.
What to do before 9 December
- Inventory your products. Which software and AI solutions you will place on the market, or substantially modify, after 9 December 2026, and who their end users are (natural persons included).
- Map your components. Third-party models, open-source libraries, related cloud services: for each, who the manufacturer is and what recourse you have against it.
- Document safety. Risk assessment, testing, vulnerability management, logs and version history: this is the material a court can order you to disclose, and what you need to rebut the presumptions.
- Define your update policy. How long you provide security updates, how you distribute them, what happens if the client does not install them.
- Review supply-chain contracts. Exclusion clauses do not work against the injured person, but they govern recourse between companies: with suppliers and clients, and with micro and small enterprises whose software components you integrate (Article 12(2)).
- Check your insurance. Make sure your product liability policy expressly covers software, data and psychological harm.
Where to start
For those who develop AI solutions, the new directive turns technical documentation from a formality into a line of defence. The first step is to understand which of your products fall under the new regime and whether the documentation you have today would withstand a disclosure order. An independent gap assessment cross-checks products, components, contracts and documentation against the obligations of the directive, the AI Act and the Cyber Resilience Act, and shows what to fix before 9 December.