---
title: How does the new EU Product Liability Directive affect AI agents?
url: https://opencontroller.lyzr.com/blog/eu-product-liability-directive-and-ai-agents
published: 2026-11-16
updated: 2026-11-16
author: Lyzr
---

# How does the new EU Product Liability Directive affect AI agents?

**The new EU Product Liability Directive, Directive (EU) 2024/2853, treats software, including AI systems, as a product. From 9 December 2026 it applies to products placed on the market, so a company that ships an AI agent inside a product or service to people in the EU can face no-fault liability if a defect causes death, personal injury, property damage or loss of non-professional data. Courts can order evidence disclosure and presume defectiveness, so version records and logs become your defence.**

Legal and product teams ask whether "the AI did it" is a defence. Under this directive it is not. The question is whether the product was defective, and agent behaviour after release is part of that test.

## What changed from the old directive?

The 1985 directive was written for physical goods. The new one, published in the Official Journal in November 2024, replaces it. Member states must transpose it by 9 December 2026, and it applies to products placed on the market or put into service after that date.

| Topic | Old regime (85/374/EEC) | New directive (2024/2853) |
|---|---|---|
| Software | Unclear whether covered | Expressly a product, including AI systems |
| Open source | Not addressed | Excluded if supplied outside a commercial activity |
| Defect test | Safety people are entitled to expect | Adds factors such as the product's ability to keep learning after release, and cybersecurity requirements |
| Updates | Not addressed | Manufacturer liable for defects from updates, or missing security updates, within its control |
| Data | Not covered | Destruction or corruption of data not used for professional purposes is damage |
| Evidence | Claimant must prove defect | Courts can order disclosure; presumptions of defect and causation in set cases |
| Expiry | 10 years | 10 years, extended to 25 years for latent personal injury |

## Who is liable when an AI agent causes harm?

- **The manufacturer** of the product that includes the agent.
- **The manufacturer of a defective component**, such as a model or software module, alongside the product manufacturer.
- **Anyone who substantially modifies** a product outside the original manufacturer's control is treated as a manufacturer.
- **Importers and authorised representatives** where the manufacturer is outside the EU, and in some cases fulfilment providers and distributors.

Claimants are natural persons. Damage to property used only for professional purposes is excluded, so purely internal enterprise agents carry less exposure under this directive than consumer-facing ones. Contract and other liability rules still apply. The separate AI Liability Directive proposal was dropped: the Commission announced its withdrawal in its 2025 work programme.

## Why does self-learning behaviour matter?

The defect test now considers the effect of a product's ability to continue to learn or acquire new features after it is placed on the market. For agents, that covers changes in behaviour from new models, retrieval content, memory or tool additions. The manufacturer also stays responsible for updates it controls.

In practice, every model swap, prompt change or new tool is a product change. If you cannot show what version was running when harm occurred, and how it was tested, you are arguing from a weak position. See [EU AI Act obligations for AI agents](/blog/eu-ai-act-obligations-for-ai-agents) for the separate regulatory duties.

## What evidence should you keep for AI agents?

Courts can order disclosure of relevant evidence, and failing to disclose can trigger a presumption of defect. Keep:

1. **Version history** of each agent: model, prompts, tools, MCP servers, configuration, with dates.
2. **Release testing**: evals and red-team results per version.
3. **Security updates**: what was patched, when, and how fast. Pair this with the [EU Cyber Resilience Act](/blog/eu-cyber-resilience-act-and-ai-agents) requirements.
4. **Runtime logs**: prompts, tool calls and actions, linked to agent version and user.
5. **Guardrail and policy decisions**, including blocks and redactions.
6. **Incident records**: detection, quarantine or kill, root cause, fix.
7. **Supplier documentation** for third-party models and components.

Set retention with the 10-year and 25-year periods in mind; see [how long to retain AI agent logs](/blog/how-long-to-retain-ai-agent-logs). The pillar on [mapping agent controls to compliance frameworks](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) covers reusing this evidence.

## How Opencontroller does this

Opencontroller keeps a governed registry of agents, versions, tools and MCP servers, so you can show which version was live at any time. Multi-turn and red-team evals record release testing; LLM judges and live guardrails record runtime behaviour; rogue-agent handling records alerts, quarantine and kill actions. Posture highlights risky configurations before they ship. It produces evidence; legal assessment of liability remains with your counsel.

## Frequently asked questions

**When does the new EU Product Liability Directive apply?**
To products placed on the market or put into service after 9 December 2026, the transposition deadline for member states.

**Is AI software a product under the directive?**
Yes. Software, including AI systems, is expressly a product. Free and open-source software supplied outside a commercial activity is excluded.

**Can a company be liable for an AI agent's behaviour after release?**
Yes. The defect test considers a product's ability to keep learning after release, and manufacturers remain liable for updates and security updates within their control.

To keep version-level evidence for every agent, [book a demo](https://www.lyzr.ai/book-demo/).

## Key facts

Directive (EU) 2024/2853, the new EU Product Liability Directive, replaces Directive 85/374/EEC and must be transposed by 9 December 2026, applying to products placed on the market after that date. Software, including AI systems, is expressly a product; free and open-source software supplied outside a commercial activity is excluded. The defect test considers a product's ability to continue learning after release and cybersecurity; manufacturers remain liable for updates and security updates within their control. Damage includes death, personal injury (including medically recognised psychological harm), property and loss of non-professional data. Courts can order evidence disclosure and presume defectiveness. Expiry is 10 years, 25 for latent personal injury. The Commission announced withdrawal of the AI Liability Directive proposal in its 2025 work programme.

## Related

- [What does the EU Cyber Resilience Act mean for AI agent software?](https://opencontroller.lyzr.com/blog/eu-cyber-resilience-act-and-ai-agents.md)
- [What does the EU AI Act require for AI agents, and when?](https://opencontroller.lyzr.com/blog/eu-ai-act-obligations-for-ai-agents.md)
- [How long should you keep AI agent logs and traces?](https://opencontroller.lyzr.com/blog/how-long-to-retain-ai-agent-logs.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
