The EU AI Liability Directive is dead: what actually governs AI agent harm now

On 6 October 2025, a short notice appeared in the EU's Official Journal confirming what the European Commission had already signalled eight months earlier: the AI Liability Directive was not coming. The withdrawal followed a Commission decision at its 2533rd meeting on 16 July 2025, and closed a proposal that had been on the table since 2022 — the one piece of EU law written specifically to make it easier to prove an AI system caused you harm.

If you run agents that touch EU users, or sell into the EU, the practical question is not "what did we lose." It is: what governs an AI agent causing harm now, and does it cover the harm your agent could actually cause?

What the directive was meant to do

The proposed AI Liability Directive would have complemented the AI Act rather than duplicated it, easing a claimant's path to compensation for AI-caused harm along two specific mechanisms: a court's power to order disclosure of evidence from a provider or deployer of a high-risk AI system, and a rebuttable presumption of a causal link between an AI system's non-compliance and the resulting harm, where the claimant could show fault was likely but proving causation was disproportionately hard given how the system worked. It was designed for exactly the case that makes agents hard to litigate: the claimant cannot see inside the system, and the operator holds all the evidence.

Why it was dropped

The Commission's 2025 work programme, adopted 11 February 2025, listed the directive for withdrawal with a one-line justification: "no foreseeable agreement." Member states and industry groups had spent two years unable to converge on the scope and mechanics, and the withdrawal landed alongside a broader simplification push that also dropped the ePrivacy Regulation. The Commission's own text left the door open rather than closing it: it "will assess whether another proposal should be tabled or another type of approach should be chosen." As of this piece, no replacement has been tabled.

What already applies instead

The gap is not empty. The EU's overhauled Product Liability Directive, Directive (EU) 2024/2853, was adopted separately and was never at risk — it entered into force in the weeks after its publication on 18 November 2024, and member states have until 9 December 2026 to transpose it. Unlike the withdrawn directive, it is not AI-specific, but it was substantially rewritten with software in mind: the definition of a covered "product" now explicitly extends to software, firmware and AI systems, closing a long-standing ambiguity about whether a defective model was a "product" at all.

The mechanics carry over some of what the withdrawn directive was trying to do. A court can compel a defendant to disclose relevant evidence once a claimant has made a sufficiently plausible case. And where the claimant can show a defect is likely but proving it is excessively difficult because of the product's technical complexity — or where the defendant fails to comply with a disclosure order — the product is presumed defective. Practically, that is close to what the AI Liability Directive would have delivered for high-risk AI: the burden moves toward whoever holds the evidence.

The gap it doesn't close

The catch is scope. The Product Liability Directive exists to protect natural persons, and the damage it compensates is death, personal injury, damage to property kept for private use, and destruction or corruption of data not used for professional purposes. It expressly excludes pure economic loss and property "used exclusively for professional purposes." A consumer whose smart device is bricked by a faulty AI update is covered. A business whose AI agent misfires a transaction, corrupts a business dataset, or ships a bad deployment is not — that is a commercial, economic loss, and it stays exactly where the withdrawn directive would have reached: general national contract and tort law, unharmonised across 27 member states.

That is the case most companies running agents actually have. Agents deployed in a business context, causing business harm, to another business. The one EU proposal aimed squarely at that gap is the one that got withdrawn.

What this means if you run an agent

Two different regimes now apply depending on who got hurt, and neither hands a B2B agent operator's counterparty an easy route:

• Consumer-facing harm — falls under the revised Product Liability Directive from December 2026 for newly placed products. Courts there can already order disclosure and will presume a defect against whoever cannot produce a credible answer.
• Business-to-business, economic harm — the case a coding agent, a data pipeline agent, or an internal automation agent is most likely to produce — stays on national contract and tort law, with no EU-wide presumption to lean on either way.

In both cases, the practical position is the same: nobody is going to take a log file's word for what an agent did. Where the Product Liability Directive applies, the mechanism rewards whoever can actually satisfy a disclosure order with something a court finds credible, not just present. Where it doesn't apply, and the dispute runs on ordinary national rules of evidence, the party who can independently prove what happened is arguing from a position the other side cannot easily contest. A signed execution receipt — a canonical record of inputs, outputs, timing and sandbox policy, signed with a key the agent never touches — is built to be exactly that kind of artefact: something you hand over under a disclosure order, or into evidence, rather than something a claimant successfully argues you should be presumed to be hiding. What such a receipt does and does not establish, and the fields an audit trail needs regardless of which regime ends up applying, are covered in How to verify what an AI agent actually did and AI agent audit trails.

What to do before December 2026

• Map your exposure. Separate agent deployments that touch consumers and personal data from ones that only produce business-to-business, economic outcomes — they now sit under genuinely different legal regimes.
• Don't wait for the transposition deadline. The Product Liability Directive only covers products placed on the market after 9 December 2026; anything shipped earlier, and anything B2B, is already governed by whatever your national law and your contracts say today.
• Build the evidence trail regardless of which regime applies. Disclosure orders and presumptions of defect both cut against whoever cannot produce a credible record. That is true whether a court is applying the new EU rules or your counterparty's home jurisdiction's ordinary contract law.

The receipt format, the verifier and the transparency log are open: the spec, the verifier on PyPI and the public log. Regulation in this area is still being written and rewritten around you. The evidence you can produce when someone finally asks is not something you can backdate.