Agent Liability Europe
The AI Act and product liability, read for operators

Who answers
when an agent acts.

The deployer answers for how an agent is used, the maker for what it is. We read the AI Act, the Product Liability Directive and national enforcement for the people who run agents.

Where each deadline stands Read the Act in order
AI ActRegulation (EU) 2024/1689 Amended byRegulation (EU) 2026/1744 Product liabilityDirective (EU) 2024/2853

What already applies, and what comes next.

The Digital Omnibus moved the high-risk dates and left the rest of the Act where it stood. The product liability deadline sits outside the Act entirely. Today is marked on the line.

  1. Prohibitions and AI literacyThe Article 5 prohibited practices and the Article 4 literacy duty apply.
  2. General purpose AI modelsProvider obligations under Articles 53 and 55 apply.
  3. The Digital Omnibus in forceRegulation (EU) 2026/1744 moves the high-risk dates.
  4. Article 50 transparencyChatbot disclosure and deepfake labelling apply. Enforcement starts for everything already in application.
  5. Article 50(2) marking, and a new prohibitionSystems already on the market must mark synthetic content, and the ban on CSAM and non-consensual intimate imagery reaches them.
  6. Revised Product Liability DirectiveTransposition deadline for Directive (EU) 2024/2853. Software, AI included, is a product.
  7. Annex III high-risk obligationsArticle 26 deployer duties, the Article 27 FRIA and Article 9 risk management.
  8. Annex I high-risk obligationsAI embedded in products under EU harmonisation legislation.

Every provision, what moved and what did not

Every date here is checked against the institution that set it.

Who answers when an agent acts.

Each answer stands on its own, so it can be read, quoted or checked without the rest of the page.

Who is liable when an AI agent causes harm in the EU?

Usually more than one party, under more than one law. The deployer answers for how the agent is used. The manufacturer answers for what the agent is.

Under Article 26 of Regulation (EU) 2024/1689, a deployer of a high-risk system must use it according to the provider's instructions, assign human oversight to people with the competence and authority to exercise it, check the input data it controls, monitor operation and keep the logs. Directive (EU) 2024/2853 treats software, AI included, as a product. It gives a claimant a route to the defendant's evidence under Article 9 and a rebuttable presumption of defect under Article 10. Member States must transpose it by 9 December 2026.

The proposed AI Liability Directive, which would have added a dedicated fault regime, was withdrawn in 2025. Fault claims run through national law. The liability framework sets out how the chain divides.

Can a deployer pass its AI Act duties to the provider by contract?

No. The Article 26 duties are owed to the supervisor and, through the supervisor, to the people the system affects.

A provider's terms of service cannot remove them, and an indemnity cannot turn a compliance failure into a recoverable commercial loss. What a deployer can and should contract for is information, cooperation and prompt notice of incidents on the provider's side.

When does a deployer become a provider?

When it puts its own name on a high-risk system, modifies it substantially, or changes its intended purpose so that it becomes high-risk. Article 25 then hands it the provider's full obligations.

Prompt layer customisation, retrieval augmentation and fine tuning can each move an operator across that line without anyone deciding that it should. Article 25, read with worked examples.

What did the Digital Omnibus change?

Regulation (EU) 2026/1744 entered into force on 27 July 2026 and moved two things: Annex III to 2 December 2027 and Annex I to 2 August 2028.

It also moved the deadline for national AI regulatory sandboxes to 2 August 2027, shortened the Article 50(2) marking transition for systems already on the market so that it ends on 2 December 2026, and added a prohibition on AI systems that generate child sexual abuse material or intimate imagery of an identifiable person without consent. The Article 5 prohibitions, Article 50 transparency and the obligations on general purpose AI models were not deferred.

Three instruments. Nothing leaves your browser.

Working tools for a deployer's file. Free, with no account, and nothing you type is sent anywhere.

Omnibus tracker

Every AI Act provision the Digital Omnibus touched, its original date, the date that applies now and the source behind it.

Open the tracker

FRIA generator

A structured first draft of the Article 27 fundamental rights impact assessment, one section for each element of Article 27(1).

Draft a FRIA

Readiness scorecard

Twenty five questions across five deployer obligations, scored to a readiness percentage with a breakdown by category.

Take the scorecard

Recent writing from the desk.

Long form analysis of the Act, the Directive and the supervisors who enforce them. Each piece is dated and revised in place when the law moves.

Everything published, newest first

The briefing

When the law moves, you hear it with the source.

A short letter on what changed under the AI Act, product liability and the market for insuring agents, each item dated and cited. Unsubscribe from any issue.

One address, used for the briefing only. Read past issues.

Thank you.

You are on the list. The next issue will reach you at that address.

That did not send. Please write to daniel@futureproofintelligence.com.