The EU AI Pact: What Signatories Actually Commit To

The AI Pact is a Commission initiative asking companies to start complying with the AI Act before they have to. It is worth understanding precisely, because it is easy to overstate in both directions: it is not a regulatory approval, and it is also not nothing.

Two pillars, one of which is the pledge

Regulation (EU) 2024/1689 entered into force on 1 August 2024 and applies in tranches under Article 113, with the bulk of the high-risk regime arriving in August 2026 and 2027. That gap between force and application is the gap the Pact exists to fill. The Commission launched it through the AI Office and held the first signing event on 25 September 2024. The current description and the live list of signatories are on the Commission’s AI Pact page, which is the only citation worth relying on for who has signed, because the list changes.

The Pact has two pillars and they are frequently conflated. The first is a network: webinars, workshops and Commission-run sessions explaining the Act, open to organisations that want to follow along without signing anything. Participating in that pillar involves no commitment at all. The second pillar is the voluntary pledge, and only the second pillar makes you a signatory.

The core commitments

The pledge is built around a small set of core commitments that every signatory takes on, plus optional additional pledges a company can choose to add. The core set is:

• An AI governance strategy. Adopt an internal governance approach that promotes the uptake of AI in the organisation and works towards future compliance with the Act. In practice this is the ownership question: who in the organisation is accountable for AI systems, and through what process.
• A mapping of high-risk use. Carry out, to the extent feasible, an inventory of AI systems the organisation provides or deploys in areas that would be considered high-risk under the Act. This is the commitment with real work behind it, because it forces the Annex III classification exercise early — and that exercise is the one most organisations discover they cannot do without first building an inventory they did not have.
• AI literacy and awareness. Promote awareness and literacy of AI among staff, so that people developing and using these systems do so with an understanding of what they are. This one has an unusual property: it shadows a duty that is already binding. The AI literacy obligation in Article 4 has applied since 2 February 2025, so a signatory pledging this in 2026 is pledging to do something it is already required to do.

Signatories may then add further pledges from a published set: labelling AI-generated content, ensuring human oversight of specified systems, mitigating identified risks in particular deployments, and similar. These are individually selected and individually published, so the public document for one company can look substantially more demanding than another’s.

What signing does and does not change

Legally, the Pact creates no obligation under the AI Act and no relief from one. There is no provision of Regulation 2024/1689 that mentions it, there is no presumption of conformity attached to it, and there is no article of the Regulation whose deadline moves for a signatory. Article 113’s dates apply identically to signatories and non-signatories. If you provide a general-purpose AI model, the Chapter V obligations arrived on 2 August 2025 whether or not you signed, and the Commission’s power to fine you under Article 101 arrives on 2 August 2026 the same way.

The Pact is also not the General-Purpose AI Code of Practice, and the distinction is the single most useful thing on this page. The Code of Practice has a statutory hook: Article 56 provides for its drawing-up, and Article 53(4) provides that a provider that adheres to an approved code of practice may rely on it to demonstrate compliance with the Article 53 obligations until a harmonised standard exists. That is a defined evidentiary route recognised by the Regulation itself. The Pact has no equivalent. Signing the Pact and adhering to the Code of Practice are entirely separate acts with entirely separate consequences, and the Pact’s branding invites people to treat them as one thing.

What signing does change is presentational and, indirectly, practical. It is a public statement to a regulator you will later be supervised by, made at a time when you did not have to make it, and it is reasonable to expect that a company which publicly pledged to map its high-risk uses will be asked, at some point, to show the map. That cuts both ways: the pledge is not enforceable as such, but a signatory that never did the work has published a document that contradicts its own conduct. Whether a national authority or the Commission would ever use that is untested. It has not been tested because enforcement has not begun.

When it is worth signing

The Pact is worth signing if you were going to do the work anyway and would benefit from an internal deadline. The three core commitments map almost exactly onto the first three tasks any organisation needs to complete before the August 2026 milestone — name an owner, build an inventory, train the people. A public pledge is a good instrument for getting those funded, and it is approximately free.

It is not worth signing as a compliance strategy, and it is actively risky to treat it as one. The failure mode is an organisation that signs, announces, and then arrives at the applicable date with no technical documentation under Article 11, no quality management system under Article 17 and no conformity assessment under Article 43, believing the Pact bought it something. It bought nothing. The pledge is a commitment to prepare; the preparation is the whole of the value.

One practical note on status and dates, because it is the thing most likely to make a summary of the Pact wrong. Signatory lists are published and updated, some prominent companies declined to sign and have said so publicly, and the pledge document itself has been revised. None of that changes any legal position, but it does mean that any page asserting a signatory count or a specific company’s status is asserting something with a shelf life. The Commission’s page is the source; treat everything downstream of it, including this paragraph, as potentially out of date.