EU AI Act third-party obligations: what suppliers and customers must agree in writing

The AI Act is written around roles. The provider is the organisation that develops a system and puts it on the market or into service. Around it sit deployers who use the system, importers and distributors who bring it to the EU market, and the third parties who supply the models, tools, data services and components that a system is built from.

Article 25 is the provision that deals with how those roles interact. It is headed responsibilities along the AI value chain, and it does two different things. It says when a party in the chain is treated as the provider of a high-risk system, and it says what the provider and its suppliers must agree between themselves.

Article 25(1) says that any distributor, importer, deployer or other third party shall be considered to be a provider of a high-risk AI system in three situations. The first is where it puts its name or trademark on a high-risk system already on the market or in service. The second is where it makes a substantial modification to a high-risk system and the system remains high-risk. The third is where it changes the intended purpose of an AI system that was not high-risk so that it becomes high-risk under Article 6.

The consequence is that the full set of provider obligations moves to that party. A reseller that puts its own brand on a third-party high-risk system, or a customer that repurposes a general tool for a high-risk use, takes on duties it may not have priced.

The contract question follows directly. Resale, white-label and customisation terms should say who is the provider for the purposes of the Act, who can modify the system, and what happens to responsibility if a customer’s use changes its intended purpose.

Article 25(4) requires the provider of a high-risk AI system and the third party that supplies AI systems, tools, services, components or processes used or integrated in it to specify, by written agreement, the necessary information, capabilities, technical access and other assistance. The purpose is to let the provider meet its own obligations under the Act.

There is a carve-out for third parties that make tools available to the public under a free and open-source licence. The Commission’s AI Act Service Desk notes that general-purpose AI models are excluded from that carve-out.

Article 25(4) also provides that the AI Office may develop and recommend voluntary model terms for these contracts. No such model terms were found published as at 18 October 2026, so the drafting is currently left to the parties.

Article 25(5) adds that the information and access duties do not affect the protection of intellectual property, confidential business information and trade secrets. Both sides should expect that tension to appear in negotiation.

Where a third party becomes the provider under Article 25(1), Article 25(2) says the initial provider shall closely cooperate with the new provider. According to the Commission’s AI Act Service Desk text, that covers making the necessary information available, informing the new provider of known limitations and failure modes, and giving targeted technical access, including for testing and validation. It does not apply where the initial provider clearly excluded its system from being changed into a high-risk one.

For general-purpose AI models, Article 53(1)(b) requires the model provider to prepare, keep up to date and make available documentation to providers of AI systems that intend to integrate the model. The documentation must help them understand the model’s capabilities and limitations and meet their own obligations, and it must include at least the elements listed in Annex XII. Intellectual property and trade secrets stay protected. Free and open-source model providers are exempt from the documentation duty under Article 53(2), but the exemption does not apply to models with systemic risk.

For a downstream builder, these provisions are leverage. A model or component supplier cannot simply refuse to share information the Act requires it to provide, and a contract should say in what form and by when the information arrives.

Article 26 sets out deployer obligations for high-risk systems. In summary, a deployer must use the system according to its instructions for use, assign human oversight to people with the necessary competence, training and authority, ensure input data is relevant and sufficiently representative where it controls that data, monitor operation and report risks and serious incidents, and keep the logs the system generates for at least six months.

Each of those duties depends on the supplier. A deployer cannot follow instructions it has not been given, cannot monitor without the right outputs, and cannot keep logs it cannot access. The practical implication is that a deployer’s supplier contract should secure the instructions for use, the log access and the incident notification route that Article 26 assumes.

This summary reflects the main paragraphs of Article 26. The exact wording of each paragraph, and the duties of authorised representatives, importers and distributors in Articles 22 to 24, should be read in the Regulation itself.

The AI Act entered into force on 1 August 2024. The Commission states that the obligations for general-purpose AI models became applicable on 2 August 2025. The Commission’s FAQ also states that the Digital Omnibus on AI entered into force on 27 July 2026, with rules for stand-alone high-risk systems applying from 2 December 2027 and rules for AI embedded in regulated products applying from 2 August 2028.

Several points should be checked against the Official Journal before relying on this article. The Official Journal reference of the amending regulation was not confirmed. It was also not confirmed whether the Digital Omnibus changes the wording of Article 25 itself. The Service Desk text of Article 99(4) lists Article 25(2) and (4) among the provisions for which administrative fines of up to EUR 15m or 3% of worldwide turnover can apply, but that reading should be confirmed against the Official Journal text.

The timetable gives time to prepare, not a reason to wait. Contracts signed now will still be in force when the high-risk obligations apply, so the written agreement Article 25(4) requires is better drafted in the original contract than added later.

The following is drafting advice, not a statement of what the Act requires of any particular agreement. It should be checked against the consolidated text of the Regulation as amended.

• Role allocation, stating which party is the provider, deployer, importer or distributor for each system, and what happens if rebranding, modification or a change of intended purpose changes that role.
• An Article 25(4) schedule, listing the information, capabilities, technical access and other assistance the supplier will provide, with timescales, formats and a named contact.
• Change and notice terms, requiring the supplier to tell the provider of model updates, new limitations or failure modes, and giving the provider the right to suspend or replace a component that no longer supports its compliance.
• Logs, instructions and incident routes, so a deployer receives the instructions for use, log access and incident notification that its own obligations assume.
• Confidentiality and cost allocation, balancing the Act’s information duties against trade secret protection, and settling who bears the cost of documentation, testing access and regulatory requests.