When an AI deployer becomes a provider under the EU AI Act

Why the role matters more than the tool

The AI Act distributes its obligations by role, whatever the technology. Deployers of high-risk systems owe real but manageable duties under Article 26: using systems per instructions, human oversight, input quality, monitoring and, for Annex III systems other than the critical-infrastructure systems in point 2, a fundamental-rights impact assessment under Article 27 where the deployer is a public body or a private provider of public services, and for every deployer of the credit-scoring and life and health insurance pricing systems in points 5(b) and (c). Providers owe the full provider programme: risk management, data governance, technical documentation, conformity assessment, registration, post-market monitoring. The difference in cost and liability is large, which is why the reclassification rules in Article 25 belong in commercial decision-making as well as in the compliance annex. The triggers are ordinary commercial acts: putting a name on a product, changing it, using it for a new purpose.

The three triggers in Article 25

Branding. Put your name or trademark on a high-risk AI system already on the market and you are its provider. White-labelling a vendor's tool under your own logo is the trigger. The same result follows without Article 25 where you have a system built for you, or build one yourself, and put it into service under your own name, including for internal use: that is the ordinary provider definition in Article 3(3), and it applies today. Article 25(1)(a) applies without prejudice to contractual arrangements stipulating that the obligations are otherwise allocated, so the work can be reallocated by contract.

What the contract cannot do is keep the original provider in the frame: Article 25(2), as recast by Regulation (EU) 2026/1744, provides that the initial provider shall no longer be considered a provider of that specific AI system, but must closely cooperate with the new provider and make available the necessary information, reasonably expected technical access and other assistance required for compliance, including technical documentation sufficient to assess compliance with the Article 16 requirements, information about known limitations and failure modes, and targeted technical access for testing and validation. That duty does not apply where the initial provider has clearly specified that its system is not to be changed into a high-risk system.

Substantial modification. The definition is technical but the idea is simple: a change the original conformity assessment did not foresee, which affects compliance or the assessed purpose. Retraining on your data, changing decision thresholds, adding functions: each needs the question asked, against the definition, in writing.

Repurposing into high-risk. Take a system that was not high-risk, including a general-purpose system, and give it a high-risk intended purpose, a chatbot repurposed for recruitment screening or a scoring engine pointed at credit decisions, and the resulting high-risk system is yours.

What is live now, and what is not

The timing needs stating with care, because the Digital Omnibus changed it. Article 25 sits in Chapter III, Section 3, and Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026, postponed the application of Chapter III, Sections 1, 2 and 3, that is Articles 6 to 27: the classification rules, the requirements and the operator obligations, including the reclassification rule itself. The dates are 2 December 2027 for systems classified as high-risk under Article 6(2) and Annex III, and 2 August 2028 for those under Article 6(1) and Annex I, which requires both that the system is a safety component of, or is itself, a product covered by Annex I legislation and that the product must undergo third-party conformity assessment (Article 113 as amended). The dates are fixed and are not conditional on harmonised standards becoming available.

Machinery has changed route: Regulation (EU) 2026/1744 moved the machinery regulation, Regulation (EU) 2023/1230, from Section A to Section B of Annex I. AI-enabled machinery keeps its high-risk classification under Article 6(1), but under Article 2(2) only that classification, the real-world testing article and the governance provisions of the AI Act apply to it; the substantive requirements are to be carried into the machinery regulation itself by delegated acts that must apply by 2 August 2028.

For high-risk systems already placed on the market or put into service before those dates, Article 111(2) as amended applies the Regulation only if the system is subsequently subject to significant changes in its design, and systems intended for use by public authorities must comply by 2 August 2030 in any case. A company that rebrands or substantially modifies an existing system after the application date should not assume it inherits that transitional protection.

Two provider tracks are already in force. First, under the Commission's guidelines for general-purpose AI providers of 18 July 2025, a company that fine-tunes or otherwise modifies a GPAI model becomes the provider of the model only where the modification leads to a significant change in the model's generality, capabilities or systemic risk; the guidelines' indicative criterion is training compute for the modification exceeding one third of the original model's. The modified model must also be placed on the Union market, which on the Commission's reading of recital 97 includes integrating it into the firm's own AI system that is made available or put into service, but not purely internal use that is not essential to a product or service for third parties, where natural persons' rights are unaffected and the model carries no systemic risk.

Where that line is crossed, the provider obligations attach to the modification: technical documentation and downstream information limited to what was changed, a copyright policy and a public training-content summary limited to the modification data, an authorised representative for modifiers established outside the Union and, where the original model carried systemic risk, the systemic-risk obligations and notification of the Commission, with the open-source exemptions available only conditionally and never for systemic-risk models. Routine fine-tuning below that threshold does not, by itself, confer provider status.

Second, a company that places a generative system on the market under its own brand owes the machine-readable marking duties in Article 50(2), covered in our Article 50 update from 2 August 2026, or from 2 December 2026 where the system was already on the market before that date (Article 111(4)), regardless of whether the underlying model came from a vendor. The provider question therefore has consequences now, and its high-risk consequences have fixed dates.

Staying a deployer, or becoming a provider deliberately

Both outcomes are legitimate; the risk is reaching one by accident. Staying a deployer takes discipline in three places. Brand discipline: the decision to put your mark on an AI feature is a legal decision, and someone outside marketing must sign it. Modification governance: a gate in the development process that asks, before any retraining or functional change to a bought system, whether the change is substantial in the Article 25 sense, and records the answer. Purpose documentation: written intended-use statements for every deployed system, so that repurposing becomes a visible decision and never a drift.

Article 25(4) also gives you a contractual lever: where a third party supplies an AI system, an AI model, tools, services, components or processes used in or integrated into a high-risk system, the provider and that third party must specify by written agreement the information, capabilities, technical access and other assistance the provider needs to comply. The duty is mutual, and it does not apply to items made available to the public under a free and open-source licence, a carve-out that does not extend to general-purpose AI models; the AI Office may publish voluntary model contractual terms for these agreements (Article 25(4)).

A fourth route sits in Article 25(3) for manufacturers: where a high-risk AI system is a safety component of a product covered by the Annex I Section A legislation and is placed on the market or put into service under the product manufacturer's name or trademark, the manufacturer is the provider. That clause belongs in your procurement standard now, in both directions, as buyer and, if you ever flip, as the provider entitled to your own suppliers' cooperation.

And for companies that choose provider status deliberately, because the white-label margin or the product ambition justifies it, the decision buys a to-do list with a 2 December 2027 deadline for Annex III systems and 2 August 2028 for Annex I systems (for a system already on the market before those dates, the obligations bite once its design changes significantly, and by 2 August 2030 in any event where it is intended for public authorities), which is exactly enough time if the budget starts now. The analysis reaches Swiss companies through the triggers we mapped in our Switzerland update: a Swiss white-labeller serving the EU market flips roles in the same way.

What this means for you

• Inventory your role per system. One organisation can be deployer of ten systems, modifier of two and provider of one; the duties follow each system.
• Make branding an approval step. Your logo on an AI feature can be a reclassification; route it through legal approval as you would a guarantee.
• Gate modifications in writing. A one-page substantial-modification assessment per change records the answer before a regulator asks the question.
• Write purpose statements and keep them current. Repurposing happens by drift; a written intended-use statement makes it a decision.
• Fix the contracts both ways. Vendor cooperation clauses per Article 25, and clarity on who carries which role, before the high-risk deadlines arrive.

Isak Legal advises companies on AI Act role analysis, white-label and modification strategies, vendor contracts and AI governance across the EU and Switzerland. See our AI, Data & Technology page, or get in touch.

This client update is provided for general information only, reflects the position as at 27 September 2026 and does not constitute legal advice or create a lawyer-client relationship. The high-risk timelines reflect Regulation (EU) 2026/1744, the Digital Omnibus on AI, published in the Official Journal on 24 July 2026 and in force since 27 July 2026; the definition of substantial modification is in Article 3(23), outside Article 25, and the conditions under which a party that fine-tunes a general-purpose model becomes its provider are set out in the Commission's guidelines, which are not binding, while the obligations that then apply come from Articles 53 to 55 of the Act itself. The current wording should be checked against the consolidated text of Regulation (EU) 2024/1689. For advice on a specific matter, please contact us.