Short answer
A deployer or distributor does not become a provider merely because it configures a model, connects it to an application or fine-tunes it. The answer depends first on whether the business is dealing with a general-purpose AI (GPAI) model or an AI system, and then on what it puts on the EU market or into service under its own name.
For a high-risk AI system, Article 25 of the EU Artificial Intelligence Act makes a downstream operator the provider if it rebrands the system, substantially modifies it while it remains high-risk, or changes a previously non-high-risk system's intended purpose so that it becomes high-risk. Integration alone is not on that list. But an integrator that builds and markets a new system under its own name can be its provider under the general definition.
Fine-tuning a GPAI model follows a different analysis. The Commission's non-binding guideline treats more than one-third of the original model's training compute as an indicative criterion for a significant change. That is not the statutory test for modifying a high-risk system.
These distinctions are time-sensitive. GPAI rules have applied since 2 August 2025. Regulation (EU) 2026/1744 postponed Chapter III Sections 1–3—including Article 25—to 2 December 2027 for Annex III systems and 2 August 2028 for Article 6(1)/Annex I systems.
Start by separating the model from the system
The AI Act gives models and systems distinct roles. A GPAI model supports many tasks; a system combines a model with components such as an interface, retrieval layer and workflow. Buying model access does not make the buyer its provider, but supplying a branded application may make the buyer provider of the resulting system while the upstream developer remains provider of the model.
Article 3(3) supplies the baseline:
“‘provider’ means a natural or legal person, public authority, agency or other body that develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark”
—
The actor must also place the model or system on the Union market—or put the system into service—under its own name or trademark. Free supply can qualify. A deployer uses a system under its authority; a distributor makes one available without being provider or importer.
When Article 25 converts a downstream operator
Article 25 is a special rule for high-risk AI systems. It applies to a distributor, importer, deployer or other third party when one of three events occurs:
- The operator puts its name or trademark on a high-risk system already placed on the market or put into service.
- It makes a substantial modification to an existing high-risk system and the modified system remains high-risk.
- It changes the intended purpose of an existing system, including a GPAI system, that was not classified as high-risk, so that it becomes high-risk under Article 6.
The consequence is direct:
“Any distributor, importer, deployer or other third-party shall be considered to be a provider of a high-risk AI system”
—
The new provider assumes Article 16 duties. The initial provider ordinarily ceases to be provider of that specific system and must cooperate. Since 27 July 2026, the amended Article 25(2) specifies that this may require sufficient technical documentation, known limitations and failure modes, and targeted technical access for testing and validation. The handover duty does not apply if the initial provider clearly prohibited changing its system into a high-risk system.
Article 25(a) preserves contractual allocation of obligations. Contracts can organise evidence, testing and access, but private wording does not prevent a regulator from applying the statutory role to the operator whose conduct triggered Article 25.
What counts as a substantial modification?
“Substantial” does not mean merely expensive or technically difficult. Article 3(23) asks whether a post-market change was unplanned in the original conformity assessment and affects high-risk compliance or changes the assessed intended purpose.
The original technical documentation is therefore central. A new model version, fine-tuning, changed decision threshold or data source may be substantial, depending on its effect on the assessed system. The engineering label does not decide the result.
There is also an express safe boundary for planned learning. For a high-risk system that continues to learn, changes to the system and its performance that the original provider predetermined and documented at conformity assessment are not substantial modifications:
“changes to the high-risk AI system and its performance that have been pre-determined by the provider … shall not constitute a substantial modification.”
—
Outside that boundary, a substantial modification requires a new conformity assessment even if the current deployer keeps using the system and never redistributes it. The modification can therefore change both the operator's role and the system's compliance path.
Integration can create one provider role without transferring another
Integration is not automatically a substantial modification. The right question is what the integrator has created and supplied.
If an upstream business makes a GPAI model available in the Union, it ordinarily remains provider of that model. A downstream business that integrates the model into an application and places the application on the market under its own brand may become provider of the resulting AI system. The Act calls such an actor a “downstream provider.” Model obligations and system obligations can coexist along the same chain.
The Commission identifies an exception: if an upstream developer supplies a model outside the Union and unequivocally excludes EU distribution and use, the downstream actor that nevertheless integrates it into an EU-bound system should be treated as model provider too. This is non-binding interpretation.
For high-risk systems, amended Article 25(4) requires written value-chain arrangements between the provider and suppliers of AI systems, AI models, tools, services, components or processes. Integration therefore calls for contractual access to documentation and testing support, even where it does not itself transfer provider status.
When fine-tuning a GPAI model creates a new model provider
The Act acknowledges that GPAI models can be modified or fine-tuned into new models, but it does not state exactly how much modification creates a new provider. The Commission's 2025 guidelines fill that practical gap without changing the legislation.
The Commission says a downstream modifier becomes provider of the modified GPAI model:
“only if the modification leads to a significant change in the model’s generality, capabilities, or systemic risk.”
—
Its indicative criterion is training compute for the modification greater than one-third of the original model's training compute. If that original figure cannot reasonably be known or estimated, the guideline substitutes one-third of 10^25 floating-point operations (FLOP) for an original systemic-risk GPAI model, or one-third of 10^23 FLOP otherwise.
The threshold is indicative and case-specific. It concerns modification of a GPAI model, not Article 25: a smaller change could still alter a high-risk system's intended purpose or compliance.
Where the modifier does become a GPAI provider, recital 109 and the guidelines limit its documentation, copyright-policy and training-summary work to the modification and its new training data. If the original model had systemic risk and the threshold is crossed, the modified model is presumed to have high-impact capabilities, bringing the additional notification and systemic-risk duties into view.
A practical role-change test
Before fine-tuning, rebranding or integrating a model, record answers to these questions:
- Is the object being changed a GPAI model, an AI system, or both?
- Who developed or commissioned it, and under whose name will it reach the Union market or be put into service?
- Is the resulting system high-risk under Article 6, and through which route?
- Does new branding trigger Article 25(1)(a)?
- Was the change predetermined and documented in the original conformity assessment?
- Does it affect Chapter III, Section 2 compliance or change the assessed intended purpose?
- Does a new purpose move a previously non-high-risk system into high-risk use?
- For GPAI fine-tuning, does the change significantly affect generality, capabilities or systemic risk, and what compute was used?
- Do written agreements secure the documentation, testing access and assistance needed for the role each party holds?
- Was the system already on the market before the relevant application date, bringing Article 111's transitional rules into play?
Support the answers with technical documentation, intended-purpose statements, change logs, compute records and value-chain contracts. A marketing label cannot replace factual analysis.
No settled court interpretation yet
As of 29 July 2026, no Court of Justice or General Court decision located in the EU case-law corpus interpreted Article 25 or the GPAI modifier threshold. The distinction therefore rests on the amended enacted wording, the system's conformity evidence and the Commission's expressly non-binding GPAI guidance. Only the Court of Justice can give an authoritative interpretation of the Regulation.
Borderline cases—such as reallocated duties after rebranding or integration that changes both model and system—require separate model-level and system-level conclusions.
Conclusion
Fine-tuning, integration and configuration do not automatically turn every deployer or distributor into a provider. Provider status follows what the actor develops or has developed, how it supplies or commissions it under its own brand, whether the resulting system is high-risk, and whether a later change affects conformity or intended purpose.
For high-risk systems, rebranding and the two modification routes in Article 25 are the decisive conversion rules. For GPAI models, the Commission's significant-change and one-third-compute approach is a useful but non-binding enforcement guide. Keeping those two regimes separate is the key to assigning responsibility accurately.
Sources
This article provides general legal information, not advice for a specific matter. Laws, facts and procedural deadlines may change the result. Follow the linked primary sources and obtain local advice before acting.