A company buys an artificial intelligence solution, connects it to its data and adapts it to its business. The model belongs to an external supplier. A specialist service provider handles the integration. Decisions about its use remain within the company.
When something goes wrong, each party understands its own part. Yet no one may be able to account for the whole.
This is where the distinction between those who provide AI and those who use it reveals a practical difficulty: some companies make significant interventions in a system without controlling all the components on which it depends.
We argue that this intervention deserves more explicit recognition. Legal classification should make it possible to identify specific responsibilities throughout the chain, including those of organisations that integrate and adapt third-party technology.
1. The space between providing and using
The AI Act distinguishes the provider from the deployer and recognises other operators, such as importers and distributors. However, it does not establish a standalone category for integrators that applies across the framework.
Those who combine components, configure behaviour or adapt a solution must determine where they fit within the existing categories. Provider status is not reserved for those who write the code: it can also cover an organisation that has a system developed and puts it into service under its own name. [1]
The difficulty lies in the distance between these categories and the actual distribution of work. An organisation may control the data and configuration, depend on third parties to test the model and take responsibility towards the client for the final solution.
It is this lack of granularity that we question. Explicit recognition of the integration function would help make its obligations visible. To be useful, that recognition would need to preserve identifiable responsibility for the system's compliance.
2. An adaptation can change a company's position
Consider a company commissioning a tool to support candidate selection. The service provider integrates an external model, the company defines the assessment criteria and the model provider retains control over its development.
Before discussing who is responsible for an inappropriate decision, there are more immediate questions: who validated the tool for that purpose? Who approved the criteria? Does a model update require further testing?
For high-risk systems, the framework set out in Article 25 covers situations in which a deployer or another party is deemed to be a provider. These include certain substantial modifications and changes to the intended purpose that make the system high-risk. The article also provides for cooperation and written agreements along the value chain. [2]
Not every configuration change has that effect. Substantial modification has a specific definition, linked to the initial conformity assessment and the impact of the change on the applicable requirements or intended purpose. [1]
The legal position must therefore be reassessed when the intervention changes. A contract signed at the start of a project may cease to describe the operation accurately.
3. What other frameworks help us see
Comparison with other frameworks helps explain how this allocation can become more precise.
ISO/IEC 42001: recognising the functions that exist
The explanatory note to clause 4.1 of ISO/IEC 42001 includes different organisational roles, among them system integrators and data providers. It also recognises that an organisation may perform several roles.
This approach allows each entity's contribution to be described in greater detail. As a management system standard, it does not change legal status under the AI Act. It does, however, offer a useful basis for organising responsibilities that a legal classification alone cannot describe in detail. [3]
GDPR: looking at who makes the decisions
In data protection, classification requires an understanding of who determines the purposes and means of processing. Joint controllership allows for different contributions without requiring equal participation.
The lesson matters for AI governance: contracts should reflect actual decisions. This requires separate analyses. Being a provider does not automatically make an organisation a controller; being a deployer does not mean it is a processor. [4]
Brazil: allocating duties across the lifecycle
The text of Bill 2338/2023 sent by the Senate to the Chamber of Deputies distinguishes between developers, distributors and deployers. It also relates governance measures to the lifecycle stage and each actor's knowledge.
The proposal also provides for a party to become a developer following a substantial modification or a change in purpose. It therefore does not eliminate the question of changing legal status.
Its contribution lies in the attention paid to each party's involvement. This is a bill under consideration, not legislation in force or three successive degrees of responsibility. [5]
4. What must be settled before integration
For a company, this discussion has consequences for procurement and project approval. Four questions should be documented:
In what capacity does each entity participate in this particular system?
Who selects the data, approves configurations and authorises changes?
Who provides the documentation and carries out the necessary tests?
Who investigates, corrects and can suspend use?
A generic clause requiring compliance with the law is unlikely to resolve these questions. Deliverables, access to information and cooperation mechanisms need to be agreed.
Nor is it enough to state that the model belongs to a third party. Where the available information is insufficient to support the intended use, that limitation should inform the decision to procure or proceed.
Responsibility must follow the intervention
We advocate more explicit recognition of the integration function, with duties proportionate to each entity's intervention and clear responsibility for the system's compliance.
Companies can already start making that allocation visible in their contracts and processes. This means knowing who decides, who verifies and who has the means to put things right.
Sources and references
- AI Act, Article 3, points 3, 4, 8 and 23. Definitions of provider, deployer, operator and substantial modification.
- AI Act, Article 25(1), (2) and (4). Responsibilities along the value chain. The obligations must be read alongside the application timetable in Article 113.
- ISO/IEC 42001:2023, clause 4.1, Note 1. Organisational context and roles in the AI ecosystem, with reference to ISO/IEC 22989.
- EDPB, Guidelines 07/2020 on the concepts of controller and processor in the GDPR, version 2.1; GDPR, Article 4, points 7 and 8, and Articles 26 and 28.
- Brazil, Bill 2338/2023, text sent by the Senate to the Chamber of Deputies in March 2025, Articles 4 and 18, paragraphs 4 and 5. Legislative progress, accessed on 4 October 2026.
General information only. This content does not constitute a legal opinion, audit, certification or recommendation applicable without an assessment of the specific circumstances.