Skip to content

Provider or operator? Correctly determining roles under the AI Act

“What obligations do we have under the AI Regulation?” Clients ask us this question time and again. To answer it, we must first clarify what role the company plays in relation to its AI tools or use case. The AI Regulation (AI Reg or AI Act) does not base its obligations on ‘AI’ per se, but rather on the system’s risk class (see ‘Risk classification under the AI Regulation’) and the company’s role in the value chain. It is not uncommon for companies to fail to assess their role – and thus their obligations – at all, or to assess it incorrectly, and as a result they overlook important obligations or even attempt to fulfil obligations that do not apply to them at all.

The classic scenario: a company uses a third-party AI tool, for example as a SaaS solution. In such cases, it is generally the operator of that tool. However, if it makes this available to its staff as a white-label solution under its own name (e.g. as a corporate chatbot) or integrates third-party AI models into its own products, the company becomes a provider more quickly than it realises. Why is this the case, and what does it mean in terms of the company’s compliance?

The role model: six actors, two main roles

The AI Regulation groups six roles under the umbrella term ‘actor’, each of which is subject to various obligations. In practice, two are particularly crucial for most companies: the provider, who places an AI system on the market, and the operator, who uses it. The remaining roles mainly concern supply chains.

Three misconceptions we need to clear up:

  1. You can only ‘wear one hat’ – provider without operator!

    No, the roles are not mutually exclusive. For example, anyone who develops an AI system for their own use, or commissions a third party to develop one, is both a provider and an operator.

  2. The role is determined on a company-by-company basis.

    No, the role is determined by the AI system. The same company can be a provider in relation to the customer service bot it makes available to customers, whilst also being the operator of the recruitment tool it has purchased.

  3. A company’s role can be set out in a contract.

    No, the role is determined by the function, not (solely) by the contract. The decisive factors are whose name the AI system bears and who has (actual) economic control. However, contracts shape the function, and with the right contract drafting, it is certainly possible to influence who assumes which role.

Who is not required to comply with the AI Regulation despite using AI:

  • Employees who operate an AI system on behalf of their employer are neither controllers nor processors. Their actions are attributed to the company. It is therefore important for companies to draw up AI guidelines and provide their staff with appropriate training. See also ‘AI Regulation: Do companies have to train their employees?’ on this topic.
  • Private individuals who use AI tools are also not required to comply with the AI Regulation.
  • Anyone who merely uses the results of an AI system operated by a service provider under their own responsibility is generally not subject to the AI Regulation either.

You become a provider as a business more quickly than many realise

A provider is anyone who develops an AI system or commissions its development and places it on the market or puts it into service under their own name or brand, whether for a fee or free of charge. This role largely corresponds to that of a manufacturer under product safety law. The provider bears the bulk of the obligations because they have the greatest influence over the nature and performance of a system – and thus the risk it poses.

This sounds like the major AI players such as OpenAI, Google or Microsoft. But that is a misconception: even ‘ordinary’ companies can quickly become providers of an AI tool.

In our consultancy work, however, we encounter three main scenarios in particular:

1. In-house development for internal use

It is not only those who sell an AI system who are considered providers. ‘Putting into service’ explicitly includes in-house use as well. Anyone who builds an AI tool internally and then makes it available to their staff is therefore both the provider of the AI system and its operator. This generally also applies if the tool is based on a purchased language model that is integrated via an API.

Practical example: The HR department commissions its own IT staff to develop a tool that pre-sorts job applications using a large language model. The company is therefore a provider (in this case, even of a high-risk AI system in accordance with Annex III, No. 4(a) of the AI Regulation). From 2 December 2027, this will entail: risk management, data governance, technical documentation, quality management, conformity assessment, etc. Had it, on the other hand, purchased a ready-made recruitment tool, it would be ‘merely’ an operator, subject to the significantly lighter obligations set out in Article 26.

The same applies if an agency takes on the development. ‘Having it developed’ makes the client the provider, not the service provider. The decisive factor is who sets the requirements and specifications and controls the process.

2. White label: third-party AI under the client’s own brand

Many companies use a chatbot or assistant from a third-party provider, but brand it with their own name and logo. It is not yet entirely clear whether this makes them a provider. Above all, experts are still debating whether simply applying one’s own brand actually fulfils the definition of a provider. One argument in favour of this is that the law has explicitly clarified this case for high-risk AI systems – but not for all others. On the other hand, given the spirit and purpose of the regulations – particularly with regard to the assessment of labelling obligations – it is reasonable to argue that this must also apply, by analogy, to systems not classified as high-risk.

In particular, companies that affix their logo to chatbots or use voicebots in customer service should consider the question of their role.

Rule of thumb: the more a system is presented as a standalone product, the more likely it is that the company is acting as a provider. Anyone using white-label solutions should also ensure, through a contract, that the manufacturer has implemented the machine-readable labelling (or ensure its implementation themselves).

3. Customisations: fine-tuning, RAG, system prompts

Not every customisation turns an operator into a provider. However, the line between the (further) development of a system and its customised use can be blurred:

  • system prompts, integration of a knowledge base (RAG), design of the user interface. Where we are talking about the configuration of a system, this generally constitutes merely a form of use. In some cases, targeted meta-prompts are even regarded as an indication of independent use as an operator. The situation may be different if this results in a new product being offered under the operator’s own name, or if the intended purpose changes.
  • Fine-tuning the system using one’s own data may already constitute ‘development’ if this leads to a significant change to the system. This is clearly the case if the fine-tuning leads to the appropriate use of AI in one of the high-risk areas (including biometrics, recruitment or access to essential services such as insurance or loans).

Rule of thumb: Anyone who merely configures the system is simply an operator. Anyone who, on the other hand, retrains, modifies or repurposes the system should consider the issue of the provider.

The high-risk trap

The consequences are particularly serious for companies if – without realising it – they go from being an operator to a provider of a high-risk AI system.

This happens when they
  1. affix their name or brand to a high-risk AI system that has already been placed on the market,
  2. make substantial changes to such a system whilst it remains a high-risk system, or
  3. change the intended purpose of a system – explicitly including a general-purpose AI system – in such a way that it becomes a high-risk system.

The third scenario is the one that is particularly relevant in practice. If, for example, a company licences a universal AI assistant for office tasks and the HR department begins to use it to systematically assess job applications and rank candidates (thus operating in the high-risk area of human resources management), it must be assumed that the company has thereby altered the intended purpose of the system and is deemed to be a provider.

Here, too, the contractual arrangements between the parties subject to the AI Regulation are relevant. In principle, the original provider must support the new provider with information and technical access so that the latter can fulfil its provider obligations. However, this duty to cooperate does not apply if the original provider has clearly stipulated that its system must not be converted into a high-risk system.

The downstream provider

Anyone who builds their own AI system on top of a GPAI model – such as a specialist assistant based on a large language model – is a so-called downstream provider. This role does not entail any additional obligations of its own. It is a sub-category of the provider: the provider’s obligations regarding their own system remain fully in force. However, integration alone does not make one the provider of the integrated model.

As such, this role actually entails fewer obligations of one’s own than it does specific rights vis-à-vis the model provider: the provision of information and documentation in accordance with Annex XII (including details on the model’s capabilities, limitations and integration), as well as a right to lodge a complaint in the event of breaches by the model provider.

As this information is necessary to fulfil one’s own provider obligations in the first place, it must be included in every contract governing the use of foundation models.

Operator: Anyone who uses AI ‘under their own responsibility’

With the concept of the operator, product safety law for the first time explicitly holds the user accountable, as many risks only become apparent during use. An operator is anyone who uses an AI system under their own responsibility, provided this is not done purely for private purposes. Whether someone is using AI on their own responsibility is determined, above all, by on whose behalf and at whose risk the use takes place.

The operator’s obligations are more manageable than those of the provider, but should not be overlooked.

AI as SaaS: Is the customer really always the operator?

Most companies now use AI as a cloud service. One might think that, in this case, it is not the customer who is the operator of the AI system, but the service provider who actually ‘operates’ the system from a technical perspective. As a rule, however, the customer is the controller of the Software-as-a-Service tool. This is because they decide on the ‘if’, ‘how’ and ‘when’ of its use, on human oversight and on the conclusions drawn from the results. Where the servers are located is irrelevant in this regard. Incidentally, a data processing agreement under Article 28 of the GDPR tends to support rather than contradict the customer’s role as the controller.

Exception: Outsourcing

The situation is different if a service provider takes over a task entirely and uses AI on its own responsibility in the process. If, for example, an agency uses an AI tool to create promotional videos or social media content, it is the controller of the AI tool. The client merely uses the results.

Even in this scenario, companies should take the drafting of contracts seriously! These must address issues of liability, intellectual property rights and indemnities, as well as the transfer of labelling obligations.

Importers, distributors, authorised representatives: the supply chain

If an AI system originates from outside the EU, further roles come into play. Third-country providers must appoint an authorised representative in the EU for high-risk systems and GPAI models (Articles 22 and 54). Importers and distributors must verify that conformity assessment, labelling and documentation are in place. For most companies that ‘merely’ use AI, these roles are of secondary importance. However, it is important to note that distributors and importers can also become suppliers under Article 25. Furthermore, the AI Regulation covers suppliers and operators from third countries as soon as the output of their system is used within the EU.

Tasks for businesses

✓ Add a role column to the AI inventory and check and document the company’s own role(s) for each system.

✓ Identify in-house developments, paying particular attention to API-based in-house tools and agency projects.

✓ Review branding and, when using your own name in connection with third-party AI, clarify and contractually ensure compliance with transparency obligations in particular.

✓ Control the intended purpose and, for example, ensure that usage guidelines for general-purpose tools only permit applications in the areas listed in Annex III (HR, credit, education, etc.) following prior review.

✓ Link changes to an approval procedure. Fine-tuning, new uses and deeper integrations should always be assessed in advance to determine whether they entail a change in role.

✓ Review contracts and, where necessary, tighten them up, as at the very least roles, information obligations, cooperation, liability and indemnities should be regulated in the context of AI. Whilst a contract cannot change a role, it can clearly assign control and responsibility.

Conclusion

The question of what role a company plays in relation to its AI is crucial in view of the associated obligations and the scope of compliance. This must be examined with particular care when using white-label solutions, SaaS and the acquisition of third-party systems.

As a general rule: anyone who uses AI is an operator. Anyone who builds, adapts, repurposes or offers AI under their own name should check whether they are (or have long since been) a provider.