A chatbot answers customer questions, an AI system produces marketing material, a digital assistant supports employees. Applications like these create concrete transparency duties: in defined cases, people must be able to tell that they are interacting with AI or looking at artificially generated content.
The European AI Act sets out these requirements in particular in Article 50. They apply, as a rule, from 2 August 2026. Which duties a company must meet depends on its role and on the specific use case. This article is based on Regulation (EU) 2024/1689 as currently in force.
The running example is the fictional Nordwerk AG, an industrial group active across Europe in production, sales and customer service. The group wants to introduce a customer chatbot, support employees in finding information, and use AI in its corporate communications.
First clarify: provider or deployer?
Article 3 distinguishes two central roles:
A provider is, in particular, whoever develops an AI system or has it developed and places it on the market or puts it into service under its own name or trademark.
A deployer is whoever uses an AI system under its own authority. A company can take different roles for different applications, and it can fulfil both functions for a single system.
Nordwerk AG, for example, buys a ready-made AI application for internal information search. Its employees use that application inside the company. Nordwerk takes the deployer role.
For customer service, by contrast, Nordwerk has its own AI assistant developed and puts it into service under its own name. In this example the group is also the provider of the system. The fact that the application is based on an external language model does not automatically change that classification.
What Nordwerk should do: Before rollout, record for every application which role the group takes. That classification is the basis for the further responsibilities.
Example 1: An AI chatbot in customer service
Nordwerk introduces a chatbot that helps customers order spare parts. It answers questions, explains products and guides the choice of suitable components.
Article 50(1) requires providers of interactive AI systems to inform the people concerned that they are interacting with AI. An exception applies where this is obvious from the context of use. Under paragraph 5, the information must be clear and accessible at the latest at the first interaction.
What Nordwerk would have to implement: The customer chatbot it provides itself must be designed accordingly. One possible implementation is a notice as soon as the chat opens:
“You are communicating with an AI assistant of Nordwerk AG.”
The information should be built directly into the user interface. The project team should check that it remains perceivable on smartphones, when the display is enlarged, and with assistive technologies.
Recital 132 explains that such notices serve to protect against deception. It expressly takes into account people who are particularly vulnerable because of their age or a disability.
Example 2: An in-house platform for AI-generated content
The marketing department wants to create product images and advertising copy. Nordwerk develops its own generative AI application for this and puts it into service under its own name.
As provider, Nordwerk must under Article 50(2) ensure that synthetic outputs are marked in a machine-readable way and are detectable as artificially generated or manipulated. The provision takes into account, among other things, technical feasibility, the cost of implementation and the generally acknowledged state of the art. Exceptions exist for certain assistive editing functions.
What Nordwerk would have to implement: Technical marking belongs in the development of the platform. The company must provide a suitable solution and assess whether it works.
In practice the project team should also test what happens when content is exported, cropped or converted. Those steps can alter or remove technical marks.
If Nordwerk instead uses a finished external application, the provider duty lies first with the provider of that system. When selecting such a solution, the marking functions on offer should therefore be reviewed specifically.
Example 3: An artificially generated video of the chief executive
For an international product launch, Nordwerk wants a video in which the chief executive appears to speak several languages. Image and voice are generated with AI. The video looks like a real recording.
Article 3(60) defines deepfakes as AI-generated or manipulated image, audio or video content that resembles existing persons or other real points of reference and would falsely appear to be authentic.
For content that meets these conditions, Article 50(4) requires disclosure by the deployer of the AI system used.
What Nordwerk would have to implement: The company must disclose the artificial origin. One possible wording is:
“Image and voice were generated with the help of artificial intelligence.”
The notice should be designed so that the audience can perceive it at first contact. According to the Commission’s explanations, a technical mark that is invisible to people is not enough on its own.
Publication should follow an agreed process between marketing, communications and the responsible approval function. The notice should also be kept in shortened or translated versions.
Example 4: AI writes an article for the company website
Nordwerk has an article drafted on the economic consequences of the energy transition. The piece is meant to inform the public about a matter of public interest.
For such AI-generated or manipulated text, Article 50(4) generally requires disclosure. The provision contains an exception: the content has undergone human review or editorial control, and a natural or legal person carries editorial responsibility.
According to the Commission’s explanation, superficial checks such as a spell-check alone are not enough.
What Nordwerk would have to implement: The group must either make the relevant disclosure or ensure that the conditions of the exception are met.
A workable editorial process could look like this: a competent person checks the statements and sources, corrects errors and assesses completeness. The responsible editorial team then decides on publication. Nordwerk expressly names the person or organisation with editorial responsibility.
Recording the review and the approval is a practical recommendation, so that this process can be organised in a traceable way.
Example 5: Emotion recognition and biometric categorisation
Special information duties apply when systems for emotion recognition or biometric categorisation are used. Article 50(3) requires deployers to inform the people concerned about the operation of the system.
If Nordwerk planned such an application, the group would have to include this duty in the project. At the same time it would have to assess separately whether the specific use is permissible at all. A transparency notice does not replace that assessment.
For project organisation, such functions should therefore be identified early and the responsible specialist functions involved.
Transition period: what applies to existing systems
Article 111(4) contains a limited transitional rule. Providers of generative AI systems that were placed on the market before 2 August 2026 must implement the requirements of Article 50(2) by 2 December 2026.
This deadline concerns the technical marking of generated outputs. It does not postpone the other transparency duties across the board.
What Nordwerk would have to check: If the group placed such an application on the market before the cut-off date, it must assess whether the transitional rule applies. The age of a system does not imply a general exemption.
How Nordwerk can organise implementation
A central inventory of its AI applications is a practical approach for the group. This is an organisational recommendation; Article 50 does not prescribe a particular template.
For each application, Nordwerk should record:
- Ownership: Which business unit runs the application, and who is accountable for providing it?
- Company role: Is Nordwerk provider, deployer, or both?
- Points of contact: Where do people interact with the system or receive its outputs?
- Transparency measures: Which notices, disclosures and technical marks are planned?
- Approval: Who checks implementation and decides on use or publication?
The company should then test the actual flows: entering the customer chat, exporting an image file, displaying the executive video, and approving a specialist article.
That makes the requirements checkable inside the concrete applications. The results can be built into the existing processes for software development, procurement and corporate communications.
Violations can have serious economic consequences
Article 99(4)(g) provides for fines of up to EUR 15 million or up to 3 percent of worldwide annual turnover in the preceding financial year for breaches of the transparency duties. For undertakings, the higher of the two caps generally applies.
For small and medium-sized enterprises, including start-ups, paragraph 6 provides that the lower of the two caps applies. Whether a sanction is imposed, and at what level, depends on the circumstances of the individual case.
Transparency as part of introducing AI
The Nordwerk AG example shows how different the tasks are: the customer chat is about a recognisable AI interaction, a generative platform is about technical marks, the artificial executive video is about disclosure, and the specialist article is about the publication process.
Companies should therefore answer these questions when they plan their applications. Clear ownership and tested processes are the basis for using AI in day-to-day business in a controlled way.
Status: 4 October 2026. Legal basis: Regulation (EU) 2024/1689, in particular Articles 3, 50, 99, 111 and 113 and recital 132. Nordwerk AG and all projects described are fictional.