The HR department pre-sorts applications with an assistant system, a chatbot answers the first inquiries in customer service, and marketing produces product texts in seconds. None of these decisions was made as a regulatory decision. Since August 2, 2026, they are regulatory decisions nonetheless, because key provisions of the EU AI Act have been applicable and enforceable since then.
Uncertainty among SMEs is high, and it is being amplified by contradictory reports. On the one hand, deadlines have been postponed; on the other, Germany has designated its supervisory authority, and fines can reach up to 35 million euros. Both are true. The question is which part of this affects you.
clience.legal is a digital business law firm for entrepreneurs. We support the legally compliant use of AI and assess which obligations actually follow from the business model. This article sorts the rules by role and risk class and shows what a company that uses AI rather than developing it actually needs to do.
The most important information at a glance
- Role before risk: Most medium-sized companies are deployers, not providers. The strict obligations fall on the provider side.
- Four risk classes: prohibited practices, high-risk systems, systems subject to transparency obligations and applications with minimal risk.
- Applicable since August 2, 2026: Transparency obligations under Article 50 as well as enforcement and sanctions.
- High-risk obligations postponed: The Digital Omnibus shifts the obligations for standalone high-risk systems to December 2, 2027, and for product-embedded systems to August 2, 2028.
- AI literacy has applied since February 2025: Article 4 requires that persons dealing with AI have a sufficient level of understanding.
- Supervision in Germany: The Federal Network Agency (Bundesnetzagentur) is the central market surveillance authority; sectoral authorities such as BaFin remain responsible for their respective areas.
The first key question: provider or deployer?
This distinction determines ninety percent of your workload. A provider is anyone who develops an AI system or has it developed and places it on the market under their own name or trademark. A deployer is anyone who uses an AI system under their own authority.
Anyone who uses a tool purchased on the market is a deployer. The extensive obligations regarding risk management, technical documentation, data governance and conformity assessment fall on the provider. Deployer obligations are considerably leaner and essentially concern use in accordance with the intended purpose, human oversight and informing the persons affected.
Caution is advised if you substantially modify a third-party system, offer it under your own name or use it for a purpose other than the intended one. In these cases, you may slip into the provider role, with all the obligations attached to it. Check this before you embed a solution in your own product.
Why the role question is so often answered incorrectly
In practice, two mistakes occur. The first is the assumption that a mere user is not affected at all. Deployer obligations are leaner, but they exist, and the transparency and AI literacy obligations apply to everyone who uses AI in the company.
The second mistake is the inadvertent shift in roles. Anyone who embeds a purchased model into their own product, offers it under their own name, or uses it for a purpose other than that intended by the manufacturer can be considered a provider. This happens particularly frequently in the case of software companies that integrate an assistance function into their application and then advertise it using their own product name.
It is therefore advisable to answer three questions in writing for every integration: who appears to the customer as the responsible party, is the manufacturer’s intended purpose being changed, and how far do the provider’s contractual assurances extend? The answers belong in the contract with the provider and in your own service description. We describe how AI tools can be integrated sensibly into contract work in AI in contract law and, with a view to the limits, in why contracts should not be drafted with AI tools such as ChatGPT.
The four risk categories
Prohibited practices
Certain applications are prohibited, such as social scoring with detrimental effects, the targeted exploitation of vulnerabilities of certain groups of persons and emotion recognition in the workplace. The last point is of practical relevance for companies, because sentiment analysis systems in call centers or job interviews may be covered by it. Violations trigger the highest range of fines, up to 35 million euros or seven percent of worldwide annual turnover.
High-risk systems
This category covers AI systems that pose a high risk due to their area of use, including recruitment and personnel management, creditworthiness assessment, critical infrastructure and education. For SMEs, the HR side is particularly relevant: systems that pre-screen applications or support promotion decisions typically fall into this category.
The Digital Omnibus has narrowed the definition of high-risk AI. Components that merely serve to support, improve efficiency or control quality are now only considered high-risk if a failure can actually create risks to health or safety. The classification itself therefore becomes more important, because it is no longer made across the board.
Systems subject to transparency obligations
This class affects most medium-sized users. Anyone who uses a system that interacts with natural persons must ensure that it is recognizable that they are interacting with a machine. Artificially generated or manipulated content must be labeled accordingly. These obligations have applied since August 2, 2026. The technical labeling of synthetic content by machine-readable markings has been postponed to December 2, 2026, because the standards for it are still maturing.
Minimal risk
The vast majority of business applications fall into this category, such as spell checkers, scheduling, translation or forecasting in logistics planning. No special obligations under the AI Act arise here. Data protection, trade secret protection, copyright and liability continue to apply regardless.
What the Digital Omnibus changed
The amendment package was published in the Official Journal at the end of July 2026 and has entered into force. Three points are essential for companies.
- Postponed deadlines: The obligations for standalone high-risk systems will only apply from December 2, 2027, and for systems embedded in regulated products from August 2, 2028.
- Narrower definition of high risk: Several systems that would have previously been considered high-risk are no longer automatically considered high-risk.
- Relief for smaller companies: Documentation and verification requirements were specifically reduced.
A postponement, however, is not a repeal. On December 2, 2027, the full catalog of obligations originally planned for August 2026 is expected to apply. Those who do nothing now will have used up the time gained instead of using it.
The obligation that already applies: AI literacy
Since February 2, 2025, Article 4 of the AI Act has required providers and deployers to take measures to ensure that their staff and other persons dealing with the operation of AI systems have a sufficient level of AI literacy. The scope depends on prior knowledge, the context of use and the groups of persons affected.
In practice, this means documented training tailored to the role. Someone who works with a chatbot in customer service needs different knowledge than the specialist department that selects a system. Record participants, content and date, because if in doubt, compliance with this obligation can only be proven through documentation. The content framework for this overlaps considerably with what we describe in AI and data protection and Liability, copyright and data protection when using AI .
What deployers specifically have to do
The deployer obligations can be summarized in four points that can be implemented without much effort.
First, use in accordance with the intended purpose. A system may only be used for the purpose for which it is intended. This sounds obvious, but it is regularly violated when a tool is used for new tasks in day-to-day business without anyone checking the intended purpose.
Second, human oversight. Results must not feed unchecked into decisions that affect people. This requires a designated person who is able and authorized to review them and who also has the necessary time to do so. A duty to review without the capacity to do so is a fiction.
Third, informing the persons affected. Anyone who interacts with a system or is affected by a decision supported by it should be able to recognize this. In the employment context, the workforce must also be informed.
Fourth, record-keeping. Where the system generates logs, these must be retained insofar as they fall within your area of responsibility. These records are also your proof that oversight actually took place.
Who supervises in Germany
Germany has regulated national implementation with the AI Market Surveillance and Innovation Promotion Act, which the Bundestag passed on June 11, 2026. The Federal Network Agency is the central market surveillance authority. A coordination and competence center has been established there, which pools cooperation with the specialist authorities and operates a service desk for companies that is expressly aimed at small and medium-sized enterprises as well.
Sectoral responsibilities remain in place. BaFin supervises the financial sector, BfArM medical devices, and the data protection supervisory authorities data protection matters. The act does not create any additional substantive obligations; it governs responsibilities, procedures and sanctions. The authorities can request documents, examine systems and require corrective measures.
For companies, the designation of the supervisory authority is more than a formality. It ends the situation in which a regulation applied but no one enforced it. From now on, there is an address for complaints, for example from applicants, customers or competitors, and a procedure that follows. At the same time, this creates for the first time a body to which companies can put questions about classification without triggering proceedings.
Five steps to an inventory
- Take inventory: List all the AI tools used, including the features that were added to existing software later. These shadow applications are the most common blind spot.
- Determine the role: For each system, clarify whether you are a deployer or, exceptionally, a provider, especially when embedding it in your own products.
- Classify: For each system, check the risk classification. Pay special attention to personnel processes and systems that interact with customers.
- Set rules: A usage policy with an approved-tools list, input rules, a duty to review before use and labeling requirements covers most of what is needed.
- Train and document: Conduct and record role-specific training. This fulfills the AI literacy obligation and at the same time lays the foundation for implementing further requirements later.
For most medium-sized companies, this can be done in a few days. It only becomes complex where high-risk applications are in use, and that is exactly why early classification pays off. We describe how this fits into your wider compliance in Compliance and managing director liability and AI: Opportunities and pitfalls.
What the AI Act does not regulate
A common misconception is that the matter is settled once the classification under the AI Act has been made. In fact, the Act leaves several areas of law untouched that are more often relevant in day-to-day business.
- Data protection: For any processing of personal data, there must be a legal basis, a contract with the provider, and clarity about the place of processing. A system with minimal risk can be highly problematic from a data protection perspective.
- Protection of trade secrets: If trade secrets are entered into external systems without appropriate protective measures, legal protection may be lost.
- Copyright: Purely machine-generated content is generally not protected, and at the same time, the output may infringe third-party rights.
- Liability: The company that uses a tool is liable for a faulty result, not the tool.
- Labor law and co-determination: The use of systems that can monitor behavior or performance affects the participation rights of the works council.
These issues can be covered in the same usage policy, which significantly reduces the effort. We cover the data protection aspects in Data protection and compliance in the company.
Related: Legal certainty in the digital enterprise.
Send us a list of the AI tools in use and a brief description of what they are used for. We will determine the role and risk class, identify the obligations that actually apply to you and draft the usage policy your teams can work with. To do so, book a free initial consultation of 15 to 20 minutes. For ongoing support, we are available as external legal department .
FAQs
Does the EU AI Act apply to small businesses as well?
Yes, the regulation does not depend on company size but on role and risk. For small and medium-sized enterprises, however, the obligations are considerably leaner, because they are usually deployers rather than providers. The Digital Omnibus has also introduced relief for documentation and record-keeping. Of practical relevance are above all the transparency obligations and the AI literacy obligation.
What is the difference between a provider and a deployer?
A provider is anyone who develops an AI system or has it developed and places it on the market under their own name or trademark. A deployer is anyone who uses a system under their own authority. The extensive obligations regarding risk management, technical documentation and conformity assessment fall on the provider. Anyone who substantially modifies a third-party system or offers it under their own name may move into the provider role.
What obligations have applied since August 2, 2026?
Since then, the transparency obligations under Article 50 in particular, as well as the provisions on sanctions and enforcement, have been applicable and enforceable. From this point on, national supervisory authorities can request documents, examine systems and demand remedial action. The obligations for general-purpose AI models have already applied since August 2, 2025. The extensive high-risk requirements, by contrast, have been postponed.
Were the high-risk obligations really postponed?
Yes. The Digital Omnibus postpones the requirements for standalone high-risk systems under Annex III to December 2, 2027, and for systems embedded in regulated products under Annex I to August 2, 2028. The content of the obligations remains unchanged; only the date has been postponed. Full implementation is expected by the respective deadline.
Is the use of AI in the recruitment process high-risk?
Systems for selecting and evaluating applications and for supporting promotion and termination decisions are among the areas listed in Annex III. Under the narrower wording introduced by the Digital Omnibus, what matters is the function the system actually performs. Purely auxiliary functions without influence on the decision may fall outside the scope. Regardless of this, data protection law and the prohibition of discrimination apply.
Do I have to label content as AI-generated?
For systems that interact with people, it must be recognizable that they are dealing with a machine. Artificially generated or manipulated content must be labeled, although the machine-readable technical marking has been postponed to December 2, 2026. Irrespective of the AI Act, labeling may be required under competition law if its absence is misleading.
What does the AI literacy obligation mean in practice?
Since February 2, 2025, Article 4 has required that persons involved in the operation and use of AI systems have a sufficient level of understanding. The scope depends on prior knowledge, area of use and the persons affected. In practice, the obligation is met through role-specific training documented by participants, content and date. No specific certification is required.
Who monitors compliance in Germany?
The central market surveillance authority is the Federal Network Agency, where a coordination and competence center with a service desk for companies has been set up. Sectoral responsibilities remain in place, such as BaFin for the financial sector and BfArM for medical devices. The basis is the national implementing act passed by the Bundestag on June 11, 2026. It does not create any additional substantive obligations.
How high can fines be?
The range is set by the AI Act. For prohibited practices, fines of up to 35 million euros or seven percent of worldwide annual turnover are provided for, with lower amounts for other infringements. In each case, the higher amount applies. For small and medium-sized enterprises, the Act provides that company size be taken into account.
When is it worth consulting an attorney about the AI Act?
As soon as AI intervenes in processes that affect people, such as recruitment, customer communication or credit scoring, a review is worthwhile. We classify your systems by role and risk class, draft the usage policy and align it with data protection and trade secret protection. The effort is manageable as long as the classification is made early. It becomes costly when a system is already in production and turns out to be high-risk.
