The four risk classes of the AI Act
The classification is the real core of the Regulation. It is based not on the technology, but on the intended purpose. The same language model can mean minimal risk in one application and high risk in another.
Risk class 1: Prohibited AI practices
Unacceptable risk: These practices are prohibited and have been banned since 2 February 2025. They include social scoring by public and private actors, manipulative techniques that influence behaviour subliminally and cause harm, the untargeted scraping of facial images from the internet to build facial recognition databases, and emotion recognition in the workplace and in education institutions. The last point is the one companies most often brush up against without realising it, for example with analysis tools for job interviews or sales calls.
Risk class 2: High-risk AI systems
High risk: This is where the full set of requirements applies: a risk management system, data quality, technical documentation, record-keeping, transparency towards deployers, human oversight, accuracy and cybersecurity, plus a conformity assessment and CE marking. Two annexes define which systems fall into this class: Annex III lists stand-alone applications in sensitive areas such as recruitment, creditworthiness assessment, education or critical infrastructure; Annex I covers AI embedded as a safety component in products that are already regulated, such as medical devices, lifts or toys. Since the Omnibus, machinery has been listed in Section B of Annex I. For AI in machinery, only certain provisions of the AI Act therefore still apply directly (Article 2(2)). What exactly falls into this class and what evidence it requires is described in detail under High-risk AI system.
Risk class 3: Limited risk and transparency obligations
Limited risk: Systems that interact with people or generate content are subject to transparency obligations under Article 50, but not to a conformity assessment. Chatbots, assistants and generative AI tools typically end up here.
Limited risk: the class where SMEs actually end up
Many typical AI applications in SMEs fall into the two lower classes. A chatbot on the website, AI-assisted generation of product texts, a voice agent on the switchboard: all limited risk. The obligations are manageable. You just need to know them.
Specifically, the following has applied since 2 August 2026: providers must design AI systems that interact directly with people in such a way that users are informed that they are dealing with an AI, unless this is obvious from the context. For a chatbot on your own website, the provider is often the company itself, namely when it develops the bot under its own name or has it developed. Deployers who publish deepfakes must disclose that the content has been artificially generated or manipulated. This means AI-generated or AI-manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to be authentic. A corresponding disclosure obligation applies to AI-generated text published with the purpose of informing the public on matters of public interest. It does not apply if the text has undergone human review and someone holds editorial responsibility for the publication. Anyone who uses emotion recognition or biometric categorisation must inform the people exposed to it. What this looks like day to day and which borderline cases exist is described in the entry AI labelling obligation.
Important when assessing articles in circulation: these obligations have applied since 2 August 2026 and have not been postponed, the chatbot notice included. A transitional period exists only for the providers’ technical marking obligation under Article 50(2), i.e. the machine-readable watermark in the generated content, and only for systems placed on the market before 2 August 2026. This grace period ends on 2 December 2026. Anyone who concludes from this that the transparency obligations as a whole are suspended until December is wrong.
Risk class 4: Minimal risk
Minimal risk: Everything else: spam filters, demand forecasts, recommendation logic in an online shop, predictive maintenance on machine data. The Regulation sets no specific requirements for these. It encourages voluntary codes of conduct, nothing more.
Obligations for providers and deployers
The full list of obligations applies only to high-risk systems. It is still worth looking at both lists, because they show where the burden lies and what actually ends up with you as a deployer.
Providers of high-risk systems must, among other things:
- run a risk management system throughout the entire lifecycle,
- manage training, validation and testing data according to quality criteria (data governance),
- draw up technical documentation before placing the system on the market and keep it up to date,
- provide for automatic recording of events (logging),
- design the system so that effective human oversight is possible,
- ensure an appropriate level of accuracy, robustness and cybersecurity,
- complete the conformity assessment and affix the CE marking.
Deployers of high-risk systems must, among other things:
- use the system in accordance with the provider’s instructions for use,
- assign human oversight to persons qualified for it,
- monitor the operation and inform the provider of incidents and risks,
- keep the automatically generated logs to the extent they are under their control,
- inform workers and their representatives before putting the system into use at the workplace,
- carry out a fundamental rights impact assessment if they are a public body or work in certain areas such as creditworthiness assessment and insurance.
Regardless of the risk class, the AI literacy obligation under Article 4 applies to all providers and deployers, and the transparency obligations under Article 50 apply to systems that interact with people, generate content or recognise emotions.
AI Act deadlines at a glance
The timetable was fundamentally revised once in summer 2026. The following table shows the status after this change.
What the Digital Omnibus changed and what it did not
On 24 July 2026, Regulation (EU) 2026/1744, the so-called Digital Omnibus on AI, was published in the Official Journal of the EU. It has been in force since 27 July 2026 and is therefore applicable law, not a draft and not a proposal. If you come across older articles that discuss the Omnibus in the conditional, they are simply out of date.
It essentially changed three things. First, the high-risk dates: Annex III moved from 2 August 2026 to 2 December 2027, Annex I from 2 August 2027 to 2 August 2028. The reason was missing harmonised standards and conformity assessment bodies that had not yet been designated. Without this infrastructure, companies would practically not have been able to meet the requirements. Second, the wording of Article 4 on AI literacy was softened: instead of taking measures to ensure, “to their best extent”, a sufficient level of AI literacy of their staff, providers and deployers now take measures “to support” the development of this literacy. Nobody has to guarantee a specific level of literacy of individual employees any more. Third, from 2 December 2026 a new prohibition applies that targets the generation of non-consensual intimate material and child sexual abuse material.
There are also smaller changes with practical effect. Providers of generative systems placed on the market before 2 August 2026 have until 2 December 2026 to implement machine-readable marking (new Article 111(4)). Some reliefs for SMEs now also apply to small mid-cap enterprises (SMCs). And each Member State only has to have at least one AI regulatory sandbox operational by 2 August 2027.
Not changed were the remaining prohibitions under Article 5, the date of application of the transparency obligations on 2 August 2026 and the maximum fines. If you based your approach on the status before July 2026, you should above all reschedule your high-risk planning, but not put the implementation of transparency on hold. The European Commission maintains a continuously updated overview of the legal framework at digital-strategy.ec.europa.eu.
This AI Omnibus should not be confused with the general Digital Omnibus, which the Commission proposed on the same day, 19 November 2025, and which concerns the GDPR, the Data Act and cookie rules, among other things. As of 10 October 2026 it is only a proposal: according to the Legislative Observatory, it is awaiting the vote in the responsible committees of the European Parliament, and the Council has not yet adopted a negotiating mandate.
National implementation: KI-MIG and the Bundesnetzagentur
An EU regulation applies directly, but it needs national authority structures for market surveillance and penalties. Germany created this structure with the Act on Market Surveillance and the Promotion of Innovation in Artificial Intelligence (KI-MIG). It is dated 22 July 2026 (Federal Law Gazette BGBl. 2026 I No. 223) and has been in force since 29 July 2026. The text of the Act is available (in German) at gesetze-im-internet.de.
The central body is the Bundesnetzagentur (Federal Network Agency). It is the market surveillance authority, single point of contact and complaints office for the AI Act, as it announced in its press release of 29 July 2026. Germany takes a hybrid approach. In product areas that are already fully harmonised, the established sectoral supervisory authorities remain responsible: BaFin for AI in regulated financial activities, the product market surveillance authorities for embedded AI, and the state authorities for bodies of the federal states. The Bundesnetzagentur coordinates the whole system and is itself the market surveillance authority for product areas at federal level that have not been regulated so far.
The second part of the Act’s name is what matters in practice: the promotion of innovation. The Bundesnetzagentur runs an AI Service Desk with easily accessible information and an online tool for a first assessment of the risk class. If you are unsure how to classify a specific system, you already have a free first point of contact before external advice becomes necessary. An AI regulatory sandbox, in which small and medium-sized enterprises in particular can develop and test systems under regulatory supervision, is something the authority still has to set up under Section 13 KI-MIG. So far there has been a completed pilot project and participation in an EU project, but no regular operation yet, as the Bundesnetzagentur’s sandbox page shows. A national sandbox has to be operational by 2 August 2027.
Fines: how much is really at stake
The penalty framework has three tiers. Infringements of the prohibited practices under Article 5 cost up to €35 million or 7 per cent of worldwide annual turnover. Infringements of other obligations, including the transparency obligations and the high-risk requirements, are subject to fines of up to €15 million or 3 per cent. For supplying incorrect, incomplete or misleading information to authorities, the ceiling is €7.5 million or 1 per cent.
The point that matters most for SMEs is often overlooked: for small and medium-sized enterprises, including start-ups, the lower of the two amounts applies in each case. For a company with €30 million in turnover, a transparency infringement therefore means a theoretical ceiling of €900,000 instead of €15 million. That is no free pass, but it puts the figures into a realistic perspective and refutes the headlines that work with a blanket €35 million. Since the Digital Omnibus, the lower-amount rule also applies to small mid-cap enterprises, but not for prohibited practices.
Example: a mechanical engineering company with 180 employees
Suppose a supplier from the Sauerland region uses three AI applications. First, a chatbot on its website that pre-qualifies enquiries and finds spare part numbers. Second, a language model in marketing that generates product texts and data sheet summaries. Third, a pre-selection tool in recruiting that sorts incoming applications by fit. The tool was bought as SaaS and has been in use for two years.
The classification is clear, the roles are not. The chatbot and text generation are limited risk. If the supplier had the chatbot built by an agency and runs it under its own name, it is the chatbot’s provider and must ensure that the bot identifies itself as an AI, most simply with a sentence in the greeting and a notice in the chat window. For the language model in marketing, it is the deployer. According to the Commission’s guidelines, product texts and data sheet summaries are generally not covered by the disclosure obligation for AI-generated text, as long as they make no claims about, for example, health, consumer safety or sustainability. An internal rule that a responsible person approves generated texts before publication still makes sense. The effort for both is manageable.
The recruiting tool is a different matter. Selecting applicants falls under Annex III and therefore into the high-risk class. The resulting obligations only apply from 2 December 2027, and they primarily affect the SaaS provider. As the deployer, the company remains responsible for human oversight of the selection decisions, following the instructions for use, retaining the logs and informing the applicants concerned. Because the tool was already in use before that date, the high-risk obligations under Article 111(2) also only apply if the system is subsequently subject to significant changes in its design. Whether a SaaS provider’s ongoing releases amount to such a change is a question for the provider. The strategic question is: will the provider present a conformity assessment by December 2027? This question belongs in the contract, not in a compliance checklist, and at the next renewal, not only in 2027.
What this means for your approach
The pragmatic way in is an inventory. List every system that contains AI, including the features that came bundled with existing software and that nobody counted as an AI project. For each system, note the intended purpose, your role, the risk class and the provider. This list is also the framework for documenting AI literacy under Article 4: if you can show who was trained in what and when, and which system it concerns, you have solid evidence for the AI literacy obligation, in both the softened and the original version. How to build a reliable roadmap from it is described in detail in our article EU AI Act for SMEs.
Four common misconceptions
- “The transparency obligations have been postponed to December 2026.” Wrong. They have applied since 2 August 2026. Only the providers’ technical watermark under Art. 50(2) has been postponed, and only for systems that were on the market before that date.
- “We only use tools we bought, so the Regulation doesn’t affect us.” Wrong. Deployer obligations exist regardless of who built the system. And anyone who substantially modifies a system or passes it on under their own name becomes a provider themselves.
- “The software supplier has to provide the AI notice in the chatbot.” Not necessarily. The notice obligation under Article 50(1) applies to the provider. If you have a bot developed and use it under your own name, that is you.
- “Since the Omnibus, AI literacy is off the table.” Wrong. The obligation remains, only the standard is softer. Documented training related to the systems in use remains the simplest evidence.
The sober finding after two years of the AI Act: for most small and medium-sized companies, the effort is lower than the public debate suggests, provided you know your systems and your role. It only gets expensive where a high-risk use case is running in the company unnoticed or where a prohibited practice was sold as a harmless analysis feature. A complete inventory reveals both early on.
Frequently asked questions about the AI Act
Has the AI Act been postponed?
Only in part. The Digital Omnibus, Regulation (EU) 2026/1744, postponed the high-risk obligations to 2 December 2027 (Annex III) and 2 August 2028 (Annex I). Most of the other rules have applied since 2 August 2026, including the transparency obligations under Article 50. The prohibitions and the AI literacy obligation have applied since 2 February 2025.
Does the AI Act also apply to small businesses?
Yes. The Regulation is based on the role (provider or deployer) and the risk class of the system, not on the size of the company. A two-person business with a chatbot on its website is subject to the same transparency obligations as a large corporation. Relief comes elsewhere: in the event of a fine, SMEs pay the lower of the two amounts in each case, they may use simplified forms of technical documentation for high-risk systems, and the Bundesnetzagentur’s AI Service Desk gives them a free first point of contact. Since the Omnibus, some of these reliefs also apply to small mid-cap enterprises. The Bundesnetzagentur’s AI regulatory sandbox is still being set up.
What happens in the event of infringements?
The penalty framework has three tiers, each with maximum amounts: up to €35 million or 7 per cent of worldwide annual turnover for prohibited practices under Article 5, up to €15 million or 3 per cent for infringements of other obligations such as the transparency and high-risk requirements, and up to €7.5 million or 1 per cent for supplying incorrect information to authorities. The higher amount applies in each case, the lower one for SMEs; since the Omnibus, this also applies to small mid-cap enterprises, except for the prohibitions. In Germany, the Bundesnetzagentur imposes the penalties as the market surveillance authority, and in fully harmonised product areas the relevant sectoral supervisory authority does.
Is using ChatGPT in a company a high-risk case?
As a rule, no. If you use a language model to draft texts, summarise documents or do research, you are in the minimal or limited risk category. What remains is the obligation to support the development of AI literacy in your team and to label AI-generated content in the cases covered by Article 50. What matters is the intended purpose, not the tool: if you use the same model to pre-sort job applications or assess creditworthiness, you end up in a use case under Annex III and therefore in the high-risk class.
What is the difference between the AI Act and the GDPR?
The GDPR protects personal data and asks whether and how it may be processed. The AI Act is product safety law and asks what harm an AI system can do to health, safety and fundamental rights. Both apply side by side: a chatbot needs a legal basis and privacy information under the GDPR and the AI notice under Article 50. The roles also differ (controller and processor there, provider and deployer here), as do the competent authorities: the data protection authorities on one side, the Bundesnetzagentur on the other.
