Key points
Directly applicable
Regulation (EU) 2024/1689 applies directly in all EU member states. No national transposition act is needed; member states set up authorities and penalty rules.
Since 2 February 2025
Prohibited AI practices are banned. Providers and deployers must take measures on the AI literacy of their staff.
Since 2 August 2026
General date of application. The transparency obligations for chatbots and certain AI content apply, and national authorities enforce the rules.
High-risk later
The obligations for high-risk systems apply from 2 December 2027 (Annex III) and from 2 August 2028 (Annex I).
The intended purpose decides
Which obligations apply depends on the intended purpose of the AI system and on your role. Small companies receive some relief, but there is no general exemption for them.
Relevant for Switzerland too
Swiss companies are covered when they offer AI systems in the EU or when the output of their systems is used in the EU.
What the EU AI Act is
The AI Act follows the model of European product safety law. It sets the conditions under which AI systems may be placed on the market, put into service and used in the EU. The approach is risk-based: the greater the risk of a use case for health, safety or fundamental rights, the stricter the requirements. Many everyday office applications fall into the lowest tier.
Key terms
- AI system (Art. 3(1)): a machine-based system that operates with some degree of autonomy and infers from its input how to generate outputs such as predictions, content, recommendations or decisions. Software that only executes rules defined by people is generally not covered.
- Provider: anyone who develops an AI system, or has one developed, and places it on the market or puts it into service under their own name or trademark, whether for payment or free of charge.
- Deployer: anyone who uses an AI system under their own authority. Purely personal, non-professional use is excluded.
- Placing on the market: making an AI system available on the EU market for the first time. Putting into service also covers first use for one's own purposes.
- General-purpose AI model (GPAI): a foundation model, for example a large language model, that can be used for a wide range of tasks.
Who is covered (Art. 2(1))
- Providers that place AI systems on the market or put them into service in the EU, wherever they are established.
- Deployers established or located in the EU.
- Providers and deployers in third countries where the output of their AI systems is used in the EU.
Exemptions include purely personal use, scientific research and development, and systems used exclusively for military purposes. Most small and medium-sized companies are deployers, because they buy AI applications and use them in their own operations.
Timeline: what applies from when
The regulation applies in stages. The table reflects the changes made by Regulation (EU) 2026/1744.
| Date | What applies | Legal basis | Who is affected |
|---|---|---|---|
| 1 August 2024 | Regulation (EU) 2024/1689 enters into force | Art. 113 | everyone |
| 2 February 2025 | Prohibited practices and AI literacy; Commission guidelines on prohibited practices of 4 February 2025 | Art. 5, Art. 4 | providers and deployers |
| 2 August 2025 | Obligations for providers of GPAI models, governance, member state penalty rules | Art. 53 et seq., Art. 99 | model providers, member states |
| 27 July 2026 | Amending Regulation (EU) 2026/1744 (Digital Omnibus on AI) enters into force | Reg. (EU) 2026/1744 | everyone |
| 2 August 2026 | General date of application; transparency obligations; supervision and enforcement by national market surveillance authorities; Commission enforcement powers over GPAI providers; Commission guidelines on Art. 50 of 20 July 2026 | Art. 50, Art. 113 | providers and deployers |
| 2 December 2026 | Machine-readable marking of synthetic content for systems placed on the market before 2 August 2026; new prohibitions on non-consensual intimate imagery and child sexual abuse material | Art. 50(2), Art. 111(4), Art. 5(1)(ba) and (bb) | providers and deployers |
| 2 August 2027 | At least one AI regulatory sandbox per member state; GPAI models placed on the market before 2 August 2025 must comply | Art. 57, Art. 111 | member states, model providers |
| 2 December 2027 | Obligations for high-risk systems under Annex III, for example in recruitment or credit scoring | Art. 6(2), Art. 113 | providers and deployers |
| 2 August 2028 | Obligations for high-risk systems under Annex I, meaning AI as a safety component of regulated products such as medical devices or toys | Art. 6(1), Art. 113 | providers and deployers |
In force or applicable?
“In force” means the regulation exists as a legal act. “Applies” or “applicable” means a specific obligation must be met from that date and can be enforced. The AI Act has been in force since 1 August 2024; its obligations have become applicable step by step since then.
Digital Omnibus: what changed in summer 2026
On 19 November 2025 the Commission proposed targeted amendments to the AI Act as part of its simplification agenda. Parliament and Council reached a political agreement in May 2026. Regulation (EU) 2026/1744 of 8 July 2026 was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026, a few days before the general date of application.
The main changes
- High-risk deadlines postponed: to 2 December 2027 for Annex III systems and to 2 August 2028 for Annex I systems. The stated reason is the delayed availability of standards, guidance and national authorities.
- AI literacy reworded: Art. 4 now requires providers and deployers to take measures to support the development of AI literacy. They do not have to guarantee any specific level of AI literacy of any individual.
- Transition period for marking: providers of generative systems placed on the market before 2 August 2026 have until 2 December 2026 to apply machine-readable marking under Art. 50(2).
- New prohibitions: from 2 December 2026, AI systems that generate realistic intimate or sexually explicit images of identifiable people without their explicit consent are prohibited, as are systems that generate child sexual abuse material.
- Small mid-caps: several relief measures that used to be reserved for SMEs now also cover small mid-cap enterprises, such as simplified technical documentation and lower fine ceilings. Under Recommendation (EU) 2025/1099, these are companies that are no longer SMEs, employ fewer than 750 people and have an annual turnover of up to EUR 150 million or a balance sheet total of up to EUR 129 million.
- Regulatory sandboxes: every member state must have at least one AI regulatory sandbox operational by 2 August 2027.
The prohibitions, the AI literacy obligation, the transparency obligations for new systems and the penalties remain unchanged. Anyone who expected a general postponement should therefore review where they stand.
The four risk classes, explained simply
The regulation classifies AI applications by their risk. Each tier comes with different rules.
Unacceptable risk
Practices considered a clear threat to safety and fundamental rights. Examples include social scoring, emotion recognition in the workplace and untargeted scraping of facial images from the internet. Obligation: such systems may neither be offered nor used.
High risk
Systems in sensitive areas listed in Annex III, such as screening job applications, evaluating employee performance, assessing creditworthiness, or risk assessment and pricing in life and health insurance. AI used as a safety component of regulated products under Annex I is also included. Obligation: providers need, among other things, risk management, suitable training data, technical documentation and a conformity assessment. Deployers have the obligations set out in Art. 26.
Transparency risk (Art. 50)
Systems that interact directly with people or generate content: chatbots, voice and phone assistants, image and text generators, deepfakes. Obligation: people must be able to tell that they are interacting with AI or that content was artificially generated.
Minimal risk
Spam filters, spell checkers, AI features in standard software, internal research aids. Obligation: the regulation sets no specific obligations. AI literacy under Art. 4 and data protection law still apply.
A system can touch several tiers at once. A chatbot that pre-sorts job applications is subject to the transparency obligations and is a high-risk system at the same time. What counts is always the concrete intended purpose.
Prohibited practices
Art. 5(1) prohibits placing on the market, putting into service and using AI systems for the following purposes:
- subliminal, manipulative or deceptive techniques that materially distort people's behaviour and can cause significant harm (point a);
- exploiting vulnerabilities due to age, disability or a specific social or economic situation (point b);
- from 2 December 2026: generating realistic intimate or sexually explicit images of identifiable people without their explicit consent (point ba) and generating child sexual abuse material (point bb);
- social scoring that leads to unjustified or disproportionate detrimental treatment (point c);
- predicting the risk that a person will commit a criminal offence based solely on profiling or personality traits (point d);
- untargeted scraping of facial images from the internet or CCTV footage to create or expand facial recognition databases (point e);
- emotion recognition in the workplace and in education institutions, except for medical or safety reasons (point f);
- biometric categorisation to infer sensitive characteristics such as race, political opinions, religious beliefs or sexual orientation (point g);
- real-time remote biometric identification in publicly accessible spaces for law enforcement, with narrow exceptions (point h).
In day-to-day business, emotion recognition in the workplace deserves the most attention, for example automated mood analysis of employees in video calls or in customer service. Collecting facial images and sales techniques that deliberately exploit people's vulnerabilities can also fall under the prohibitions.
Infringements of the prohibitions can be fined up to EUR 35 million or 7 % of worldwide annual turnover. The Commission published guidelines with explanations and examples on 4 February 2025.
Obligations by role: provider or deployer?
Your obligations depend on the role your company plays for an AI system and on the risk class of that system. The table shows typical situations.
| Situation in the company | Role | Typical obligations |
|---|---|---|
| Employees use ChatGPT, Copilot or similar tools | Deployer | Art. 4; Art. 50(4) if you publish deepfakes or texts on matters of public interest without editorial control |
| A purchased chatbot answers questions on your website | Deployer; the vendor is the provider | Art. 4; the provider must build in the notice that users are interacting with AI (Art. 50(1)). Make sure it stays visible in your setup. |
| You have an AI application developed and offer it to customers under your name | Provider | depending on the risk class; at least Art. 4 and, where it interacts with people, Art. 50 |
| You have an internal AI tool developed and only use it yourself | Provider (putting into service for own use) and deployer | Art. 4; Art. 50 where it interacts with people; high-risk obligations only for a corresponding purpose |
| You use software that pre-sorts job applications | Deployer of a high-risk system | Art. 4; from 2 December 2027 also Art. 26 |
When a deployer becomes a provider (Art. 25)
A deployer, distributor or importer is considered the provider of a high-risk system, and takes on the provider obligations, if it
- puts its name or trademark on a high-risk system already placed on the market,
- makes a substantial modification to a high-risk system so that it remains high-risk, or
- changes the intended purpose of a system that was not high-risk, including a general-purpose system such as a chat assistant, so that it becomes a high-risk system.
The third case is the most relevant in practice. A company that deliberately uses a general AI assistant to evaluate job applications can become the provider of a high-risk system itself.
Obligations of deployers of high-risk systems (Art. 26)
- use the system in accordance with the provider's instructions for use;
- assign human oversight to people with the necessary competence, training and authority;
- ensure that input data under your control is relevant and sufficiently representative;
- monitor operation, inform the provider and the market surveillance authority of risks and report serious incidents;
- keep automatically generated logs under your control for at least six months;
- inform workers' representatives and affected employees before use in the workplace;
- inform people when an Annex III system makes or assists in decisions about them.
AI literacy (Art. 4): what has been required since 2025
Art. 4 has applied since 2 February 2025 to all providers and deployers, regardless of their size and of the risk class of the systems they use. As amended by Regulation (EU) 2026/1744, the obligation reads in essence: providers and deployers take measures to support the development of AI literacy of their staff and of other persons dealing with AI systems on their behalf. They take into account technical knowledge, experience, education, the context of use and the people on whom the systems are used. No specific level of AI literacy of any individual has to be guaranteed.
The Commission answers several practical questions in its AI literacy Q&A (last updated 27 July 2026). No certificate is required; organisations can keep internal records of training and other measures. Companies whose employees use ChatGPT, for example for advertising copy or translations, are also covered by Art. 4; employees should be informed about specific risks such as hallucinations. Art. 4 does not require an AI officer or a governance board. Supervision lies with the national market surveillance authorities, which have been able to act since 2 August 2026.
Deployers of high-risk systems additionally remain obliged to train the staff responsible for human oversight.
Putting it into practice in four steps
- Inventory: who uses which AI application for what?
- Role-based training: basics for everyone, more depth for people who review outputs, configure systems or base decisions on them.
- Usage policy: which tools are approved, which data may be entered, and who reviews results?
- Documentation: record the date, content and participants of each measure and update it when new tools arrive.
Transparency obligations (Art. 50): chatbots, generated content, deepfakes
The transparency obligations have applied since 2 August 2026. Depending on the case, they fall on the provider or on the deployer. The information must be given clearly and distinctly at the latest at the time of the first interaction or exposure (Art. 50(5)).
| Case | Obligation | Who | From |
|---|---|---|---|
| AI systems that interact directly with people, such as chatbots and phone assistants (para. 1) | People are informed that they are interacting with an AI system, unless this is obvious | Provider | 2 August 2026 |
| Systems that generate synthetic audio, image, video or text content (para. 2) | Mark outputs in a machine-readable format and make them detectable as artificially generated | Provider | 2 August 2026; for systems already on the market, 2 December 2026 |
| Emotion recognition and biometric categorisation, where not prohibited (para. 3) | Inform the people exposed about the operation of the system | Deployer | 2 August 2026 |
| Deepfakes and published AI texts on matters of public interest (para. 4) | Disclose that content was artificially generated or manipulated; for texts the obligation does not apply if they undergo human review or editorial control and someone holds editorial responsibility | Deployer | 2 August 2026 |
The Commission published guidelines on the transparency obligations on 20 July 2026. In addition, there is a Code of Practice on Transparency of AI-Generated Content, which the Commission assessed in July 2026.
Everyday examples
- Website chat: the provider must build in the AI notice. As the deployer, check that it is actually displayed in the widget you embed.
- Phone assistant: callers should learn at the start of the call that they are talking to an AI system, unless this is obvious.
- AI-generated images on social media: machine-readable marking is the tool provider's job. You have to disclose under para. 4 if the image is a deepfake, meaning it resembles real people, places or events and would falsely appear authentic.
- Internal summaries and drafts: they are not published, so the disclosure obligation under para. 4 does not apply. Art. 4 still does.
High-risk AI: when an SME is affected
Annex III lists eight areas: biometrics, critical infrastructure, education and vocational training, employment and workers management, access to essential private and public services, law enforcement, migration and border control, and the administration of justice and democratic processes. Two of them are the most likely to matter for small and medium-sized companies.
Employment and workers management
Systems for placing targeted job advertisements, analysing and filtering applications and evaluating candidates; systems for decisions on promotion or termination, for allocating tasks based on behaviour or personal traits, and for monitoring and evaluating performance and behaviour.
Essential private services
Evaluating the creditworthiness of natural persons, except for detecting financial fraud, and risk assessment and pricing for natural persons in life and health insurance.
Exceptions under Art. 6(3)
An Annex III system is not considered high-risk if it does not pose a significant risk, for example because it only performs a narrow procedural task, improves the result of a previously completed human activity, detects decision-making patterns without replacing human assessment, or performs a preparatory task. Systems that profile natural persons are always considered high-risk. A provider relying on an exception must document its assessment.
Sensible next steps
- Check your AI inventory for applications in these areas, including AI features in HR and finance software.
- Ask vendors how they will meet the requirements by 2 December 2027 and which evidence of conformity they will provide.
- Plan the deployer obligations under Art. 26: responsibility for human oversight, log retention, and informing employees and workers' representatives.
GPAI: what SMEs need to know about foundation models
The obligations for providers of general-purpose AI models (Art. 53 et seq.) have applied since 2 August 2025. They include technical documentation, information for downstream providers, a copyright compliance policy and a summary of the training content. Models with systemic risk have additional obligations. These obligations fall on the developers of the models, such as OpenAI, Google or Mistral. Companies that use such models through an app or an API do not have to meet them.
A company that integrates a foundation model into its own application and offers it under its own name, or puts it into service itself, becomes the provider of that application. The obligations for the model itself stay with the model provider. The Commission has published guidelines on the scope of the GPAI obligations and a voluntary General-Purpose AI Code of Practice. Since 2 August 2026 it can also enforce the obligations of model providers with fines.
Penalties
| Infringement | Maximum fine |
|---|---|
| Prohibited practices (Art. 5) | EUR 35 million or 7 % of worldwide annual turnover |
| Infringements of other obligations, including those of providers (Art. 16), deployers (Art. 26) and the transparency obligations (Art. 50) | EUR 15 million or 3 % |
| Incorrect, incomplete or misleading information to authorities or notified bodies | EUR 7.5 million or 1 % |
As a rule, the higher amount applies. For SMEs, including start-ups, the lower amount applies at all three tiers (Art. 99(6)). Since the amending regulation, the lower amount also applies to small mid-caps at the second and third tier (Art. 99(6a)). Member states lay down the penalty rules and take into account the interests and economic viability of SMEs and small mid-caps.
Fines are not the only risk. Infringements can also have contractual and liability consequences or lead to data protection proceedings.
Germany, Austria, Switzerland: who is responsible and what else applies
Federal Network Agency as the central body
The German act on market surveillance and innovation for artificial intelligence (KI-MIG) entered into force on 29 July 2026. It makes the Federal Network Agency (Bundesnetzagentur) the market surveillance authority, single point of contact and complaints body for the AI Act. Among other things, it supervises AI systems in sensitive areas such as human resources and, within its remit, the transparency obligations and prohibitions.
In sectors that are already regulated, the established supervisory authorities remain responsible, for example BaFin in financial services or the state media authorities in the media sector. A coordination and competence centre at the Federal Network Agency supports the authorities. Companies can find information and an online tool for a first assessment of the risk class at the agency's AI Service Desk. An AI regulatory sandbox is intended in particular to let SMEs test new systems with regulatory guidance.
AI Service Desk at RTR
In Austria, the AI Service Desk (KI-Servicestelle) at the regulator RTR has been the central information point since 2024, with FAQs and guidance. According to RTR, the national competences under the AI Act have not yet been implemented in law; this applies to the market surveillance authorities as well as to the notifying authority (research as of 15 September 2026). The Austrian Economic Chamber (WKO) also provides information on obligations and deadlines. The obligations under the regulation still apply directly in Austria.
No AI act of its own, but EU links
Switzerland has no overarching AI law so far. On 12 February 2025 the Federal Council decided to ratify the Council of Europe's AI Convention and to make the necessary legal adjustments on a sector-specific basis wherever possible. A draft for consultation is due by the end of 2026, covering in particular transparency, data protection, non-discrimination and supervision.
The Data Protection Act already applies today: according to the Federal Data Protection and Information Commissioner (FDPIC, 9 November 2023), the FADP applies directly to data processing using AI. Manufacturers, providers and users must, among other things, be transparent about the purpose, functioning and data sources, and users have a right to know whether they are communicating with a machine.
When the EU AI Act applies to Swiss companies
The AI Act covers Swiss companies if they place AI systems on the market or put them into service in the EU (Art. 2(1)(a)) or if the output of their AI systems is used in the EU (Art. 2(1)(c)). Examples:
- a chatbot that advises customers in Germany;
- software that pre-screens applications for positions at a branch in Austria;
- an AI application that a Swiss software company sells to businesses in the EU.
Purely domestic use with no link to the EU is not covered by the regulation. It is governed by the FADP and Switzerland's sector-specific rules.
Practical checklist for SMEs
These ten points help you review where you stand. The ticks are only for orientation on this page and are not saved.
Frequently asked questions
When does the EU AI Act apply?
Regulation (EU) 2024/1689 has been in force since 1 August 2024. The prohibitions and AI literacy have applied since 2 February 2025, and the obligations for providers of GPAI models since 2 August 2025. The general date of application, including the transparency obligations, was 2 August 2026. The high-risk obligations apply from 2 December 2027 (Annex III) and from 2 August 2028 (Annex I).
Does the EU AI Act apply to small businesses?
Yes. The obligations depend on the intended purpose of the AI system and on the company's role; company size does not change that. SMEs and small mid-caps receive relief, such as simplified technical documentation and lower fine ceilings. There is no general exemption.
Does the AI Act affect us if we only use ChatGPT or Copilot?
Yes, your company is then a deployer. At a minimum, Art. 4 applies: you must take measures on your employees' AI literacy, according to the Commission for example on risks such as hallucinations. If you use these tools to publish deepfakes or texts on matters of public interest without editorial control, the disclosure obligation under Art. 50(4) also applies.
What does the AI literacy obligation under Art. 4 require?
Providers and deployers must take measures that support the AI literacy of their staff and of people acting on their behalf, tailored to knowledge, role and context of use. No specific level has to be guaranteed, and certificates are not required. Internal records of training are advisable. National authorities have supervised the obligation since 2 August 2026.
Do I have to label AI-generated content?
That depends on your role. Providers of generative AI must mark audio, image, video and text outputs in a machine-readable way (Art. 50(2)); for systems placed on the market before 2 August 2026, by 2 December 2026. Deployers must disclose deepfakes and published AI texts on matters of public interest (Art. 50(4)). Chatbots must make clear that they are AI. Internal, unpublished content is not subject to the disclosure obligation.
What did the Digital Omnibus postpone, and what not?
It postponed the high-risk obligations to 2 December 2027 and 2 August 2028, the marking obligation for generative systems placed on the market before 2 August 2026 to 2 December 2026, and the deadline for national AI regulatory sandboxes to 2 August 2027. The prohibitions, AI literacy, the transparency obligations for new systems and the penalties apply unchanged.
Does the EU AI Act apply to Swiss companies?
Yes, if they place AI systems on the market or put them into service in the EU, or if the output of their AI systems is used in the EU (Art. 2(1)(a) and (c)). Switzerland has no AI act of its own so far. According to the FDPIC, the Federal Act on Data Protection applies directly to data processing using AI.
What are the penalties?
Up to EUR 35 million or 7 % of worldwide annual turnover for prohibited practices, up to EUR 15 million or 3 % for infringements of other obligations such as those of deployers or the transparency obligations, and up to EUR 7.5 million or 1 % for incorrect information to authorities. As a rule the higher amount applies, for SMEs the lower. For small mid-caps the lower amount applies at the two lower tiers.
What is a high-risk AI system?
An AI system used in one of the areas listed in Annex III, such as screening job applications, evaluating employee performance or assessing creditworthiness, or used as a safety component of a product under Annex I, for example in medical devices. Extensive obligations for providers and deployers apply to these systems from 2 December 2027 and 2 August 2028 respectively.