Key takeaways
- The EU AI Act (Regulation (EU) 2024/1689) entered into force on August 1, 2024. As a regulation, it applies directly in every member state and needs no national transposition law.
- It applies in stages: bans and the AI literacy duty since February 2, 2025, rules for general-purpose AI models since August 2, 2025, and transparency duties since August 2, 2026. Regulation (EU) 2026/1744 moved the high-risk obligations to December 2, 2027 (Annex III) and August 2, 2028 (Annex I).
- It covers far more than developers: deployers that use AI professionally, importers, distributors, and providers from non-EU countries are in scope once an AI system is placed on the EU market or its output is used in the EU.
- Maximum fines are EUR 35 million or 7% of worldwide annual turnover for prohibited practices, EUR 15 million or 3% for most other obligations, and EUR 7.5 million or 1% for supplying incorrect information. SMEs pay whichever amount is lower.
- In Germany, the Federal Network Agency (Bundesnetzagentur) has been the central market surveillance authority, single point of contact, and complaints body since July 29, 2026. BaFin and sector regulators keep their own remits.
What is the EU AI Act?
The EU AI Act is the world's first comprehensive law on artificial intelligence. Its official title is Regulation (EU) 2024/1689. The European Parliament adopted it on March 13, 2024 by 523 votes to 46 with 49 abstentions, the Council approved it on May 21, 2024, and it entered into force on August 1, 2024.
A common misconception concerns the legal form: the AI Act is neither a directive nor a national law. Unlike NIS2, no member state has to transpose it first. Like the GDPR, it applies directly. What Germany did have to settle was which authorities enforce it (see the section on enforcement below). How to get your organization ready in practice is covered in our guide to EU AI Act compliance.
The EU AI Act at a glance
| Feature | Detail |
|---|---|
| Official title | Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (CELEX 32024R1689) |
| In force since | August 1, 2024 |
| Legal form | EU regulation, directly applicable |
| Application | In stages from February 2, 2025 to August 2, 2028 |
| Roles covered | Providers, deployers, importers, distributors, product manufacturers, and authorized representatives |
| Risk tiers | Four: unacceptable, high, limited, minimal |
| Maximum fine | EUR 35 million or 7% of worldwide annual turnover |
| Enforcement | Commission AI Office (GPAI models) and national market surveillance authorities, in Germany the Bundesnetzagentur |
The AI Act does not replace the GDPR. If an AI system processes personal data, both frameworks apply side by side. The data protection basics are summarized in our GDPR guide for businesses.
Who does the EU AI Act apply to?
The EU AI Act applies to providers, deployers, importers, distributors, product manufacturers, and authorized representatives (Art. 2(1)). What matters is not where a company is based but whether an AI system is placed on the EU market or its output is used in the EU. A provider in a non-EU country is therefore covered just like one in Germany.
| Role | What the role means | Example |
|---|---|---|
| Provider | Develops an AI system or model, or has it developed, and places it on the market under its own name | Software vendor with an AI feature in its own product |
| Deployer | Uses an AI system under its own authority in a professional context | Company running AI-assisted applicant screening |
| Importer | Places an AI system from a non-EU provider on the EU market for the first time | EU sales entity for a US vendor's AI product |
| Distributor | Makes an AI system available on the market without being provider or importer | Reseller of a finished AI product |
| Product manufacturer | Places an AI system on the market together with its product under its own name | Maker of a medical device with built-in AI |
Many articles speak of "users", but the regulation uses the term deployer. The distinction matters because the obligations differ sharply. If you use someone else's AI system in your business, you are a deployer. If you put your own name on a high-risk system or substantially modify it, you become a provider under Art. 25.
Excluded are AI systems used exclusively for military, defense, or national security purposes, systems developed solely for scientific research, and purely personal, non-professional use by individuals (Art. 2(3), (6), and (10)). The wording is available in the official legal text.
Which AI practices have been banned since February 2, 2025?
Article 5 bans eight groups of AI practices, and these bans have applied since February 2, 2025. Violations can be fined up to EUR 35 million or 7% of worldwide annual turnover. The banned practices are:
- subliminal or manipulative techniques that materially distort behavior and cause significant harm
- exploiting vulnerabilities based on age, disability, or social or economic situation
- social scoring that leads to unjustified or disproportionate detrimental treatment
- predicting criminal offenses solely from profiling or personality traits
- untargeted scraping of facial images from the internet or CCTV to build facial recognition databases
- emotion recognition in the workplace and in education, except for medical or safety reasons
- biometric categorization that infers sensitive attributes such as political opinions, religion, or sexual orientation
- real-time remote biometric identification in publicly accessible spaces for law enforcement, with narrow exceptions
Not every form of biometric identification is banned. The prohibition targets real-time remote identification for law enforcement, and even there narrow exceptions exist with prior authorization from a judicial or independent authority. Two further bans apply from December 2, 2026: AI systems that generate child sexual abuse material or non-consensual intimate imagery.
Which risk tiers does the EU AI Act define?
The EU AI Act sorts AI systems into four risk tiers: unacceptable, high, limited, and minimal. The higher the risk, the stricter the rules. Many everyday applications, such as spam filters, fall into the lowest tier and carry no specific obligations.
| Risk tier | Rule | Examples | Applies from |
|---|---|---|---|
| Unacceptable | Banned (Art. 5) | Social scoring, untargeted facial image scraping | February 2, 2025 |
| High | Requirements for risk management, data governance, documentation, human oversight, and conformity assessment (Art. 8 to 15) | AI for applicant selection, credit scoring | December 2, 2027 (Annex III), August 2, 2028 (Annex I) |
| Limited | Transparency duties (Art. 50) | Chatbots, deepfakes, AI-generated content | August 2, 2026 |
| Minimal | No specific obligations, voluntary codes of conduct | Spam filters, AI in video games | not applicable |
The transparency duties under Art. 50 require, among other things, that people can tell they are talking to an AI and that deepfakes are disclosed. For machine-readable marking of synthetic content under Art. 50(2), systems already on the market have a grace period until December 2, 2026.
High-risk: the eight areas of Annex III
Annex III lists eight areas in which AI systems count as high-risk unless the Art. 6(3) exemption applies: biometrics, critical infrastructure, education and vocational training, employment and workforce management, access to essential private and public services (such as credit scoring), law enforcement, migration, asylum, and border control, and the administration of justice and democratic processes. AI that serves as a safety component in products covered by Annex I, such as medical devices or toys, is also high-risk. Which systems are affected in a given case, and when the exemption applies, is explained in our post on high-risk AI under the EU AI Act.
A widespread myth is that the high-risk duties only concern providers. Deployers have their own: they must use the system according to its instructions, assign human oversight to competent people, keep the automatically generated logs for at least six months, and inform workers' representatives and affected employees before using the system in the workplace (Art. 26). A company that deploys someone else's applicant management system is not off the hook just because the vendor runs the conformity assessment.
AI literacy under Article 4: a duty for everyone
Article 4 applies to all providers and deployers regardless of risk tier, and it has applied since February 2, 2025. The Omnibus reworded the duty: providers and deployers must take measures to promote AI literacy. The Commission and member states are to support this, in particular for SMEs.
The duty is often read as stricter than it is. The regulation does not require a certificate or a guaranteed competence level for each individual. It requires appropriate measures that take into account people's knowledge, experience, training, and the context of use. A company running sales chatbots therefore needs different measures than one using AI for applicant selection.
What applies to general-purpose AI models?
Provider obligations for general-purpose AI (GPAI) models, such as large language models, have applied since August 2, 2025. Models trained with more than 10^25 floating-point operations are presumed to carry systemic risk and face additional duties (Art. 51(2) and Art. 55). Models placed on the market before August 2, 2025 must comply by August 2, 2027.
Since August 2, 2026, the Commission can fine GPAI model providers up to EUR 15 million or 3% of worldwide annual turnover (Art. 101). If you only connect to someone else's model through an API, you are usually the deployer of the AI system, not the provider of the model. The model obligations then sit with the model provider.
Who enforces the EU AI Act?
At EU level, the Commission's AI Office supervises general-purpose AI models. National market surveillance authorities supervise everything else. Germany settled the responsibilities in its AI Market Surveillance and Innovation Promotion Act (KI-MIG), signed on July 22, 2026 and in force since July 29, 2026.
| Authority | Responsibility under the KI-MIG |
|---|---|
| Federal Network Agency (Bundesnetzagentur) | Default market surveillance authority (Sec. 2(1)), coordination and competence center (Sec. 5), single point of contact (Sec. 6), central complaints body (Sec. 8) |
| AI market surveillance chamber within the Bundesnetzagentur | Certain Annex III systems, including law enforcement, border management, and justice and democratic processes (Sec. 4 and Sec. 2(5)) |
| BaFin | AI systems at the financial institutions it supervises (Sec. 2(3)) |
| Sector regulators under product law | AI in products covered by Annex I Section A (Sec. 2(2)) |
| State authorities | Public bodies of the federal states and certain media providers (Sec. 2(6) and (8)) |
For companies, the Bundesnetzagentur is the first address. It runs an AI service desk aimed mainly at SMEs and start-ups, including an online tool for a first risk classification, and is setting up an AI regulatory sandbox (Sec. 13). The KI-MIG itself provides for fines of up to EUR 50,000 (Sec. 15(3)), for example for breaches of certain cooperation and documentation duties, and federal and state authorities are exempt (Sec. 17(2)). The much higher maximums of the AI Act itself apply in addition.
How high are the fines under the EU AI Act?
The EU AI Act sets three fine tiers, and in each the higher of the two amounts applies. For SMEs, including start-ups, the lower amount applies instead (Art. 99(6)).
| Violation | Maximum | Provision |
|---|---|---|
| Prohibited AI practices | EUR 35 million or 7% of worldwide annual turnover | Art. 99(3) |
| Obligations of providers, deployers, importers, distributors, and notified bodies, including the Art. 50 transparency duties | EUR 15 million or 3% | Art. 99(4) |
| Incorrect, incomplete, or misleading information to authorities | EUR 7.5 million or 1% | Art. 99(5) |
| Violations by providers of general-purpose AI models (fined by the Commission) | EUR 15 million or 3% | Art. 101 |
An illustrative scenario: a company with EUR 400 million in annual turnover breaches a deployer obligation. Three percent of turnover would be EUR 12 million, but because the higher amount applies, the cap is EUR 15 million. At EUR 1 billion in turnover, it would be EUR 30 million instead.
What applies when? The timeline after the Omnibus
The EU AI Act has applied in general since August 2, 2026, but the high-risk obligations only start in December 2027. Regulation (EU) 2026/1744, the Digital Omnibus on AI, entered into force on July 27, 2026 and adjusted the deadlines.
| Date | What applies |
|---|---|
| August 1, 2024 | EU AI Act enters into force |
| February 2, 2025 | Bans under Art. 5 and the AI literacy duty under Art. 4 |
| August 2, 2025 | Obligations for general-purpose AI models, governance structures, and penalty rules |
| July 27, 2026 | Regulation (EU) 2026/1744 (Omnibus) in force |
| July 29, 2026 | German KI-MIG in force, Bundesnetzagentur takes up its roles |
| August 2, 2026 | General application, Art. 50 transparency duties, Commission fining powers over GPAI providers |
| December 2, 2026 | Two new bans (abuse imagery), end of the grace period for Art. 50(2) for systems already on the market |
| August 2, 2027 | Deadline for GPAI models placed on the market before August 2, 2025, and for at least one regulatory sandbox per member state |
| December 2, 2027 | High-risk obligations for Annex III systems |
| August 2, 2028 | High-risk obligations for Annex I systems (AI embedded in products) |
A widespread myth is that the AI Act only applies from 2027. In fact, the bans, AI literacy, GPAI rules, and transparency duties already apply. If you still have August 2, 2026 in mind as the start of the high-risk obligations, that information is out of date: the date was postponed.
What the Omnibus changed beyond the deadlines
The Omnibus did more than move dates. Four changes affect companies directly:
- Small and medium-sized enterprises and small mid-caps get simplified technical documentation, a proportionate quality management system, milder penalties, and priority access to regulatory sandboxes.
- Machinery under Regulation (EU) 2023/1230 is carved out of Annex I. Medical devices and toys stay fully covered.
- Notified bodies go through a single designation procedure, and bodies already designated under sector legislation get a transitional regime.
- The AI Office now also supervises AI systems built on a model from the same provider, as well as AI in very large online platforms and search engines under the Digital Services Act.
How does the EU AI Act relate to the GDPR, ISO 42001, and NIS2?
The EU AI Act complements these frameworks and replaces none of them. Three points of contact matter most for companies.
On data protection, certain deployers must carry out a fundamental rights impact assessment before putting a high-risk system into use (Art. 27). This covers public bodies, private entities providing public services, and deployers of systems for credit scoring and for risk assessment in life and health insurance. If a data protection impact assessment under Art. 35 GDPR already exists, the fundamental rights assessment complements it. Our GDPR guide explains the data protection side in detail.
On management systems, ISO 42001 gives an AI management system its structure of responsibilities, risk assessment, and documentation. The standard is not harmonized, however, so a certificate does not automatically prove conformity with the AI Act. Its main value is that it helps you meet the requirements continuously rather than once. What the AI Act makes mandatory and what ISO 42001 adds as good practice is compared in our article on ISO 42001 and the EU AI Act. Our ISO 42001 guide shows how to build and certify an AI management system.
On security, Art. 15 requires high-risk systems to reach an appropriate level of accuracy, robustness, and cybersecurity. If you already run an ISMS under ISO 27001 or implement the duties of the NIS2 directive, you have a foundation for that. Our post on ISO 27090 explains what the new AI security standard means for an ISO 27001 ISMS.
Where to start, and how Kertos supports you
Start with three steps: list every AI system in your organization, including tools nobody formally approved. Determine your role and the risk tier for each system. Then name one person who keeps track of systems and deadlines. Our guide to EU AI Act compliance walks you through the full checklist.
Most obligations assume you know which AI systems your organization runs. With the Kertos AI Risk Assessment, you record and assess your AI systems, and AIMS maps an AI management system to ISO 42001. For a framework overview, see our page on the EU AI Act. Certified Kertos experts guide you along the way. If you would like to see how your AI systems can be classified, book a demo.
Frequently asked questions
Is the EU AI Act the same as the KI-Verordnung?
Yes. KI-Verordnung is the German name for Regulation (EU) 2024/1689, and EU AI Act is the English one. Legally it is a regulation, not an act of national law, even though it is often called the EU AI law.
Is the EU AI Act already in force?
Yes, it has been in force since August 1, 2024 and applies in stages. Bans and AI literacy have applied since February 2, 2025, GPAI rules since August 2, 2025, and transparency duties since August 2, 2026. High-risk obligations apply from December 2, 2027 (Annex III) and August 2, 2028 (Annex I).
Who enforces the EU AI Act in Germany?
By default, the Bundesnetzagentur, which has acted as market surveillance authority, single point of contact, and complaints body since July 29, 2026. BaFin is responsible for financial institutions, and the relevant sector regulator for AI in products under Annex I Section A.
Does the EU AI Act apply to companies outside the EU?
Yes. Providers and deployers based in a non-EU country are covered if the output of their AI system is used in the EU (Art. 2(1)(c)). Non-EU providers must also appoint an authorized representative in the EU in many cases.
How high are the fines under the EU AI Act?
Up to EUR 35 million or 7% of worldwide annual turnover for prohibited practices, up to EUR 15 million or 3% for most other obligations, and up to EUR 7.5 million or 1% for incorrect information. The higher amount applies, and for SMEs the lower one.
