EU AI Act Guide 2026: Risk Classes, GPAI, Deadlines, Fines

Overview: this guide shows what the AI Act regulates, whom it applies to and from when. How to implement the obligations step by step is set out in the EU AI Act Compliance Roadmap. All articles on the topic can be found in the EU AI Act topic area.

1. What is the EU AI Act?

Regulation (EU) 2024/1689 on artificial intelligence (AI Regulation, AI Act) of 13 June 2024 was published in the Official Journal on 12 July 2024. It entered into force on the twentieth day following publication, i.e. on 1 August 2024 (Art. 113 sentence 1). As a regulation it applies directly in Germany and Austria, without an implementing act. The Member States mainly regulate competences and penalties (Art. 70, Art. 99(1)). Since 27 July 2026 it has applied as amended by Regulation (EU) 2026/1744 (Digital Omnibus on AI). That amendment above all shifted the timetable of the high-risk obligations.

The AI Act follows a risk-based approach: the type and content of the obligations depend on the intensity and scope of the risks that an AI system can generate (Recital 26). Unlike the GDPR, the AI Act regulates products and their use along the value chain; the GDPR continues to apply alongside it (Art. 2(7)). Territorially, the regulation is tied to the Union market: it covers providers that place AI systems on the market or put them into service in the Union, irrespective of where they are established. It also covers providers and deployers from third countries where the output of the AI system is used in the Union (Art. 2(1)(a) and (c)).

Core concepts:

• AI system (Art. 3(1)): a machine-based system that is designed to operate with varying levels of autonomy, that may exhibit adaptiveness after deployment and that infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments.
• General-purpose AI model (GPAI, Art. 3(63)): a model that displays significant generality, is capable of competently performing a wide range of distinct tasks and can be integrated into a variety of downstream systems, for example large language models. Models used solely for research, development or prototyping before being placed on the market are excluded.
• Placing on the market (Art. 3(9)): the first making available of an AI system or a GPAI model on the Union market.
• Putting into service (Art. 3(11)): the supply of an AI system for first use directly to the deployer or for own use in the Union for its intended purpose.

2. The 4 risk classes

The regulation expressly governs three categories: prohibited practices, high-risk AI systems and AI systems with specific transparency obligations. Everything else falls into the practical tier “minimal risk”. A system can be subject to several tiers at the same time. A high-risk system that interacts with people must, for example, also comply with Art. 50.

2.1 Prohibited practices (Art. 5)

The prohibited AI practices have applied since 2 February 2025. Infringements fall under the highest tier of fines of up to EUR 35 million or 7 % of worldwide annual turnover (Art. 99(3)). Art. 5(1) prohibits:

• (a) subliminal techniques or purposefully manipulative or deceptive techniques that materially distort behaviour and cause significant harm
• (b) the exploitation of vulnerabilities due to age, disability or a specific social or economic situation
• (ba) the generation or manipulation of realistic image, video or audio content depicting the intimate parts or sexual activities of an identifiable person without that person's explicit consent, and (bb) material within the meaning of Art. 2(c) and (e) of Directive 2011/93/EU; both apply only from 2 December 2026 (inserted by Regulation (EU) 2026/1744, subject to the restrictions of Art. 5(1a))
• (c) social scoring, not limited to public bodies
• (d) risk assessments of whether a person will commit a criminal offence based solely on profiling
• (e) the untargeted scraping of facial images from the internet or CCTV footage for facial recognition databases
• (f) emotion recognition in the workplace and in education institutions, except for medical or safety reasons
• (g) biometric categorisation to infer sensitive characteristics
• (h) real-time remote biometric identification in publicly accessible spaces for the purposes of law enforcement, with narrowly limited exceptions

2.2 High-risk AI (Art. 6 with Annexes I and III)

Two paths lead to high-risk classification:

• Annex I path (Art. 6(1)): the AI system is a safety component of a product covered by the harmonisation legislation listed in Annex I (for example machinery, medical devices, toys), or is itself such a product. In addition, it must be subject to a third-party conformity assessment. Since 27 July 2026, AI used exclusively for aspects not relevant to safety, such as user assistance, performance optimisation, automation or quality control, is not considered a safety component (Art. 6(1a)). This does not apply where its failure would endanger health and safety (paragraph 1b).
• Annex III path (Art. 6(2)): the system falls into one of the eight areas of Annex III: 1. biometrics, 2. critical infrastructure, 3. education and vocational training, 4. employment and workers' management, 5. access to essential private and public services (including creditworthiness, life and health insurance, emergency calls), 6. law enforcement, 7. migration, asylum and border control, 8. administration of justice and democratic processes.

Under Art. 6(3), an Annex III system is by way of exception not considered high-risk if it does not pose a significant risk to health, safety or fundamental rights. This exception never applies where the system performs profiling of natural persons. How to classify a system step by step is shown in the classification workbook for Annex III.

2.3 Transparency obligations (Art. 50)

Art. 50 has applied since the general date of application on 2 August 2026. No conformity assessment is required for it:

• Interaction (paragraph 1): providers ensure that systems interacting directly with people inform them that they are interacting with an AI system.
• Synthetic content (paragraph 2): providers mark generated audio, image, video and text content in a machine-readable format as artificially generated or manipulated. For systems placed on the market before 2 August 2026, a transitional period runs until 2 December 2026 (Art. 111(4)).
• Emotion recognition and biometric categorisation (paragraph 3): deployers inform the persons concerned, in so far as the use is not already prohibited under Art. 5.
• Deep fakes and texts on matters of public interest (paragraph 4): deployers disclose that the content has been artificially generated or manipulated.

In depth: Transparency obligations under Art. 50 and Chatbot labelling in practice.

2.4 Minimal risk

Everything that is neither prohibited nor high-risk and does not trigger a transparency obligation, for example spam filters, recommendation functions in e-commerce or spelling and writing aids with an assistive function. The AI Act sets no specific requirements for the system here. Art. 4 (AI literacy) does, however, apply to providers and deployers. The GDPR, copyright law and sectoral rules apply independently of this.

3. Provider vs. deployer

Most obligations under the AI Act depend on the role. A company can have different roles for different systems, and both roles for the same system.

3.1 Provider (Art. 3(3))

A provider is anyone who develops an AI system or GPAI model or has it developed and places it on the market or puts it into service under its own name or trademark, whether for payment or free of charge. Anyone who puts a self-developed system into service only internally under its own name is also a provider. The regulation knows no criterion of whether the system “leaves the premises”.

Obligations for high-risk AI (Art. 16): meet the requirements of Art. 9 to 15, quality management system (Art. 17), technical documentation under Annex IV (Art. 11), conformity assessment (Art. 43), CE marking, registration in the EU database (Art. 49), post-market monitoring (Art. 72) and reporting of serious incidents (Art. 73).

3.2 Deployer (Art. 3(4))

A deployer is anyone who uses an AI system under its own authority, except in the course of a personal non-professional activity. Anyone who buys a recruiting tool, credit-scoring software or a chatbot and uses it in operations is therefore a deployer.

Obligations for high-risk AI (Art. 26): use in accordance with the instructions for use, human oversight by competent persons, representative input data in so far as they are under the deployer's control, monitoring of operation, retention of the automatically generated logs for at least six months (paragraph 6), information of the workers' representatives and the affected workers before use in the workplace (paragraph 7) and information of affected persons in the case of Annex III decisions (paragraph 11). Certain deployers must additionally carry out a fundamental rights impact assessment under Art. 27 (see Myth 5).

3.3 Watch out: Art. 25 – the role changes

Under Art. 25(1), distributors, importers, deployers or other third parties are considered providers of a high-risk AI system and are subject to Art. 16 if they

• put their name or trademark on a high-risk AI system already placed on the market (point (a)),
• make a substantial modification to a high-risk AI system in such a way that it remains high-risk (point (b)), or
• modify the intended purpose of a system that is not high-risk, including a general-purpose AI system, in such a way that it becomes high-risk (point (c)).

The initial provider is then no longer considered the provider of that system, but must cooperate with the new provider (Art. 25(2)). Practical consequence for API integration: anyone who builds a language model into its own application via an interface and offers that application under its own name is as a rule itself the provider of that AI system (Art. 3(3)). Whether high-risk obligations follow depends on the purpose of use. Lead scoring or text suggestions do not fall under Annex III, applicant selection does. Art. 50(2) may apply to generated texts. The model provider owes you the information under Art. 53(1)(b) and Annex XII. More on delimiting the roles in the Roadmap, section provider vs. deployer.

4. High-risk AI: provider obligations Art. 9-15

The requirements for high-risk AI systems (Chapter III Section 2) apply to Annex III systems from 2 December 2027 and to Annex I systems under Section A from 2 August 2028; for Section B (machinery) see 9.4. The seven core requirements:

1. Risk management system (Art. 9): a continuous iterative process across the entire lifecycle, with regular systematic review: identify, assess and mitigate risks, evaluate residual risks.
2. Data and data governance (Art. 10): training, validation and testing data must be relevant, sufficiently representative and, to the best extent possible, free of errors and complete in view of the intended purpose (paragraph 3). This includes documented procedures on origin, preparation and assumptions (paragraph 2).
3. Technical documentation (Art. 11, Annex IV): it is drawn up before placing on the market and contains at least the information in the nine points of Annex IV. SMEs and small mid-caps may provide the information under Annex IV in a simplified manner; for this purpose the Commission establishes a simplified form, which notified bodies accept (Art. 11(1), second subparagraph).
4. Record-keeping (Art. 12): the system must be able to log events automatically over its lifecycle.
5. Transparency and instructions for use (Art. 13): deployers must be able to interpret the outputs and use them appropriately. To that end, the instructions for use provide information on capabilities, limitations, accuracy and maintenance.
6. Human oversight (Art. 14): the system is designed so that natural persons can effectively oversee it and intervene.
7. Accuracy, robustness, cybersecurity (Art. 15): an appropriate level throughout the lifecycle, declared in the instructions for use.

Added to this are the quality management system (Art. 17, proportionate to the size of the organisation under paragraph 2), the conformity assessment (Art. 43, internal or with a notified body), the CE marking, registration in the EU database (Art. 49), post-market monitoring (Art. 72) and the reporting of serious incidents (Art. 73).

5. GPAI Art. 53-55 – General Purpose AI Models

For providers of general-purpose AI models, Chapter V has applied since 2 August 2025. For models placed on the market before 2 August 2025, a catch-up period runs until 2 August 2027 (Art. 111(3)). Supervision lies with the Commission's AI Office. Since 2 August 2026, the Commission can also impose fines on GPAI providers (Art. 101).

5.1 Obligations of all GPAI providers (Art. 53)

• technical documentation of the model, at least in accordance with Annex XI (paragraph 1(a))
• information and documentation for providers integrating the model into their own systems, at least in accordance with Annex XII (point (b))
• a policy to comply with Union copyright law, in particular to observe a reservation of rights under Art. 4(3) of Directive (EU) 2019/790 (point (c))
• a publicly available, sufficiently detailed summary of the training content according to the template of the AI Office (point (d))

For models released under a free and open-source licence with publicly available parameters, points (a) and (b) do not apply, except in the case of systemic risk (Art. 53(2)). Codes of practice under Art. 56 are voluntary; a presumption of conformity arises only from a harmonised standard (Art. 53(4)).

5.2 GPAI with systemic risk (Art. 51 and 55)

Systemic risk is presumed where the cumulative amount of computation used for training exceeds 10^25 floating-point operations (FLOP) (Art. 51(2)). The Commission can also classify a model on the basis of the criteria in Annex XIII (Art. 51(1)). In addition to Art. 53, the following then apply under Art. 55(1):

• model evaluation in accordance with standardised protocols, including adversarial testing
• assessment and mitigation of systemic risks at Union level
• tracking of serious incidents and reporting to the AI Office without undue delay
• an adequate level of cybersecurity for the model and the physical infrastructure

In depth: GPAI provider obligations, the documentation template GPAI compliance template Art. 53 and Open-source models under the AI Act.

6. Deadlines 2024–2030

The timetable is laid down in Art. 113 as amended by Regulation (EU) 2026/1744. The proposal had still sought to tie application to Commission findings on standards. The adopted version sets fixed calendar dates. Background to the amendment: Digital Omnibus: Regulation (EU) 2026/1744 in force.

Note: 2 December 2026 applies to two different things, the new prohibitions and the legacy systems under Art. 50(2). All dates at a glance are provided by the Compliance Deadlines Calendar 2026 to 2028.

7. Fines

For SMEs including start-ups, the lower of the two amounts applies to every fine under Art. 99 (Art. 99(6)). Since 27 July 2026, this has also applied to small mid-cap enterprises for the fines under paragraphs 4 and 5 (Art. 99(6a)). When imposing penalties, the Member States also take into account the interests of these undertakings and their economic viability (Art. 99(1)). Art. 4 is not in the catalogue of Art. 99(4). Penalties for it are governed by national law under Art. 99(1).

Germany: the KI-MIG contains its own administrative offences for infringements of Art. 21, 27, 45 and 86(1) AI Act with fines of up to EUR 50,000 (§ 15 KI-MIG). For the infringements under Art. 99(3) to (5), the law on administrative offences applies accordingly (§ 16 KI-MIG). No fines are imposed on authorities and other public bodies (§ 17(2) KI-MIG). An order of magnitude for your own case is provided by the fine calculator.

8. First steps for SMEs (short list)

The order follows the deadlines: first what already applies (Art. 4, Art. 5, Art. 50), then preparation for Annex III. The detailed twelve-month plan with phases and budget is contained in the EU AI Act Compliance Roadmap.

1. AI inventory: record all systems with purpose, model, provider and area of use; instructions in AI system inventory in 8 steps, for HR departments in AI inventory in HR.
2. Role per system: provider, deployer or both (Art. 3(3) and (4)), check for a change of role under Art. 25.
3. Rule out prohibitions: check every system against Art. 5, from 2 December 2026 also against points (ba) and (bb).
4. Risk class per system: Annex I or III, give reasons for the exception under Art. 6(3) and document the classification.
5. AI literacy: define and implement measures under Art. 4; keep an internal record (recommendation, not a statutory obligation).
6. Transparency: implement chatbot notices, marking of synthetic content and deep fake disclosure under Art. 50; legacy systems under paragraph 2 by 2 December 2026.
7. High-risk providers: prepare Art. 9 to 15, quality management, technical documentation, conformity assessment and registration by 2 December 2027 (Annex III).
8. High-risk deployers: organise Art. 26, clarify the fundamental rights impact assessment under Art. 27, inform employees in good time.
9. Contracts: agree in writing with suppliers of models, components and services on the necessary information and access (Art. 25(4)).
10. Ongoing operation: post-market monitoring (Art. 72), reporting channels for serious incidents (Art. 73) and review of the inventory with every system change.

A quick start is offered by the AI System Inventory Quickcheck.

9. Sector practice: EU AI Act in common use-cases

The most common fields of use in SMEs, each with the specialist article for depth. An overview with a review path is provided by EU AI Act: Top 5 SME Use Cases.

9.1 Recruiting and HR (Annex III no. 4)

AI for applicant selection, promotion, termination or task allocation is high-risk, from 2 December 2027. Employers as deployers inform the workers' representatives and the affected workers before use (Art. 26(7)). More in AI recruiting AGG-compliant: 8 safeguards and in the FRIA template for Annex III no. 4 (FRIA obligation only for bodies governed by public law and private entities providing public services, Art. 27(1)).

9.2 Education (Annex III no. 3)

AI for admission, evaluation of learning outcomes or proctoring of tests is high-risk. Emotion recognition in education institutions has already been prohibited since 2 February 2025, unless for medical or safety reasons (Art. 5(1)(f)). Details: AI Act in education.

9.3 Creditworthiness and insurance (Annex III no. 5 (b) and (c))

The evaluation of creditworthiness falls under no. 5 (b), risk assessment and pricing in life and health insurance under no. 5 (c). Affected persons have a right to an explanation of the individual decision (Art. 86). In depth: AI Act and credit scoring (Annex III no. 5 (b)) and Life and health insurance.

9.4 Mechanical engineering and embedded AI (Annex I)

Since Regulation (EU) 2026/1744, the Machinery Regulation (EU) 2023/1230 is listed in Annex I Section B (Art. 1 point 41 of Regulation (EU) 2026/1744); for AI as a safety component of machinery, only Art. 6(1), Art. 60a and Art. 102 to 112 of the AI Act apply (Art. 2(2)). The AI requirements come via delegated acts to Annex III of the Machinery Regulation, which apply from 2 August 2028 (Art. 3 point 1 of Regulation (EU) 2026/1744). AI used exclusively for aspects not relevant to safety is not a safety component (Art. 6(1a)), unless its failure would endanger health and safety (paragraph 1b). Details: AI Act and robotics.

9.5 Chatbots and customer service (Art. 50)

A service chatbot is as a rule not high-risk, but since 2 August 2026 must disclose that it is an AI system (Art. 50(1)). Alongside this, the information obligations of Art. 13 and 14 GDPR apply. Implementation examples: Chatbot labelling in 5 variants.

Special rules apply to public bodies, for example the fundamental rights impact assessment under Art. 27 and the deadline of 2 August 2030 for legacy systems (Art. 111(2)). On this: AI Act in public administration.

10. Germany and Austria: supervision and national rules

The AI Act applies in the same way in both countries. What differs is who enforces it.

Germany: KI-MIG

The German AI Market Surveillance and Innovation Promotion Act (KI-MIG) entered into force on 29 July 2026. The Federal Network Agency (Bundesnetzagentur) is the competent market surveillance authority, unless the KI-MIG provides otherwise (§ 2(1) KI-MIG). It is also the single point of contact under Art. 70(2) AI Act (§ 6(1) KI-MIG). Sectoral competences remain in place:

• the existing market surveillance authorities for products under Annex I Section A (§ 2(2))
• the Federal Financial Supervisory Authority (BaFin) for AI directly connected with regulated financial activities (§ 2(3))
• the Länder authorities for public bodies of the Länder (§ 2(6))
• an independent AI market surveillance chamber at the Federal Network Agency for Annex III no. 1, in so far as law enforcement, border management, justice and democracy are concerned, and for nos. 6 to 8 (§ 2(5), § 4)

The Federal Office for Information Security (BSI) is the provisional notifying authority for Annex III no. 1 (§ 3(2) KI-MIG). Since 29 July 2026, infringements of the AI Act also fall within the material scope of the Whistleblower Protection Act (HinSchG) (§ 2(1) no. 10 HinSchG). Reports on them can therefore be made via the internal reporting office.

Austria: no designation under Art. 70

As at 17 August 2026, Austria has not regulated the competences under the AI Act at national level. No market surveillance authority or notifying authority under Art. 70 AI Act has been designated. Under § 20c KOG, the AI Service Desk of RTR is an information and advisory body without supervisory powers. Under Art. 74(8) AI Act in conjunction with §§ 18 and 31 DSG, the Austrian Data Protection Authority is the market surveillance authority for high-risk AI in areas including law enforcement, border management, justice and democracy. That is not a designation under Art. 70. For companies, the obligations of the regulation nevertheless apply directly. What this means for implementation is set out in the Roadmap, section Germany and Austria.

11. Open interpretation questions

On some points the text of the regulation sets the framework, but practice is still lacking:

• Substantial modification and fine-tuning (Art. 25): from when does fine-tuning a model turn the deployer into a provider? Guidelines on this are drawn up by the Commission, not the European Data Protection Board (Art. 96(1)(a) and (c)).
• Depth of the fundamental rights impact assessment: the AI Office is developing a template for a questionnaire, which also allows references to an existing DPIA (Art. 27(4) and (5)). Until then, it remains open which level of detail is expected.
• Open-source systems and models: Art. 2(12) exempts open-source AI systems, but not where they are high-risk or fall under Art. 5 or 50. For GPAI models only the limited exception of Art. 53(2) applies, and not in the case of systemic risk.
• Interface with Art. 22 GDPR: an automated individual decision can at the same time be an Annex III case. Both sets of rules require their own assessments. More in GDPR and EU AI Act for AI applications.
• Legacy systems: Art. 111(2) covers high-risk systems placed on the market before the start of Chapter III only in the event of significant changes in their design. Where this threshold lies has not been clarified. Systems of public authorities must catch up by 2 August 2030.
• Division of labour in the supply chain: Art. 25(4) requires a written agreement on information and technical access between the provider and the supplier. What content supervisory authorities expect for this is open. The AI Office can recommend voluntary model terms.

12. 5 EU AI Act myths

Myth 1: “We only use ChatGPT, we are not affected”

Fact: anyone who uses an AI system professionally under its own authority is a deployer (Art. 3(4)). Art. 4 therefore applies: measures to support the AI literacy of staff. Anyone publishing AI-generated deep fakes must disclose this under Art. 50(4). Anyone who builds the model into its own customer application and offers it under its own name becomes the provider of that system (Art. 3(3)) and must, for example, ensure the chatbot information under Art. 50(1). “API use only” therefore does not mean “no obligations”.

Myth 2: “GPAI only applies to OpenAI and Google”

Fact: Art. 53 applies to every provider placing a GPAI model on the market in the Union, irrespective of its size and place of establishment. The additional obligations of Art. 55 apply only to models with systemic risk. This is presumed above training compute of more than 10^25 FLOP (Art. 51(2)) or determined by the Commission (Art. 51(1)). Anyone releasing a medium-sized model is also a GPAI provider, provided the model displays significant generality (Art. 3(63)). For open-source release, the obligations under Art. 53(1)(a) and (b) and the obligation to appoint an authorised representative (Art. 54(6)) do not apply.

Myth 3: “Internal use is exempt from the AI Act”

Fact: the exemptions are narrow: exempt are deployers who are natural persons using AI in the course of a purely personal non-professional activity (Art. 2(10)). Further exemptions apply to research, testing and development before placing on the market or putting into service (Art. 2(8)), to AI specifically developed for scientific research (Art. 2(6)) and to exclusively military, defence or national security purposes (Art. 2(3)). The internal use of a purchased system makes you a deployer. Anyone who develops in-house and puts the system into service internally under its own name is at the same time a provider (Art. 3(3)) and bears both sets of obligations.

Myth 4: “The Digital Omnibus postpones all obligations”

Fact: Regulation (EU) 2026/1744 has been in force since 27 July 2026 and postpones the high-risk obligations: Annex III from 2 December 2027, Annex I from 2 August 2028. Not postponed are the prohibitions of Art. 5, Art. 50 and the GPAI obligations of Art. 53 to 55. Details: Digital Omnibus on AI.

Myth 5: “FRIA is just a risk assessment” – or the same as the DPIA

Fact: the fundamental rights impact assessment under Art. 27 is a procedure of its own with its own group of addressees. It applies to deployers of high-risk systems under Art. 6(2) that are bodies governed by public law or private entities providing public services, and to deployers of systems under Annex III no. 5 (b) and (c), i.e. creditworthiness and life or health insurance. Annex III no. 2 (critical infrastructure) is excluded. A DPIA under Art. 35 GDPR does not replace the FRIA. In so far as it already covers individual points, the FRIA may refer to it (Art. 27(4)). The obligation applies to Annex III systems from 2 December 2027. Practice template: FRIA template.

Templates and training: which documents belong to which obligation is shown in the overview of the EU AI Act templates. The EU AI Act Kit brings them together as 65 editable Word, Excel and PowerPoint templates, from EUR 990 net. For the measures under Art. 4 there is the “AI Literacy” e-learning with a certificate of participation. Standardised documentation, no legal advice in the individual case.