The AI Act applies to any company that develops, sells, or uses an AI system in the European Union. Its obligations, however, depend on two variables: your role, provider or deployer, and the system’s risk level. Since 2 August 2026, the majority of the regulation has been applicable, while the rules for high-risk systems have been postponed to 2 December 2027 and 2 August 2028 by the 2026 amending regulation, known as the ‘omnibus’. This guide gives you a comprehensive overview: who is targeted, how to classify a system, which obligations to comply with, and by what date.
Regulation (EU) 2024/1689, better known as the AI Act, entered into force on 1 August 2024. It was amended by Regulation (EU) 2026/1744, the ‘Digital AI Omnibus’, proposed by the Commission on 19 November 2025 and entered into force on 27 July 2026, according to the European Commission. The dates and articles cited here were verified in November 2026 on the consolidated version published by the Commission’s AI Act Helpdesk. This article is an information summary: it does not replace legal advice in a real case.
Who is covered by the AI Act
Article 2 of the regulation sets a very broad scope. It covers providers who place an AI system or a general-purpose AI model on the European market, whether established in the Union or elsewhere, deployers established in the Union, as well as providers and deployers located outside the Union when the system’s outputs are used within the Union. Importers, distributors, manufacturers who integrate AI into their product under their own brand, and authorized representatives of foreign providers complete the list.
In practice, a French SME using a CV screening software, an American SaaS publisher whose tool is sold to European clients, or an agency delivering a chatbot to a client all fall within the scope. The question is therefore almost never: ‘Am I covered?’, but rather: ‘In what capacity and for which system?’. A satellite article in this dossier will detail, case by case, which companies and uses are covered. We detail this point in ‘AI Act: Which Companies and Uses Are Covered?’.
Definitions govern the rest
The regulation defines an AI system as an automated system designed to operate with varying levels of autonomy, which may adapt after deployment and which infers, from the inputs it receives, how to generate predictions, content, recommendations, or decisions. Recital 12 of the regulation expressly excludes systems that automatically execute operations according to rules set solely by humans; conversely, a language model integrated into a business tool does infer its responses.
Two roles then structure the obligations. The provider is the one who develops a system or model, or has it developed, and then places it on the market or into service under its own name or brand, whether for payment or free of charge. The deployer is any person or organization using an AI system under its authority, excluding personal non-professional use. A single company can therefore be a provider of a tool it markets and a deployer of another tool it purchases. To go further, read our article ‘AI Act: Obligations for AI Providers and Deployers’.
Exclusions remain limited
The regulation does not apply to systems used exclusively for military, defence, or national security purposes, nor to systems developed and put into service solely for scientific research. Research and testing activities conducted before market placement are also excluded, except for real-world testing. Natural persons using AI purely for personal purposes are also exempt from deployer obligations.
Systems released under an open-source licence also benefit from an exclusion, but this falls away as soon as the system is placed on the market as a high-risk system, falls under a prohibited practice, or triggers a transparency obligation under Article 50. For an initial diagnosis, the interactive test ‘Am I covered by the AI Act?’ on the site asks you to choose your role and use cases, before redirecting you to a risk level and associated deadlines.
How to classify your systems according to risk level
The Commission summarises the AI Act into four risk levels on its page dedicated to the regulatory framework, updated on 19 October 2026. Classification follows a precise order: first check if the use is prohibited, then if it is high-risk, then if it triggers a transparency obligation. General-purpose AI models follow a separate track, imposing obligations on their providers, not end-users.
Prohibited practices are an absolute stop
Article 5 prohibits, among other things, manipulation through subliminal or deceptive techniques causing significant harm, exploitation of vulnerabilities related to age, disability, or social situation, social scoring, predicting offences based solely on profiling, creating facial recognition databases through untargeted scraping of images, emotion recognition in the workplace and education (except for medical or safety reasons), certain biometric categorisation, and real-time remote biometric identification by law enforcement (except under strict exceptions). The omnibus added two prohibitions applicable from 2 December 2026: systems that generate intimate or sexually explicit images, videos, or audio of identifiable persons without their consent, and those that generate child sexual abuse material.
High-risk systems are subject to strict rules
Article 6 provides two entry gates. The first concerns AI that is a safety component of a product covered by the harmonisation legislation listed in Annex I (e.g., machinery, toys, or lifts) and subject to third-party conformity assessment. The second covers uses listed in Annex III, divided into eight areas: biometrics, critical infrastructure, education and training, employment and worker management, access to essential public and private services (including creditworthiness assessment and life/health insurance pricing), law enforcement, migration and border control, justice and democratic processes.
A system listed in Annex III may escape qualification if it does not pose a significant risk to health, safety, or fundamental rights, for example because it performs a narrow procedural or preparatory task. This exception never applies when the system performs profiling of natural persons. The provider invoking it must document its assessment before market placement and remains obliged to register the system in the EU database, according to the consolidated Article 6. This topic is covered in a dedicated article: ‘AI Act: How to Classify an AI System According to Its Risk?’.
Transparency covers only certain uses
The transparency obligations under Article 50 cover systems that interact with people, those that generate synthetic content (text, image, audio, or video), emotion recognition and biometric categorisation systems, as well as deepfakes (deepfakes). Everything else falls under minimal risk, with no specific obligations: the Commission cites spam filters and AI in video games. A satellite article dedicated to AI Act risk classification will outline the full decision tree. We detail this point in ‘Generative AI: What Transparency and Labelling Obligations?’.
| Level | Example uses | Main consequence | Application |
|---|---|---|---|
| Prohibited (Article 5) | Social scoring, emotion recognition at work, non-consensual sexual deepfakes | Use prohibited | 2 February 2025; 2 December 2026 for the two new prohibitions |
| High risk, Annex III | CV screening, credit scoring, education admission | Conformity requirements, documentation, human oversight | 2 December 2027 |
| High risk, Annex I | AI as a safety component of machinery, a toy, or a lift | Same requirements, integrated into product assessment | 2 August 2028 |
| Transparency (Article 50) | Chatbot, image generator, published deepfake | Information of persons, content labelling | 2 August 2026 (labelling: 2 December 2026 for systems already on the market) |
| Minimal risk | Spam filter, video game AI | No specific obligation, aside from AI literacy | Not applicable |
Source: Consolidated Regulation (EU) 2024/1689, Articles 5, 6, 50, and 113, and Commission AI Act page. Data verified in November 2026.
Three cases to illustrate
A company that purchases an automated candidate analysis and filtering tool falls under Annex III, point 4: it is a high-risk system, and the company is the deployer. If it adds a chatbot to its website to answer customer questions, the tool falls under Article 50: the provider must design it so the user knows they are talking to an AI, unless this is obvious. Finally, the spam filter in its email system triggers no specific obligations under the regulation.
What obligations apply to providers and deployers
The bulk of obligations falls on the provider of a high-risk system. Article 16 requires it to comply with the technical requirements of Chapter III (risk management, data quality, technical documentation, logging, transparency, human oversight, accuracy, robustness, and cybersecurity), implement a quality management system, have the conformity assessment performed, draft the EU declaration of conformity, affix the CE marking, and register the system in the EU database.
The deployer has shorter but very concrete obligations, listed in Article 26: use the system in accordance with the instructions, entrust human oversight to competent persons with the necessary authority, ensure the relevance of input data under its control, monitor operation and report serious risks and incidents, keep logs for at least six months, and inform employee representatives before using such a system at work. Public bodies, private entities providing public services, and deployers of credit scoring or life/health insurance pricing systems must additionally conduct a fundamental rights impact assessment, as per Article 27.
| Obligation | Provider | Deployer |
|---|---|---|
| AI literacy of personnel (Article 4) | Yes | Yes |
| Risk management, technical documentation, quality system | Yes (high risk) | No |
| Conformity assessment, EU declaration, CE marking | Yes (high risk) | No |
| Registration in the EU database | Yes (high risk and systems exempted via Article 6(3)) | Only authorities and public bodies |
| Use according to instructions, human oversight | Designs the oversight | Exercises it |
| Log retention | Yes, if under its control | At least six months |
| Fundamental rights impact assessment | No | Public bodies, public services, credit and insurance |
| Transparency (Article 50) | Chatbots, labelling of generated content | Deepfakes, texts of public interest, emotions and biometrics |
A satellite article will detail the AI Act obligations for providers and deployers, point by point.
The deployer can become a provider
Article 25 provides three situations where a distributor, importer, or deployer assumes provider obligations: if they place their name or brand on a high-risk system already on the market, if they substantially modify it, or if they change the intended purpose of a system, including a general-purpose AI, such that it becomes high-risk. A company connecting a general-purpose model to a candidate selection process must therefore consider its role before putting it into service.
AI literacy also covers general-purpose models
The omnibus rewrote Article 4 on AI literacy: providers and deployers must take measures to support the development of their personnel’s skills, taking into account their experience and context, without being required to guarantee a given level for each person. The Commission and Member States now have a supporting role, especially for SMEs. Providers of general-purpose AI models must, since 2 August 2025, maintain technical documentation, inform downstream providers, apply a copyright compliance policy, and publish a summary of training data.
Sanctions, set by Article 99, can reach up to €35 million or 7% of worldwide annual turnover for a prohibited practice, and up to €15 million or 3% for violating operator or transparency obligations, whichever is higher. For SMEs and start-ups, the lower amount applies; the omnibus extended this rule to small mid-caps, the small mid-caps, for a portion of the fines.
What timeline applies after the 2026 amendments
Article 113, amended by the omnibus, now sets a general application date of 2 August 2026, with exceptions before and after that date. The table below outlines the deadlines from the consolidated version and the Commission’s official timeline. The site analysed the practical consequences of this partial postponement of the AI Act timeline upon its adoption. To delve deeper, consult ‘AI Act Timeline: Which Obligations Apply and When?’.
| Date | What applies |
|---|---|
| 1 August 2024 | Regulation enters into force |
| 2 February 2025 | Definitions, AI literacy, prohibited practices |
| 2 August 2025 | General-purpose AI models, governance, sanctions regime |
| 2 August 2026 | General application, including Article 50 transparency and supervision by authorities |
| 2 December 2026 | New prohibitions, sexual deepfakes and child sexual abuse material; end of labelling grace period for generative systems placed on the market before 2 August 2026 |
| 2 August 2027 | Compliance for general-purpose models placed on the market before 2 August 2025 |
| 2 December 2027 | Annex III high-risk systems |
| 2 August 2028 | Annex I product-related high-risk systems |
| 2 August 2030 | Compliance for existing high-risk systems intended for public authorities |
Source: Consolidated Articles 113 and 111, Commission implementation timeline. Verified in November 2026.
Common reading errors
First error: believing the omnibus postponed the entire regulation. Only the rules for high-risk systems were delayed; prohibitions, AI literacy, rules for general-purpose models, and Article 50 transparency already apply. Second error: thinking the deadline until 2 December 2026 covers all transparency. It only concerns the technical labelling of generated content, under Article 50(2), and only for systems placed on the market before 2 August 2026.
Third error: considering a high-risk system already in service is permanently safe. Article 111 only exempts it as long as it does not undergo a significant design change, and sets a deadline of 2 August 2030 for those intended for public authorities. A satellite article will keep the AI Act timeline updated, deadline by deadline.
How to organise compliance and security
AI Act compliance starts with an inventory, not a legal audit. Without a list of AI tools actually used, including AI features integrated into purchased software and undeclared employee usage, no classification is possible. The Shadow AI phenomenon is massive: the site reported that two out of three professionals report unauthorised AI usage in their company, according to the Forvis Mazars and ESSCA barometer published on 1 October 2026. To delve deeper, consult ‘Shadow AI: How to Manage Undeclared AI Usage?’.
Once the inventory is established, each system is assigned a role, a risk level, a named responsible person, and a review date. This per-system sheet becomes the basis for documentation that authorities can request. An AI usage charter for companies, which the hub will dedicate a guide to, then sets common rules: authorised tools, prohibited data, human validation of published content. This topic is covered in a dedicated article: ‘AI Usage Charter for Companies: Template and Method’.
Security must go beyond the text
Article 15 requires high-risk systems to have an appropriate level of accuracy, robustness, and cybersecurity throughout their lifecycle. It explicitly cites AI-specific attacks: training or model data poisoning, adversarial examples designed to deceive the model, confidentiality attacks. For systems that are not high-risk, the regulation does not impose equivalent requirements, but the risks remain: leakage of confidential data sent to an online assistant, injection of hidden instructions in a document (prompt injection), an agent with overly broad permissions. To go further, read our article ‘Can You Send Confidential Documents to ChatGPT or Claude?’.
AI agents concentrate these risks because they act within the information system. Our guide to AI agents explains how they work, and AI agent governance (identity, permissions, responsible person) will be the subject of a dedicated article. When data is too sensitive to leave the company, running a model internally is an option: the guide to local AI and RAG presents possible architectures. The link between GDPR and artificial intelligence, distinct from the AI Act, will be treated separately. This topic is covered in a dedicated article: ‘GDPR and Artificial Intelligence: How to Protect Data?’.
Compliance must be provable
A compliance setup is verified by evidence, not intentions. For each high-risk system deployed, you must be able to present the provider’s instructions, the names of supervisors and their training, retained logs, proof of staff information, and, where applicable, the impact assessment. For each generative tool, you must be able to show how the user is informed and how published content is reviewed. A quarterly review of the inventory limits the risk of drift, without guaranteeing the absence of undeclared usage.
Tools, checklist, and official legal sources
The following checklist summarises the steps, in the order where they condition the next ones. It does not cover sectoral obligations (finance, health, regulated products), which are additional to the AI Act. A more detailed AI Act compliance audit, aimed at SMEs and publishers, will be published in this hub. We detail this point in ‘AI Act Audit: Compliance Checklist for SMEs and Publishers’.
- Inventory all AI systems used or marketed, including features integrated into off-the-shelf software and individual usage.
- Check the definition: does the system infer its results, or apply fixed rules?
- Assign the role: provider, deployer, importer, distributor. Check the three cases under Article 25 (branding, substantial modification, change of intended purpose).
- Exclude prohibited practices under Article 5, including the two prohibitions applicable from 2 December 2026.
- Compare the use to Annex III and Annex I; document any derogation invoked under Article 6(3).
- Identify transparency obligations: chatbot, generated content, deepfake, published text of public interest.
- Train the relevant personnel and keep a record of actions taken (Article 4).
- Organise human oversight and log retention for high-risk systems.
- Plan the 2 December 2027 and 2 August 2028 deadlines into procurement and development roadmaps.
- Appoint a responsible person and set a review date for each system.
Where to check the applicable law
The reference text is Regulation (EU) 2024/1689 in its consolidated version on EUR-Lex, with the amending Regulation (EU) 2026/1744. The Commission’s AI Act Helpdesk publishes this consolidated text article by article, indicating passages amended or added by the omnibus, as well as an implementation timeline. The Commission’s guidelines (definition of AI system, prohibited practices, classification, transparency) complement the text without replacing it.
These sources have a limit: they describe the law, not your situation. A borderline case, like a candidate pre-screening tool presented as a simple aid or a general-purpose model adapted for a sensitive use, merits advice from a specialised lawyer. The timeline may still change via delegated acts or new guidelines: recheck the dates at each quarterly review.
