Does the AI Act really apply to you?
This is the question almost every executive asks. Your company uses ChatGPT or Claude, an assistant in its email, perhaps a CV-screening tool or a chatbot on its website. Does the AI Act require anything of you, and by what date? The short answer: yes, almost always, but far less than you may have heard, and not right away for the heaviest part.
The AI Act is the European regulation on artificial intelligence, Regulation (EU) 2024/1689. It entered into force on 1 August 2024 and applies in stages. In July 2026, a second text, the Digital Omnibus on AI, shifted part of the timetable. This guide reviews what already applies at the end of 2026, what is coming in 2027 and 2028, and what an SME or a mid-sized company needs to do in practice. If your first priority is to train your teams in the safe use of AI, our page on AI training for companies presents the available programmes.
The AI Act in two minutes
The AI Act does not regulate "AI" in general. It regulates uses, according to the risk they pose to people's health, safety and rights. The higher the risk, the heavier the obligations.
The text also distinguishes two roles:
- the provider, who develops an AI system and places it on the market (a model developer, an HR software vendor, etc.);
- the deployer, who uses that system in the course of their professional activity.
The vast majority of companies are deployers. Their obligations exist, but they are lighter than those of providers. One nuance matters: a company that has a custom AI tool developed and puts it into service under its own name may be considered the provider of that tool. The official text is available on EUR-Lex.
The four risk levels
Unacceptable risk: prohibited practices
Since 2 February 2025, certain practices have been banned in the Union: social scoring of individuals, manipulation that exploits the vulnerabilities of an audience, emotion recognition in the workplace or at school (except for medical or safety reasons), building facial recognition databases by scraping images from the internet, and real-time remote biometric identification in public spaces for law enforcement purposes, outside strictly defined exceptions.
The Digital Omnibus adds two bans, applicable from 2 December 2026: systems designed to generate intimate content without the consent of the people depicted, and those that produce child sexual abuse material.
High risk: heavy obligations, but postponed
Two families of systems are concerned.
- Annex III: systems used in sensitive areas. For a company, the most frequent cases are recruitment and personnel management (CV screening, candidate assessment, promotion or dismissal decisions), creditworthiness assessment of individuals, pricing in life and health insurance, and education and vocational training (admission, grading).
- Annex I: AI that serves as a safety component in products that are already regulated, such as machinery, medical devices or toys.
The provider must then manage risks, control data quality, draw up technical documentation, keep logs and provide for human oversight. The deployer must use the system according to its instructions, entrust oversight to competent people, retain the logs and, if they are an employer, inform staff representatives and the employees concerned before the system is put into service.
Limited risk: transparency
This is Article 50, applicable since 2 August 2026. A person interacting with a chatbot must know they are talking to an AI, unless it is obvious. AI-generated content (text, image, sound, video) must be marked in a machine-detectable way: this is mainly the providers' responsibility. On the deployer side, a deepfake must be disclosed as such, and an AI-generated text published to inform the public on a matter of public interest must be disclosed too, unless it has been reviewed by a human who takes editorial responsibility for it. Generative systems already on the market before 2 August 2026 have until 2 December 2026 for marking.
Minimal risk
Spam filters, spell checkers, content recommendations, internal writing assistance: most everyday uses fall here. There is no specific obligation, apart from the AI literacy requirement set out in Article 4.
The timetable, updated at the end of 2026
| Date | What applies |
|---|---|
| 1 August 2024 | Regulation enters into force |
| 2 February 2025 | Prohibited practices and AI literacy (Article 4) |
| 2 August 2025 | Rules on general-purpose AI models and penalty regime |
| 27 July 2026 | Digital Omnibus enters into force, new wording of Article 4 |
| 2 August 2026 | Transparency obligations (Article 50) |
| 2 December 2026 | End of the marking deadline for generative systems already on the market, and two new bans |
| 2 December 2027 | High-risk systems under Annex III (HR, credit, education…) |
| 2 August 2028 | High-risk systems embedded in regulated products (Annex I) |
What the Digital Omnibus changed
The Digital Omnibus on AI is Regulation (EU) 2026/1744 of 8 July 2026. It was voted by the European Parliament on 16 June 2026, adopted by the Council on 29 June, published in the Official Journal on 24 July and entered into force on 27 July 2026.
It changes three things for companies. First, the obligations on Annex III high-risk systems move from 2 August 2026 to 2 December 2027, sixteen months later. Those for Annex I move to 2 August 2028. Second, Article 4 on AI literacy is rewritten (see below). Finally, two new practices are banned.
It does not change the rest. The 2025 bans remain in place, the transparency rules have indeed applied since August 2026, and providers of large models remain subject to their obligations since August 2025. The postponement buys time; it exempts you from nothing.
Article 4: training your teams is still expected
In its 2024 version, Article 4 required providers and deployers to ensure a sufficient level of AI literacy among their staff. Since 27 July 2026, they must "take measures to support" the development of that literacy. There is no longer an obligation of result, no level is required for each employee, and Article 4 has never carried a penalty of its own.
Should you therefore drop training? No, for two practical reasons. First: in the event of an incident (customer data pasted into a consumer tool, false content published without review), the company will have to show what it had put in place. Second: the real risk in 2026 is not the fine, it is unsupervised use. In many companies, some employees already use AI tools with their personal accounts, with no common rules. A short charter and training tailored to each job take care of most of it. Our guide to training your teams in AI details the method, and our AI training courses can be funded by your OPCO.
The penalties provided for
| Breach | Maximum fine |
|---|---|
| Prohibited practice (Article 5) | €35M or 7% of annual worldwide turnover |
| Other obligations (high risk, transparency…) | €15M or 3% of annual worldwide turnover |
| Incorrect or misleading information to the authorities | €7.5M or 1% of annual worldwide turnover |
For a large company, the higher of the two amounts serves as the cap. For an SME or a start-up, it is the lower one. These are maximums: the authority takes into account the seriousness, the duration and the company's cooperation.
Actions to take now
Take the example of a 60-person services SME. The sales team uses a conversational assistant, customer service is testing a chatbot, HR is considering an application pre-screening tool. Here is the order in which we would help it move forward.
- Take stock of uses. List the AI tools in use, officially or not, department by department. An AI maturity audit is a good starting point.
- Classify each use. Here: the sales team's assistant is minimal risk, the chatbot falls under transparency, the HR tool will be high risk.
- Rule out any prohibited practice. For example, software that would analyse employees' emotions during their calls.
- Apply transparency. A clear notice at the start of the chatbot conversation, and a human review rule for published content.
- Set rules and train. A usage charter (authorised tools, prohibited data, review) and a short training course for each job.
- Prepare for December 2027. Before signing with an HR vendor, ask for its compliance documentation, plan who will check the results and inform the works council (CSE).
To test a new use without taking risks, an AI POC that builds in these criteria from the start avoids rebuilding later. And to choose between several projects, calculating the ROI of AI helps you keep the most useful ones.
Mistakes to avoid
- Believing the postponement exempts you from everything. It only concerns high risk.
- Forgetting transparency. A chatbot that does not present itself as an AI has been in breach since August 2026.
- Letting everyone choose their own tools. Without a common rule, company data ends up in personal accounts.
- Treating the AI Act without the GDPR. Both texts apply at the same time as soon as personal data is involved.
Other classic pitfalls are detailed in our article on mistakes to avoid in an AI project.
Going further
- AI risk classification under the AI Act
- AI Act compliance checklist for SMEs
- GDPR and AI: how to stay compliant
- AI governance: setting up an ethics committee
Frequently asked questions
Does the AI Act apply to an SME that only uses ChatGPT?
Yes, but lightly. An SME that uses a conversational assistant to draft or summarise is the deployer of a minimal-risk system. It mainly has to take measures so that its teams use it properly (Article 4) and comply with transparency if it publishes generated content on matters of public interest.
Did the Digital Omnibus remove the obligation to train employees?
No. It softened it: since 27 July 2026, the company must take measures to support AI literacy, with no obligation of result. Training teams remains the simplest way to demonstrate this.
When do the rules on high-risk AI apply?
On 2 December 2027 for Annex III systems, such as recruitment or credit assessment. On 2 August 2028 for AI embedded in products that are already regulated.
What penalties does an SME risk?
The caps range from €7.5M to €35M, or from 1% to 7% of worldwide turnover. For an SME, the lower of the two amounts applies.
Do you have to disclose that a chatbot is an AI?
Yes, since 2 August 2026, unless it is obvious to the user. A clear sentence at the start of the conversation is generally enough.



