Updated in July 2026 to include the changes made by the Digital Omnibus on AI.
The AI Act is a European regulation that applies to every deployer of AI, including a company that simply uses ChatGPT, Microsoft Copilot or Claude in its daily work.
Among other obligations, it introduces a very operational requirement: every employee who works with an AI system must reach a sufficient level of AI literacy.
In practice, the AI Act changes things on the ground far more than in the legal department. Here is how.
The AI Act: what the regulation says and who it applies to
The AI Act, adopted in June 2024 and in force since August 1, 2024, is the world’s first legal framework for artificial intelligence. It applies in stages: the ban on “unacceptable risk” AI and the AI literacy obligation (Article 4) have applied since February 2025, the obligations on general-purpose AI models since August 2025, and the transparency obligations (Article 50) and the penalty regime apply from August 2, 2026. Following the Digital Omnibus on AI (adopted by the Council of the EU on June 29, 2026), the obligations for high-risk AI systems have been postponed: to December 2, 2027 for stand-alone systems (Annex III) and to August 2, 2028 for systems built into products that are already regulated (Annex I).
The regulation classifies each AI system by its level of risk, then imposes proportionate obligations: from a simple duty to inform up to CE marking (Conformité Européenne) for the most sensitive uses.
Who does it apply to?
Far more organizations than you would think. The text covers AI providers (those who develop and sell AI), but also deployers: any organization that brings an AI system into its business. That includes a French company that uses an HR tool to sort résumés, a chatbot for its customer service, or simply ChatGPT, Copilot or Claude within its teams. The regulation also applies beyond Europe’s borders: a US vendor that sells an AI solution to European customers is subject to it too.
In France, the CNIL (the French data protection authority) has been designated as the national lead authority for supervising the AI Act, alongside the DGCCRF (consumer protection), Arcom (the media regulator) and ANSSI (the national cybersecurity agency), each within its own scope.
The 4 risk levels and what they change for your teams
The regulation defines four levels. Each comes with its own obligations.
Unacceptable risk: banned since February 2, 2025
These are the uses considered contrary to fundamental rights: generalized social scoring, exploiting people’s vulnerabilities, emotion recognition at work or at school, mass scraping of facial images. These systems are banned from the European market. A violation can lead to a fine of up to €35 million or 7% of worldwide turnover.
The Digital Omnibus also adds a new ban on non-consensual intimate content generated by AI (“nudifiers”) and on child sexual abuse material (CSAM), applicable after a transition period that runs until December 2, 2026.
High risk: full application in December 2027
Originally set for August 2, 2026, these obligations have been postponed by the Digital Omnibus: to December 2, 2027 for stand-alone high-risk systems (Annex III) and to August 2, 2028 for systems built into products that are already regulated (Annex I). Eight areas are covered:
- biometrics,
- critical infrastructure,
- education and vocational training,
- employment,
- access to essential private services and to essential public services and benefits,
- law enforcement,
- migration,
- the administration of justice.
For these systems, the provider must obtain CE marking before putting the AI on the market. The deployer, for its part, must set up risk management, check the quality of the data, guarantee human oversight and keep technical documentation. Penalties: up to €15 million or 3% of worldwide turnover.
Limited risk: transparency is mandatory
This is the category that covers the most concrete cases at work: chatbots, conversational assistants, consumer generative AI. The obligation is simple: tell users that they are interacting with an AI, or that a piece of content was generated by one. The obligation becomes fully applicable on August 2, 2026, together with the penalty regime (up to €15 million or 3% of worldwide turnover for breaching the obligations of Article 50).
Note: unlike the high-risk obligations, these transparency obligations have not been postponed: they do apply from August 2, 2026, as does the penalty regime.
Minimal risk: a voluntary code of conduct
Spam filters, spell checkers, AI in video games. No obligation, but a code of conduct is still recommended. This is the category that covers most of today’s uses.
Article 4 of the AI Act: the training obligation (AI literacy)
This is probably the most underestimated part of the regulation. Article 4 places an AI literacy obligation on every deployer of AI: providers and deployers must take appropriate measures to support the development of their staff’s AI literacy. The Digital Omnibus softened the original wording (“ensure a sufficient level”) into an obligation of means, but the requirement has applied since February 2025.
In practice, as soon as one of your employees uses an AI tool in their work, you are required to have trained them to a level suited to that use.
Why AI literacy is an operational issue, not a legal one
Article 4 can’t be handled with a memo. It is handled with a real AI literacy program, anchored in the way each job uses AI. Three reasons:
- Literacy is measured in use, not in theoretical knowledge. An employee who can define generative AI but can’t use it on a concrete case isn’t “trained” in the sense of the regulation.
- The risks vary by role: a buyer, a salesperson and an HR professional don’t have the same exposure. One-size-fits-all training doesn’t hold up.
- And the regulation asks you to keep records: who was trained, on what, and when. Without records, there is no proof.
This obligation calls for a real upskilling plan. We explain where to start: training your people on AI.
Map the real uses first
When they do the exercise, many companies will discover that their teams already have dozens of uses of AI, often in the form of shadow AI (tools used without the IT department’s approval). The first step toward compliance isn’t training: it is mapping the tools, the use cases and the employees concerned, job by job.
Train with concrete cases, not with a catalog
AI training that works starts from the teams’ daily pain points (reviewing a contract, preparing a response to a request for proposal, handling a recurring email), not from a tool’s catalog of features.
That is exactly the approach we describe in our complete guide to a successful AI rollout: start from the front line, identify concrete cases, train through practice, measure the gains.
Conclusion
The companies that take the AI Act seriously won’t just avoid penalties: they will come out of it with a map of their AI uses, AI-literate teams and concrete use cases already identified.
The AI Act isn’t a constraint. It is a chance for everyone to level up.
Turn the Article 4 obligation into a lever for AI adoption.
FAQ: your questions about the AI Act for companies
Does the AI Act apply to a company that simply uses ChatGPT?
Yes. The AI Act covers AI providers, but also deployers: any organization that brings an AI system into its business, including a company that uses ChatGPT, Microsoft Copilot or Claude in its daily work, an HR tool to sort résumés or a chatbot for its customer service. In France, the CNIL (the French data protection authority) is the national lead authority for supervising it, alongside the DGCCRF, Arcom and ANSSI within their respective scopes.
What is the AI Act’s timeline?
The regulation came into force on August 1, 2024 and applies in stages: the ban on unacceptable-risk AI and the AI literacy obligation since February 2025, the obligations on general-purpose AI models since August 2025, and the transparency obligations and the penalty regime on August 2, 2026. After the Digital Omnibus on AI, the obligations for high-risk systems are postponed to December 2, 2027 (Annex III) and August 2, 2028 (Annex I).
What are the 4 risk levels of the AI Act?
Unacceptable risk (social scoring, emotion recognition at work…) is banned, with fines of up to €35 million or 7% of worldwide turnover. High risk (employment, education, biometrics, critical infrastructure…) requires risk management, human oversight and documentation. Limited risk (chatbots, generative AI) requires telling users that they are interacting with an AI. Minimal risk (spam filters, spell checkers) brings no obligation, though a code of conduct is still recommended.
What does Article 4 of the AI Act require on training?
Article 4 requires providers and deployers to take appropriate measures to develop their staff’s AI literacy, an obligation in force since February 2025. In practice, an employee who uses an AI tool in their work must be trained to a level suited to that use, and the company must be able to document who was trained, on what and when. The first step is to map the real uses, often shadow AI.
