Since July 27,2026, Article 4 of the AI Act, the European regulation on artificial intelligence, no longer says the same thing. Regulation (EU) 2026/1744, published in the Official Journal on July 24, replaced its text in its entirety, without deleting it. Companies must now read a reworded obligation, the scope of which requires careful reading.
The 2024 version required providers and deployers of AI systems to take measures to “ensure, to the best extent possible, a sufficient level” of AI proficiency among their personnel. The new version covers the same subject and the same criteria: the knowledge, experience, education, and training of individuals, the context of use, and the persons concerned. The verb, however, changes. It is now about taking measures to “promote the development” of this proficiency, and the text specifies that the obligation does not require guaranteeing a specific level for each individual.
Two paragraphs have been added. The Commission and Member States must support companies, notably through training, information resources, and the exchange of best practices, while the European AI Board will adopt recommendations. Recital 8 indicates that strict obligations are not suitable for all operators and weigh especially heavily on small businesses.
An obligation of means, still applicable
Reading this change as a removal would be a mistake. Article 4 still applies to all providers and deployers, regardless of their size. A company whose employees use a conversational assistant, a writing tool, or an agent integrated into their messaging system remains a deployer under the regulation.
The obligation has shifted from a logic of result to a logic of means, judged on the actions actually implemented. If a client, auditor, contractor, or authority asks the question, the answer “we did nothing, the text has been lightened” does not hold. The rest of the regulation has otherwise kept its timeline: the transparency obligations of Article 50 have applied since August 2,2026, while the obligations for high-risk systems are postponed to December 2,2027 for those in Annex III, such as recruitment or creditworthiness assessment, and to August 2,2028 for AI embedded in products from Annex I.
What a company must be able to show
The text does not prescribe any format. In practice, three elements suffice to answer most requests. The first is an inventory of uses: which AI tools are used, by which teams, for which tasks, and with which data. Without this inventory, it is impossible to verify that training corresponds to actual uses.
The second is a register of actions. For each action, the date, audience, content, duration, format (training, workshop, internal guide, or usage charter) and the link to an identified use should be noted, supported if possible by an attendance or completion certificate. The third is a responsible person, a designated individual who keeps this register up to date and reviews it upon the arrival of each new tool. In an SME, this is often the manager or the head of information systems.
Frequent mistakes
The first is to stop everything, whereas the easing relates to the level to be guaranteed and not to the existence of the obligation. The second is to train everyone in the same way: a generic one-hour webinar poorly addresses the question of a lawyer using a contract summarization tool, or an HR team testing a pre-selection tool. The third confuses training and documentation. Trained employees without written records can demonstrate nothing, and a register filled out without actual training does not withstand a specific question. Finally, a register frozen in 2025 does not cover an agent deployed in 2026: it must follow the inventory of uses.
A company starting from scratch can proceed in three steps: inventory actual uses, including those not approved by management; plan a proportionate action for each; then record these actions in a register with a designated responsible person. This work, which takes a few days according to the article’s presentation, also serves beyond Article 4, when facing a client questioning the use of AI in a service or when preparing for obligations on high-risk systems.
