Back to Insights
AI regulation10 min read

AI Act after the Omnibus: what already applies, what waits until December 2027

TokenShift Executive Note

AI-translated from the original French version.

AI Act after the Omnibus: what already applies, what waits until December 2027

“The AI Act has been postponed” is a poor summary of Regulation (EU) 2026/1744, known as the Digital Omnibus on AI. The text moves the rules that apply to high-risk systems. It does not move the transparency obligations, and it rewrites the AI literacy obligation without removing it. An executive committee that had frozen its entire compliance programme until 2027 would therefore have put on hold part of what is already enforceable. Planning starts with sorting the company's AI uses into three categories: what already applies, what has been postponed, and what is worth documenting now.

The postponement covers only high-risk systems

The text was published in the Official Journal on 24 July 2026 and entered into force on the third day after publication, 27 July. It amends Article 113 of the AI Act, which sets the dates of application (consolidated text of the Omnibus, AI Act Explorer):

  • high-risk systems listed in Annex III (recruitment, creditworthiness assessment and access to essential services, among others) become subject to the Regulation's obligations from 2 December 2027, instead of 2 August 2026;
  • high-risk systems embedded in products that are already regulated (Annex I) become subject to them from 2 August 2028.

The postponement applies to this block only. The prohibited practices, which have applied since February 2025, and the obligations of providers of general-purpose AI models, which have applied since August 2025, keep their timetable. So does Article 50. The consequence for a company is clear: only the part of the programme dealing with high-risk systems can slip to 2027.

Article 50 has applied since 2 August 2026

This is the part that talk of a “postponement” makes easiest to overlook. Article 50 sets out the transparency obligations: telling people they are interacting with an AI system, marking generated content, and disclosing deepfakes and certain texts. Its date of application has not changed. The Omnibus amended only one paragraph, the seventh, which covers the codes of practice the Commission encourages for detecting and labelling generated content (Article 50, the Commission's AI Act Service Desk). Paragraphs 1 to 5, which set out the obligations and how they apply, are unchanged.

Provider or deployer: the role determines the obligation

Paragraphs 1 and 2 apply to providers. A provider that designs a system intended to interact with people must ensure that those people are informed of it. A provider of a system that generates text, images, audio or video must ensure that the outputs are “marked in a machine-readable format”. Paragraphs 3 and 4 apply to deployers, meaning the company that uses the system. Deployers must inform people when emotion recognition or biometric categorisation is used, and disclose deepfakes and certain generated texts.

The company therefore needs to know which role it holds for each use. Take a hypothetical case: an insurer that develops its customer service assistant in-house and puts it into service under its own name is the provider of that assistant within the meaning of the Regulation. The paragraph 1 obligation then falls on the insurer, not on the vendor of the model it uses.

Chatbots: obviousness is judged from the user's point of view

Disclosure is not required where it is obvious to a reasonably well-informed, observant and circumspect person that they are dealing with an AI system, taking into account the circumstances and the context of use. The test is what that reference person would understand in that situation, not what the company assumes its customers know. An assistant with a first name and a conversational tone, placed in a customer journey where the customer expects a human adviser, would struggle to rely on this exception. As a precaution, an explicit notice removes the need to prove obviousness.

Timing matters too. The information must be provided at the latest at the time of the first interaction or exposure, and this rule applies to all of paragraphs 1 to 4. A notice buried in the terms and conditions is unlikely to meet this requirement.

Texts on matters of public interest: the exception requires review and a responsible party

Paragraph 4 requires the deployer to disclose that a text has been generated or manipulated by AI when it is published to inform the public on matters of public interest. Article 50 does not specify which corporate content falls into this category: that is the first question a communications department must settle.

The exception then requires two conditions to be met together: the content has undergone human review or editorial control, and a natural or legal person holds editorial responsibility for the publication. Review without an identified responsible party is not enough, and neither is a responsible party without review. To rely on the exception, the company must therefore be able to show who reviewed the text and who takes responsibility for its publication.

The grace period ends on 2 December 2026 and covers only marking

The Omnibus added a paragraph 4 to Article 111. Providers of systems generating synthetic content that were placed on the market before 2 August 2026 “shall take the necessary steps to comply with Article 50(2) by 2 December 2026”. This four-month period covers only machine-readable marking, and only for systems already on the market. It does not extend to informing chatbot users or to deployers' obligations. For a system placed on the market on or after 2 August 2026, marking is required immediately.

The deadline is close. A company that deploys a generation tool placed on the market before 2 August 2026 would be well advised to ask its provider now how it intends to be compliant by 2 December 2026.

Article 4 has been rewritten, not deleted

The 2024 version required providers and deployers to take measures to ensure, “to their best extent, a sufficient level” of AI literacy among their staff. The new text requires them to take measures to “promote the development” of that literacy, and states that the obligation does not require them to ensure a specific level for each individual (Article 4, AI Act Service Desk). Two paragraphs have been added: the Commission and the Member States support these efforts, and the European Artificial Intelligence Board will adopt recommendations, for which no timetable has been published.

The obligation is lighter, but it remains: providers and deployers must take measures. The text does not say which measures, or how to demonstrate them. A cautious reading, put forward in particular by a Journal du Net column of 13 November 2026, comes down to three items: an inventory of AI uses (tools, teams, tasks, data), a register of training activities linked to those uses, and a named owner responsible for keeping both up to date. This is a recommendation, not a requirement of the Regulation. It has a practical advantage: the inventory also helps identify the uses covered by Article 50 and those that will fall under Annex III. It is the document that makes it possible to fill in all three columns.

Recruitment: the AI Act waits until 2027, the GDPR does not

Recruitment systems are listed in Annex III, so their obligations under the AI Act wait until December 2027. The GDPR, by contrast, already applies. Its Article 22 gives a person the right not to be subject to a decision based solely on automated processing where that decision produces legal effects concerning them or significantly affects them, and the right of access covers the existence of such a decision.

Data protection authorities do not follow the AI Act timetable. According to the CNIL, on 10 November 2026 the European Data Protection Board chose automated decision-making, including AI-assisted recruitment systems, as the topic of its sixth coordinated action, with a launch announced for 2027 (CNIL, 12 November 2026). Because participation is voluntary, the national authorities that will take part are not yet known, and the precise scope of the action has yet to be confirmed. Nothing links this action to the Omnibus. It does mean, however, that an application screening tool may be examined under the GDPR before it becomes subject to the high-risk rules. The question to document now is how much the human actually contributes to the decision.

The Commission's annual report: a signal, not a rule

On the Commission side, the annual report submitted to the Parliament and the Council by Vice-President Henna Virkkunen, covering the period from 1 August 2025 to 31 July 2026, highlights simplification and the launch of the AI Act Service Desk, according to the summary in the EU AI Act Newsletter No. 114 (Future of Life Institute, 10 November 2026). It creates no obligations. Its practical value lies in the Service Desk itself, which is the source of the versions of Articles 4 and 50 used here; the legal reference remains the text published in the Official Journal.

Three columns for classifying uses

| Topic | Already applicable | Postponed | To document now | | --- | --- | --- | --- | | Transparency (Art. 50) | Since 2 August 2026: informing people who interact with an AI system; informing people when emotion recognition or biometric categorisation is used; disclosing deepfakes and generated texts on matters of public interest; machine-readable marking for systems placed on the market since that date | Machine-readable marking for systems placed on the market before 2 August 2026: until 2 December 2026 only | List of the assistants and content concerned, the company's role (provider or deployer), proof of review and the editorial owner | | High risk (Annexes III and I) | Nothing under the high-risk rules | Annex III on 2 December 2027; Annex I on 2 August 2028 | Inventory of systems likely to fall under Annex III, to size the 2027 workstream | | AI literacy (Art. 4) | Taking measures to promote AI literacy | No postponement; recommendations from the European Artificial Intelligence Board expected, no date set | Inventory of uses, training register, named owner (cautious reading, not required by the text) | | Recruitment and automated decision-making | GDPR: right not to be subject to a decision based solely on automated processing, right of access | AI Act high-risk rules on 2 December 2027 | How much the human actually contributes to application screening, ahead of the EDPB coordinated action announced for 2027 |

The order of work follows from the table. First, review the uses covered by Article 50, starting with customer-facing assistants and with the marking question to put to providers before 2 December 2026. Next, maintain the inventory, which feeds the other rows. Finally, plan the Annex III work across 2027, as the amended timetable allows. Before this plan is finalised, three points still need to be confirmed: the text published in the Official Journal, for which the consolidated versions cited here are no substitute; the European Artificial Intelligence Board's recommendations on Article 4, along with the codes of practice provided for in Article 50(7); and the exact scope of the EDPB action, for which neither the launch date nor the participating authorities have been set.

Sources

Continue reading

View all insights