The EU is postponing part of its AI obligations through the Digital Omnibus, but that does not mean companies can relax. Some duties have long been in force, others are merely delayed. The Digital Omnibus on AI is EU Regulation (EU) 2026/1744, in force since 27 July 2026, which postpones only the timing of certain EU AI Act obligations without touching substantive chapters such as prohibited practices, risk management, data governance or technical documentation. Many companies therefore talk about the delayed AI Act while part of its obligations has been enforced for some time. Since 2 August 2026 the transparency rules apply, and supervisory authorities have started enforcement [3]. The Digital Omnibus postpones only the obligations for high-risk AI systems, meaning AI that decides on settings, loans, health or safety [9]. Reading the delay as a quiet period confuses room to move with standing still.
What does the Digital Omnibus change about the EU AI Act deadlines?
The Omnibus shifts two dates: obligations for high-risk AI under Annex III move from August 2026 to 2 December 2027, and obligations for Annex I systems move from August 2027 to 2 August 2028 [9]. Everything else stays as it was. Regulation (EU) 2024/1689 already entered into force on 1 August 2024 [10], and the amendments of 27 July 2026 affect only the timelines, not the substantive chapters on prohibited practices, risk management, data governance (i.e. rules for recording, checking and documenting training data) and technical documentation [7].
This distinction defines the character of the Omnibus. It redistributes time rather than exiting regulation. The staged approach aims to balance innovation with legal certainty and lets organisations put their management and oversight structures in place step by step [2]. Anyone celebrating an extension without checking the remaining duties has not read the substance of the regulation.
Which obligations still apply despite the postponement?
Three sets of duties remain in force. Article 4 on AI literacy, meaning the duty that staff understand how an AI system works and where its limits lie, has applied since 2 February 2025, the prohibited practices under Article 5 (such as social scoring by public authorities or AI that manipulates people's decisions) apply as well, and the transparency duty under Article 50 has applied since 2 August 2026 [9]. On top of that, another newly added ban (numbered 9) enters into force in December 2026 as part of the AI Omnibus; what concrete AI use it targets will become clear in practice [8]. The European Commission has published two documents to support the practical application of the prohibited practices [8].
The transparency duty requires companies to clearly mark content in certain cases when it was generated or substantially modified using AI. This affects AI-generated text, images and other content in particular [4]. It reaches marketing teams, product teams and anyone working with generative tools. How to implement the labelling duties in practice is set out in our guide to the Article 50 AI Act transparency duty. Article 4 on AI literacy requires that staff using AI systems have sufficient understanding of how those systems work and where their limits lie [1]. What that means for training and evidence is explained in our piece on AI literacy under Article 4 of the AI Act.
Our position: these duties affect more companies than the later high-risk rules. Everyday use of AI tools is already widespread, while genuine high-risk systems remain the exception in most mid-sized organisations. Treating the postponement as a reason to ignore the duties already in force moves a problem forward rather than solving it.
Which AI systems count as high-risk under Annex III, and who is affected?
High-risk AI covers the systems listed in Annex III as well as AI systems used as safety components of products. This includes applications in recruitment, credit scoring, healthcare, education, law enforcement and critical infrastructure [5]. High-risk applies to applications with significant impact on safety, health or fundamental rights, for example in medicine, human resources or critical infrastructure [1].
These systems must meet strict requirements for transparency, documentation, human oversight and risk management [5]. From 2 December 2027 they are subject to these duties before they may be placed on the market [8]. The regulation addresses providers, deployers, importers and distributors of AI systems used or marketed in the European Union, meaning everyone who develops, uses, imports or resells AI systems, from the manufacturer to the online shop that uses AI [1].
For most small and mid-sized companies without high-risk use, the delay changes little. For the few with high-risk use, it is preparation time [9]. Every organisation should make that determination for itself before drawing up a deadline plan. As we set out in our piece on the high-risk deadlines of the AI Act, the concrete preparation effort depends heavily on your own portfolio.
What should companies have prepared by 2 December 2027?
By 2 December 2027, companies should have three areas prepared: data governance with documented recording and checking of training data, complete technical documentation, and organised human oversight with clear escalation paths. The extra time is only worth something if it is actually used for these areas. These requirements have been known since the original version of the regulation and do not become easier through the delay [7] [5]. A sensible preparation path has four steps:
- Build an inventory of all AI systems in use and in procurement, with purpose, data categories and owners. A step-by-step guide is available in our guide to the AI inventory.
- Clarify roles between providers and deployers, because that determines which duties fall on your organisation [1].
- Establish data governance and technical documentation, making the origin, quality and processing of training and operational data traceable [7]. How this works is shown in our data governance guide for AI systems.
- Organise human oversight with clear escalation paths and named owners for edge cases [5].
These steps make sense regardless of any deadline. They improve operational quality, the traceability of decisions and the ability to react when something goes wrong. The regulatory benefit comes as a side effect.
Why is an AI inventory the prerequisite for any deadline plan?
A company that does not know today which of its systems count as high-risk under Annex III has an inventory problem, not a deadline problem. Without an overview, you cannot determine which duty applies when. The high-risk classification decides documentation duties, conformity assessment, meaning the formal check that the system meets the requirements before it may be placed on the market, and market access [5]. Without knowing the classification, you plan either too much or too little.
An inventory is more than a list. It links each system to a purpose, a data source, an owner and an assessment under the regulation. That link reveals where work is needed and where documentation is already sufficient. It is also the basis for AI literacy under Article 4, because only those who know the systems in use can target training needs [1].
We consider the inventory the single most effective measure in the entire preparation. It can start regardless of any deadline, it costs little and it provides the basis for every further decision. Anyone starting the stocktake in autumn 2027 will not have enough time left for the actual work on data governance and documentation.
What does the delay mean for mid-sized companies?
For most small and mid-sized companies without high-risk AI, the delay changes little. The duties in force now affect the broad everyday use of AI tools, not only the few high-risk cases [9] [4]. That is the central shift in perception: public debate revolves around the postponed deadlines, while the duties already in force act quietly and broadly.
Our position: the delay rewards those who use it to prepare and exposes those who mistake it for a break. A company that builds its inventory today, clarifies its roles and structures its documentation will not be rushed in December 2027. A company that trusts the delay and ignores the duties in force risks fines and reputational damage long before the new deadline.
The time gained is a tool. Whether it is worth something depends on what it is used for. We recommend starting with the inventory, because it is the prerequisite for every further measure. Everything else follows from it.