Article · Governance & Risk

The Digital Omnibus on AI: what changes in the AI Act, and from when

Arkatai 8 min

When someone on a board says Europe has delayed its AI regulation, they are almost always running two things together. What moved is one block of duties, the ones for high-risk systems, and it was moved by a regulation that amends the AI Act and is usually called the Digital Omnibus on AI. The rest of the regulation did not move, and the rest is the part that already applies to you. This piece is the calendar I use when I have to separate the two in front of a management team.

Everywhere else on this site I avoid quoting application dates, because the detail changes and a precision error is expensive. I quote them here because the Commission’s regulatory framework publishes them and I can show you where, and I checked the regulation’s number and publication date against the authentic Official Journal text, which I do not link because it is not one of the sources this site cites. I still do not quote article numbers, because the regulation itself is where the exact wording lives.

What the Omnibus is and where it stands

It is a regulation, not a proposal, and that is the first confusion worth clearing up, because a lot of what circulates on the topic was written while it was still in progress. The Commission adopted the proposal on 19 November 2025, and the final text was published in the Official Journal as Regulation (EU) 2026/1744 on 24 July 2026 and entered into force on 27 July 2026. Its official short name is the Digital Omnibus on AI.

What it does is amend the AI Act, not replace it, and along the way it also amends the civil aviation regulation and the machinery regulation. The AI Act is the same one, with the same risk-tier approach I describe in the EU AI Act and AI agents. What changed is a handful of dates and a handful of requirements.

The only thing that moved is the high-risk calendar

High-risk obligations now arrive in two waves. The rules for uses in certain high-risk areas, which the Commission lists as including biometrics, critical infrastructure, education, employment, migration, asylum and border control, will apply from 2 December 2027. The rules for high-risk AI systems embedded into regulated products, and the Commission’s own examples are lifts and toys, will apply from 2 August 2028.

Two caveats worth carrying into a meeting. That list comes after an “including”, so it is a selection, because the regulation also covers access to essential services, law enforcement and the administration of justice. And what got deferred is the block of classifying the system, meeting the requirements and taking on each operator’s obligations, not every rule that touches high risk.

The stated reason is that the scaffolding was missing. The Commission puts it as making sure the rules apply once companies have the right support tools, technical standards among them. The duty is the same one, with more time to get to it.

February 2025prohibitions and AI literacyAugust 2025governance and GPAI modelsAugust 2026applicable, transparency, authoritiesDecember 2026new prohibitions and markingDecember 2027high-risk under Annex IIIAugust 2028high-risk inside a productlast to arrive
The full calendar after the Omnibus, top to bottom: what has applied for years, what started in August 2026, and the only two dates the Omnibus pushed forward.

What did not move, which is almost everything that applies to you

This is where the real money gets lost, because a management team that hears “delay” and stops looking at the file ends up skipping the part that is actually theirs. According to the Commission’s framework, the AI Act entered into force on 1 August 2024 and has been applicable since 2 August 2026. Prohibited practices and AI literacy obligations apply from 2 February 2025, and that includes the duty for the people working with these systems to understand what they are handling, which I treat separately in the mandatory AI literacy obligation. Governance rules and the obligations for general-purpose models apply from 2 August 2025. Transparency obligations apply from August 2026, unchanged.

Two warnings so that “unchanged” is not read wider than it is. The AI literacy date did not move, but the Omnibus did rewrite the wording of the duty, which now speaks of taking measures to support the development of AI literacy rather than ensuring a level of it. And on transparency there is a real transitional period, four months by the regulation’s own words. Anyone who already had a system generating synthetic audio, image, video or text on the market before 2 August 2026 has until 2 December 2026 to meet the marking obligation.

And from 2 August 2026 somebody is watching. The Commission states that from that date the AI Office and Member State authorities are responsible for implementing, supervising and enforcing the regulation, and that the AI Office can request technical documentation, evaluate models, require corrective measures and issue fines.

MovedUnchangedAnnex III high-risk, to December 2027Transparency, applicable from August 2026AI literacy and GPAI models, same datesHigh-risk inside a product, to August 2028the delay reaches high-risk duties only
The two columns to keep apart in a board meeting. The Omnibus pushed two high-risk dates and moved none of the others, including the ones that already applied before 2026.

The other changes a board should know about

The deferral took the headlines, and the Omnibus carries four more things the Commission lists that will change a decision in some offices.

  • New prohibitions. The Commission presents the regulation as a list of nine prohibited practices and counts the ninth one, systems generating non-consensual sexually explicit and intimate content or child sexual abuse material, as the Omnibus’s contribution. That count is the Commission’s; in the Official Journal text there are two new cases, with conditions narrowing when they bite. They apply from 2 December 2026.
  • More weight for the AI Office. Its powers are strengthened and supervision of AI systems built on general-purpose models is centralised, with the stated aim of reducing governance fragmentation. The centralisation carries exceptions, sector-specific supervision among them, so it does not empty out the national authority.
  • Paperwork relief that reaches further down. The simplified requirements small and medium-sized enterprises already had are extended, where appropriate, to small mid-caps, simplified technical documentation among them. For those companies the regulation also caps fines at the lower of the two limits that applied to them, and gives them priority in regulatory sandboxes, including an EU-level one. If you sit in the 10 to 500 million revenue band, this is the line of the four that affects you most.
  • Less duplication with product safety. The interaction between the AI Act and EU product safety law is clarified, the Machinery Regulation in particular. For a manufacturer that already certifies machine by machine, it avoids running two files that contradict each other.

What I do with this calendar in front of a management team

First, check whether the delay even reaches you. Most of the agents I deploy in internal operations reconcile invoices, classify orders or assemble case files, and they do not fall into the high-risk areas the Commission lists. For those agents the deferral changes nothing: transparency, AI literacy and the duty to know what you have deployed still apply from 2026 and from 2025. The ones who gain real time are the ones with a use in hiring, credit access, education or critical infrastructure, and that time gets spent preparing the file, not parking it.

Next, classify anyway, whether the delay reaches you or not. The first duty is still knowing which agents you have, what they decide and about whom, because without that inventory you cannot even claim the delay applies to you. Who answers for which obligation when a provider is involved is in provider or deployer under the AI Act, and the wider frame I use to order this is in AI agent governance.

And there is a third point, about not mistaking a deadline for accreditation. An obligation arriving in 2027 tells you nothing about how you will demonstrate compliance, and a management system certificate does not settle it on its own, which is what I separate out in do you need ISO 42001 certification?. The other risks around this one, from the silent error to provider dependency, are ordered in AI risks for business, and the entry point to the whole cluster is AI agents for business.

One caveat on method before closing. The official article-by-article explorer can lag behind the amended text, something I already flag in the piece on roles, and it is worth keeping in mind before resting a specific decision on literal wording.

Frequently Asked Questions

Does the Digital Omnibus on AI delay the whole AI Act?

No. It delays the obligations for high-risk systems, in two waves, and leaves the rest of the calendar where it was. The regulation has been applicable since 2 August 2026, and prohibitions, AI literacy and the duties for general-purpose models apply from earlier than that. The only things the Omnibus adds going forward are the two new prohibited cases and the marking deadline, both in December 2026. Reading it as a general reprieve is the most expensive mistake I have seen made with this file.

From when do the high-risk rules apply to me?

It depends on the type of use. For uses in the areas the Commission lists as high-risk, including biometrics, critical infrastructure, education, employment, migration, asylum and border control, from 2 December 2027. For high-risk AI systems embedded into regulated products, from 2 August 2028.

Does a mid-sized company gain anything from the Omnibus?

Yes, and it is the least discussed line. The simplified requirements SMEs already had are extended, where appropriate, to small mid-caps, with simplified technical documentation among them, and fines for those companies are capped at the lower of the two limits that applied. For a company in that band deploying agents in its operation, that cuts paperwork without cutting the underlying duty.

If my agents are not high-risk, does this calendar affect me?

What affects you is the part that did not move, which is the part that applies. Transparency towards customers and employees, AI literacy for whoever works with the systems, the prohibitions, and an inventory and classification of your uses so you can sustain the claim that you are not high-risk. That last part is what an authority asks for first, and it does not depend on any future date.