The AI Act, Cut in Half: What the Digital Omnibus Actually Changed on 27 July

Anyone who had circled 2 August 2026 in their calendar was overtaken on 27 July. That was the day an amending regulation entered into force and reshuffled the AI Act timetable — thoroughly enough that advisory material from spring 2026 has been wrong on material points ever since. Regulation (EU) 2026/1744, published in the Official Journal on 24 July, goes by the name „Digital Omnibus on AI”.

The result is a deadline cut in half. The transparency obligations are live; the high-risk regime is still more than a year away. For most companies this means less immediate burden, but a considerably more awkward question: how do you prepare for requirements when neither the standards nor the guidelines for them exist yet?

At a glance

• What applies now: since 2 August 2026, the transparency obligations in Art. 50 — chatbot disclosure, deepfake labelling, marking of synthetic content.
• What is still waiting: the high-risk regime has moved — Annex III to 2 December 2027, Annex I to 2 August 2028. That includes the fundamental rights impact assessment under Art. 27.
• What got lighter: Art. 4 on AI literacy has been softened, a great deal of industrial AI drops out of Annex I, and the fine cap now also covers mid-sized companies.

The new timetable

The postponement does not affect the whole regulation, but specifically Sections 1 to 3 of Chapter III — classification, requirements for high-risk systems, and deployer obligations. What applies when:

Date	What becomes applicable
02/02/2025	Prohibitions (Art. 5), AI literacy (Art. 4)
02/08/2025	General-purpose models (Ch. V), governance, penalties (Ch. XII)
02/08/2026	Transparency obligations (Art. 50), innovation support, Art. 6(5)
02/12/2026	Two new prohibitions in Art. 5; marking duty for generative legacy systems
02/12/2027	High-risk under Annex III — including Art. 26 and 27
02/08/2028	High-risk under Annex I
02/08/2030	Final date for high-risk legacy systems in public bodies

In practice: a mid-sized company using applicant screening or employee evaluation — Annex III No. 4 — now has until December 2027 for the high-risk regime. It has no time at all for Art. 50, and never had any for the prohibitions in Art. 5. That distinction is the whole point: what was postponed is a regime, not the regulation.

What genuinely got easier

Three changes provide real relief, and one of them is currently being misrepresented on a regular basis.

Art. 4 on AI literacy has been softened. The duty is now to take measures that support the development of AI literacy. And the regulation states expressly that neither providers nor deployers must guarantee any particular level of literacy in any individual. If someone is selling you a documented, auditable training level, they are arguing from the old wording. I have revised my own article on the AI literacy duty accordingly — it predates the change.

A great deal of industrial AI falls outside the high-risk bracket. The new Art. 6(1a) clarifies that AI used solely for non-safety-related purposes — user assistance, performance optimisation, service efficiency, automation, convenience, quality control — does not count as a safety component. The counter-exception applies where a failure would endanger health or safety. For manufacturing, this is the economically largest single change in the package.

The fine cap now covers mid-sized companies too. Previously only small and medium-sized enterprises benefited from the rule that the lower of the two figures applies — percentage or absolute amount. The new Art. 99(6a) extends this to the small mid-cap category. That is not an exemption, but it is relief.

One deadline, two speeds: the transparency obligations run straight through, while the high-risk regime only picks up again at the end of 2027. The mark sits at the break — the point where advisory material from spring 2026 quietly becomes wrong.

Why it was postponed — the EU says so itself

The most remarkable sentence in the package is in the recitals. The legislator names the cause as the delayed availability of standards, common specifications and alternative guidance — and the delayed establishment of the national competent authorities. Put differently: the postponement did not happen because companies were too slow, but because the regulatory machinery itself was not ready to deliver.

Two pieces of evidence for how seriously that is meant:

There are no final classification guidelines. Art. 6(5) obliged the Commission to produce guidelines on high-risk classification by 2 February 2026. What exists is a draft from 19 May 2026, roughly 148 pages, its status unchanged as „draft for stakeholder feedback”. The piquant detail: that very Art. 6(5) is expressly excluded from the postponement and has applied since 2 August 2026. The Commission’s duty is running; the delivery is missing.

There are no cited harmonised standards. The responsible standardisation body is working on drafts covering risk management and data quality, but no deliverable has yet been cited in the Official Journal. Without citation there is no presumption of conformity under Art. 40 — and therefore no reliable way to demonstrate conformity. The amended standardisation request runs until February 2027.

What this means for your company

First, re-examine whether you are affected at all. If your assessment dates from spring 2026, it rests on the old timetable and possibly on a safety-component evaluation that the new Art. 6(1a) has since disposed of. This is not a formality: the distance between „high-risk from August 2026” and „not high-risk” is measured in annual budgets.

Use the time you gained for what is actually doable, not for certificates. What already holds today is independent of the missing standards: an inventory of the AI systems in use, an AI policy, clear responsibilities, documented human oversight and — wherever personal data is involved — the data protection impact assessment, which follows from the GDPR anyway and is untouched by the postponement.

Treat 2 December 2026 as a real deadline. It gets overlooked because it sits between the headline dates. If you use generative AI in a product that was already on the market before 2 August 2026, your grace period for machine-readable marking expires on that day.

Conclusion

The Digital Omnibus is good news with an uncomfortable aftertaste. Good, because it lifts real burden from companies that would otherwise have had to certify against unfinished requirements. Uncomfortable, because the reason is the unfinished state of the regulation itself — and because postponements have a poor track record of encouraging preparation.

For practical purposes I would suggest the sober reading: the postponement is not a free pass, it is an extension of the preparation window for obligations that are coming. What applies today — transparency, prohibitions, penalties — applies in full, and is supervised by an authority that has only existed since the end of July.

If you want to know whether your spring assessment still holds, and what the postponement means for your specific case, let us talk. I read questions like these as a business lawyer, and I build the systems they are about myself.

FAQ

Has the EU AI Act applied in full since 2 August 2026?

No. What became applicable on that date were mainly the transparency obligations in Art. 50. The high-risk regime, originally scheduled for the same day, was postponed by Regulation (EU) 2026/1744: to 2 December 2027 for systems under Annex III, and to 2 August 2028 for systems under Annex I. The prohibitions in Art. 5 and the rules for general-purpose models have applied unchanged since 2025.

Do I still have to train my staff on AI, or not?

The duty still exists, but it is worded more weakly than before. After the amendment, Art. 4 only requires measures that support the development of AI literacy, and it states expressly that neither providers nor deployers must guarantee any specific level of literacy for any individual. No certification requirement follows from that. Training remains worthwhile operationally — as liability precaution, and because untrained use is the actual risk.

So how do I become AI Act compliant?

For high-risk systems, the honest answer right now is: prepare in a structured way, do not try to certify. The harmonised standards have not yet been cited in the Official Journal, so there is no presumption of conformity under Art. 40. The classification guidelines also still exist only in draft form. Anyone assessing a classification today is working from unfinished material — and should document that.

What happens on 2 December 2026?

Two things. First, two new prohibitions enter into application in Art. 5, aimed at AI systems that generate intimate depictions without consent and at material depicting child sexual abuse. Second, the grace period for legacy systems ends: anyone who had generative systems on the market before 2 August 2026 must have implemented the machine-readable marking under Art. 50(2) by that date.

---

Sources — as of 05/08/2026

• Regulation (EU) 2026/1744 (Digital Omnibus on AI), full text — https://publications.europa.eu/resource/celex/32026R1744
• ELI identifier of the regulation — http://data.europa.eu/eli/reg/2026/1744/oj
• Art. 113 AI Act timetable, consolidated view — https://ai-act-law.eu/de/artikel/113/
• AI Act implementation timeline — https://artificialintelligenceact.eu/implementation-timeline/
• Draft Commission guidelines on high-risk classification (19/05/2026, status: draft) — https://digital-strategy.ec.europa.eu/en/library/draft-commission-guidelines-classification-high-risk-ai-systems
• Commission: guidelines on transparency obligations — https://digital-strategy.ec.europa.eu/en/policies/guidelines-transparency-ai-generated-content
• CEN/CENELEC on AI standardisation — https://www.cencenelec.eu/areas-of-work/cen-cenelec-topics/artificial-intelligence/
• Gibson Dunn on the political agreement of May 2026 — https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/

This article is general information and does not constitute legal advice in an individual case. As of 5 August 2026; the legal position on the AI Act continues to move, so please check the current state before making decisions.