The AI Act Delay Is Not a Reprieve: What the Digital Omnibus Actually Changed for High-Risk AI

A DACH manufacturer has an AI-based CV-screening tool scheduled to go live in autumn 2026. In November 2025 the compliance lead built a plan around one date: 2 August 2026, when the AI Act’s high-risk obligations were set to bite¹. Then the headlines arrived — “EU delays high-risk AI rules to 2027” — and the project Slack lit up: “Great, we have another year. Park the AI Act work until Q1.”

That instinct is the single most expensive mistake an organisation can make this summer. As of 27 July 2026 the Digital Omnibus on AI is law, and it does move the high-risk dates². But a deferral of some obligations is not a reprieve from the regulation — and in a few places the Act actually became stricter. This article separates what changed from what didn’t, and translates it into what a CISO/DPO in Germany, Austria or Switzerland should be doing right now.

Key Takeaways

• It is now law. Regulation (EU) 2026/1744 of 8 July 2026 was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. The deferral is no longer a proposal or a provisional political agreement².
• The deferral is narrower than the headlines. What moved is Chapter III, Sections 1, 2 and 3 (with the exception of Article 6(5)): stand-alone Annex III systems now apply from 2 December 2027, Annex I product-embedded systems from 2 August 2028. Chapter III, Section 4 has applied since 2 August 2025 and is unaffected — as is the rest of the Act².
• The prohibitions (Art. 5), the GPAI obligations, the AI-literacy duty and the Article 50 transparency rules are untouched and live — and Art. 5 was expanded with new prohibitions applying from 2 December 2026³.
• Article 50 arrived on 2 August 2026 as planned, allocated by role: providers owe the interaction disclosure and the machine-readable marking of synthetic output; deployers owe disclosure for emotion recognition, biometric categorisation, deepfakes and public-interest text. Only the provider’s Art. 50(2) marking duty for systems already on the market gets a transition to 2 December 2026⁴.
• GDPR is a separate legal track. Your DPIA obligations under Article 35 did not move one day⁵. NIS2, DORA and sector law are likewise unaffected.
• The deferral exists because the standards and conformity infrastructure aren’t ready — not because your compliance programme is. Building one takes 12–18 months, which is roughly the headroom the delay provides⁶.

What actually changed

The European Commission tabled the Digital Omnibus on AI on 19 November 2025, the first substantive amendment to the AI Act since its entry into force on 1 August 2024⁷. After a first trilogue round on 28 April 2026 collapsed, negotiators reached a provisional political agreement in early May, announced by the Council presidency on 7 May 2026; the European Parliament’s IMCO and LIBE committees endorsed it on 2 June 2026 and the plenary gave final approval on 16 June 2026. The text was published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744 and entered into force on 27 July 2026⁸.

The headline is the timeline. The Commission’s original proposal used a conditional “stop-the-clock” trigger — high-risk rules would apply once standards and support tools were confirmed ready. The adopted text dropped that mechanism in favour of two fixed dates, and it did so by replacing a single point of Article 113⁹:

Around this core, the Omnibus bundled targeted changes: the AI-literacy duty under Article 4 was softened from guaranteeing a level of literacy to supporting its development; the legal basis for processing special-category data for bias detection was extended under a strict necessity test; the deadline for Member States to stand up regulatory sandboxes slipped to 2 August 2027; a new Article 111(4) gave providers of generative systems already on the market until 2 December 2026 for machine-readable marking; and Articles 102 to 110 were made applicable from 27 July 2026¹⁰.

Why it is not a reprieve

First: what moved is narrower than “high-risk obligations.” The amended Article 113 defers Chapter III, Sections 1, 2 and 3 — classification, the requirements for high-risk systems, and the obligations of providers and deployers — with Article 6(5) expressly excepted. Section 4 of the same chapter has applied since 2 August 2025. Reading the change as “the high-risk regime is off until 2027” is how organisations end up missing the parts that never left².

Second: the live obligations never moved. The prohibitions in Article 5 have applied since 2 February 2025, and the Omnibus adds to them, with the new prohibitions applying from 2 December 2026³. GPAI obligations have applied since 2 August 2025, and the Commission’s enforcement powers over GPAI model providers became applicable on 2 August 2026 — a date the Omnibus did not touch, which we take apart in The Date That Did Not Move. And the Article 50 transparency duties arrived on 2 August 2026 as scheduled.

That last point deserves precision, because it is the one most often garbled. Article 50 does not impose one duty on everyone. Providers of systems intended to interact directly with natural persons must design them so people are informed they are dealing with an AI system (Art. 50(1)). Providers of systems generating synthetic audio, image, video or text must mark the output in a machine-readable format, detectable as artificially generated or manipulated (Art. 50(2)) — watermarking is one technique, not the legal standard. Deployers owe disclosure only in specified cases: emotion recognition and biometric categorisation (Art. 50(3)), deepfakes, and AI-generated text published to inform the public on matters of public interest (Art. 50(4)). The new Article 111(4) transition to 2 December 2026 applies solely to the provider’s Art. 50(2) marking duty for generative systems placed on the market before 2 August 2026 — it does not postpone any deployer duty⁴. What that means, step by step, for an ordinary organisation — inventory, role, notice, publication check, evidence — is set out in Article 50 Implementation Strategy.

Third: GDPR is a parallel universe. Nothing in the Omnibus touches the GDPR. If your system processes personal data and is likely to result in a high risk to natural persons, your Data Protection Impact Assessment obligation under Article 35 is exactly where it was⁵. For high-risk AI used in credit scoring, insurance pricing or public services, the AI Act’s Fundamental Rights Impact Assessment (FRIA) layers on top of the DPIA — and the smart move is to build the two as one unified assessment now, while you have time, rather than under deadline pressure in 2027. (We mapped exactly how these two instruments interlock in DPIA vs. FRIA.) NIS2, DORA and TISAX obligations run on their own clocks and are unaffected by any of this.

Fourth: the delay is about their readiness, not yours. The co-legislators were explicit that the postponement reflects the fact that the harmonised standards, national competent authorities and conformity-assessment tooling needed to make high-risk obligations operable were not ready on the original timeline⁶. That is a statement about the regulatory infrastructure — not a verdict that organisations are over-prepared. Building a high-risk AI compliance programme — data governance, a risk-management system, technical documentation, logging, human oversight, post-market monitoring and a conformity assessment — is 12 to 18 months of work for a single system. The extra months are the runway, not the holiday.

What a DACH CISO/DPO should do this quarter

A note on scope before the actions: Germany and Austria are Member States, so the Regulation applies directly. A Swiss organisation is not covered merely by being in the DACH region — it falls within scope only where Article 2 is engaged, typically by placing an AI system on the Union market, putting one into service in the Union, or where the output is used in the Union¹¹.

Treat 2 December 2027 as the date by which the programme must be operating, then work backwards. Three concrete moves:

Inventory first. Build or refresh your AI system register, classify each system against Annex III, and record for each one whether you act as provider, downstream provider or deployer — most organisations underestimate how many internal tools (HR screening, credit decisioning, access control) already fall in scope, and almost all of them under-record the role. Second, start the assessment layer that did not move: run DPIAs now for any AI touching personal data, and where FRIA applies, draft it as the same unified document. Third, close the 2 August 2026 gaps that are genuinely live — the Article 50 duties that attach to your role, the literacy programme under Article 4, and a check that none of your generative tooling drifts into the expanded Article 5 prohibitions before they apply on 2 December 2026.

The organisations that will struggle in 2027 are the ones that read “delay” as “stop.” The Act’s architecture — risk tiers, governance, conformity assessment, penalties — is fully intact¹². The clock didn’t stop. It just told you how much time you have to do the work properly.

A deferral of the deadline is not a dismissal of the duty. Use the runway — don’t park on it.

What’s the one Annex III system in your estate you can’t yet prove is compliant? That’s where the next sixteen months should start.