The Digital Omnibus did not “gut” the AI Act, and it did not leave it untouched. It moved a specific set of dates and left another set exactly where they were — and a programme that reads the headline instead of the diff will misallocate the next eighteen months.
Most of the coverage of the Omnibus fell into one of two errors. Either it reported “the AI Act has been delayed” as though the whole instrument had been pushed back, or it reported “nothing has really changed” because the prohibitions and general-purpose AI rules were untouched. Both are wrong in ways that cost money. The first tells a team to stand down obligations that are still live. The second tells a team to keep building conformity machinery for duties that have genuinely slipped by more than a year. What the technology function actually needs is a clean before-and-after, so the roadmap re-baselines against the real change and not the mood of the reporting.
The Commission proposed the Digital Omnibus on 19 November 2025. The Council and Parliament reached political agreement on 7 May 2026 and the co-legislators subsequently adopted the text; treat the numbering and any last-minute drafting against the Official Journal version as final, but the substance below is settled. This is the diff.
Free · 4 minutes
Do you know where AI is already being used in your business — and what it can see?
Fourteen questions on shadow AI, data exposure, oversight, and governance debt — the gap between how fast AI is arriving and how much control you have over it. Banded finding on screen, full sheet by email.
What actually moved
High-risk obligations for Annex III systems. The use-case-based high-risk category — the one that catches recruitment, credit scoring, biometric categorisation, education, essential-services access and the rest of the Annex III list — was due to apply from 2 August 2026. It now applies from 2 December 2027. That is a slip of roughly sixteen months on the single obligation most in-scope firms had been racing towards.
High-risk obligations for Annex I product systems. Where an AI system is a safety component of, or is itself, a product already covered by the EU harmonisation legislation listed in Annex I — machinery, medical devices, lifts and the like — the high-risk rules were due from 2 August 2027. They now apply from 2 August 2028.
Regulatory sandboxes. Member States were to have AI regulatory sandboxes operational by 2 August 2026. That obligation moves to 2 August 2027, which matters if your plan depended on supervised testing being available before you deploy.
Those three are the real deferrals. Everything a supplier told you about “the AI Act being pushed back” should reduce to those, plus the content-marking grace period below. If your programme’s critical path ran through Annex III conformity for an August 2026 date, you have genuinely bought time — the risk-management system, the technical documentation, the quality management system and the conformity assessment now have a later, firmer deadline to hit.
What did not move — and still bites
The prohibitions. The Article 5 banned practices have applied since 2 February 2025. The Omnibus did not touch that date. If your estate contains anything resembling untargeted facial-image scraping, workplace or education emotion inference, social scoring or manipulative subliminal techniques, it was unlawful eighteen months ago and remains so. There is no deferral to lean on here.
General-purpose AI. The obligations on GPAI model providers have applied since 2 August 2025. If you build on, fine-tune or distribute a general-purpose model, the documentation, copyright-policy and — for systemic-risk models — evaluation and incident duties are already in force. The Omnibus clarified the AI Office’s competence over GPAI-based systems; it did not postpone the underlying obligations.
Transparency under Article 50. This is the one that catches teams out. The transparency duties — telling people they are interacting with an AI system, and marking synthetic audio, image, video and text as artificially generated — still apply from 2 August 2026. The only relief is a short grace period: systems already on the market get until 2 December 2026 to bring content-marking into line. So the obligation most likely to touch a customer-facing product did not move to 2027 with the high-risk rules. If you had bundled “AI transparency” into your high-risk workstream and re-planned it for late 2027, you have just created a gap.
A new prohibition. The co-legislators added a ban on AI systems that generate child sexual abuse material or non-consensual intimate imagery of an identifiable person, applying from 2 December 2026. It is narrow, but it is new, and it is a prohibition rather than a high-risk duty — meaning no conformity route, just a hard line.
The trigger that quietly got replaced
The Commission’s original November proposal did something more interesting than moving dates: it tried to make high-risk applicability conditional on the supporting standards and tools actually being ready — a “standards-readiness” trigger rather than a calendar date. The co-legislators rejected that mechanism and went back to fixed dates. This is the detail that should change how you plan. A conditional trigger would have left you unable to name the day you had to be ready. Fixed dates give you 2 December 2027 and 2 August 2028 to work back from, which is exactly what a delivery plan needs. Do not build your programme around the assumption that the deadline will drift again if the standards are late; the legislators just chose certainty over that flexibility.
Re-baselining without over-correcting
The re-plan is narrow and specific. Hold every transparency deliverable on the 2 August 2026 line and treat 2 December 2026 as the last date to retrofit content-marking on anything already shipped. Keep the prohibition controls exactly where they are; nothing there earned a reprieve. Confirm GPAI obligations are being met today, not next year. Only then move the Annex III high-risk workstream — the risk-management system, the quality management system and the conformity assessment route — to the 2 December 2027 date. The mistake to avoid is treating the whole file as deferred and letting the retained obligations rot for a year while the team relaxes. For the wider technology-function checklist against these deferrals, we set out what still has to be done before 2 August 2026 separately.
The Omnibus did not buy you a quiet year. It bought you a later deadline on one workstream and left three others precisely where they were. Re-plan the one, and guard the three.
Most technology problems are not technology problems. They are control problems.
The systems exist. The investment has been made. The question is whether leadership can understand, direct, evidence, and sustain what those systems produce. Find out where control exists — and where it only appears to.
Full Governance by Sixteen Pillars
Govern your business. Prove your compliance.
A board assurance cockpit for EU-regulated financial firms — tamper-evident, hash-chained proof of governance across DORA, GDPR, NIS2, ISO 27001, the EU AI Act and MiCA. In development.
See what's coming