← All articles

AI Act Omnibus Talks Stalled: The August 2026 Deadline Still Stands

· Updated
The 29 April 2026 trilogue between the Parliament, Council, and Commission ended without an agreement on the AI Act Omnibus, leaving the original 2 August 2026 deadline for high-risk obligations in force.
The 29 April 2026 trilogue between the Parliament, Council, and Commission ended without an agreement on the AI Act Omnibus, leaving the original 2 August 2026 deadline for high-risk obligations in force.

Update — 11 May 2026: The resumed trilogue produced an agreement. The Omnibus deal is done and the 2 December 2027 deadline for Annex III high-risk obligations is confirmed, pending formal adoption and publication in the Official Journal. The analysis below reflects the legal position as of 6 May, when the deadline still stood. See the follow-up: AI Act Omnibus Agreed: 2 December 2027 Deadline Confirmed.

In April, the Parliament’s vote to delay the high-risk AI deadline to December 2027 gave teams an obvious reason to slow down. The vote was 569–45. The amended dates looked locked in. Much of the press treated the delay as a done deal. It is not a done deal.

On 29 April 2026, after roughly 12 hours of talks, the trilogue between the Parliament, Council, and Commission broke down without an agreement. Negotiations are scheduled to resume in mid-May, but the original AI Act timetable is still law. The 2 August 2026 application date for high-risk systems, and for Article 50 transparency obligations, is what is legally binding this morning. Teams that put the compliance work down in April have burned a month and have under three months to ship.

What happened on 29 April

The Commission proposed the AI Omnibus simplification package on 19 November 2025. The Parliament adopted its negotiating position on 26 March 2026 — 569 in favour, 45 against, 23 abstentions. The proposed enforcement-date changes were:

ObligationCurrent lawParliament’s position
Annex III high-risk systems2 August 20262 December 2027
Annex I sectoral safety-component systems2 August 20272 August 2028
Article 50 watermarking for synthetic content2 August 20262 November 2026

The 29 April trilogue was meant to bridge the Parliament’s position with the Council’s. Twelve hours of negotiation produced no agreed text and no commitment beyond a return date in May. The published readouts from civil-society observers, industry groups, and the institutions themselves agree on one thing: there is no Omnibus deal.

Why the talks broke

The headline reason was not the delay dates, on which there was broad alignment. The breaking point was a sectoral carve-out for AI embedded in regulated products — medical devices, machinery, toys, connected vehicles, radio equipment — where the Act’s horizontal obligations would overlap with sector-specific safety law.

The Parliament pushed for a broader exemption: where existing sectoral safety law already covers an AI component, the AI Act’s high-risk obligations should apply more lightly, or not at all on top of the sectoral regime. The Council and Commission resisted, arguing that routing AI governance through sectoral law would, in practice, deregulate AI in those sectors rather than simplify the overlap. Civil-rights organisations sided with the Council, warning that broad sectoral carve-outs could weaken the Act’s biometrics, education, and law-enforcement protections.

This is a structural disagreement, not a drafting one. It will not be resolved by editing margins. The May trilogue is going to need a political decision on how the AI Act and EU sectoral safety regimes coexist, and there is no obvious midpoint.

What this means legally

The Omnibus is a proposal. The Parliament’s position is a proposal. Until the trilogue concludes and the amended text is adopted by Parliament and Council, signed, and published in the Official Journal, the original AI Act text governs. That text requires:

Enforcement attaches to the law as it stands on 2 August. The proposed text is not enforceable. National competent authorities will not enforce against a draft amendment that has not been adopted, and they will not be inclined to soft-pedal against the law that has been. Companies treating December 2027 as their effective deadline are accepting an enforcement risk that has no legal basis until the Omnibus completes.

Industry pressure is not law

On 5 May 2026, the chief executives of seven major European technology companies — Airbus, ASML, Ericsson, Mistral AI, Nokia, SAP, and Siemens — published a joint open letter urging EU policymakers to revise their AI regulatory approach, calling for “flexible guardrails” rather than rigid regulation ahead of the resumed trilogue. The TÜV Association and other testing bodies pushed back, arguing that dismantling the AI Act’s horizontal framework would create a regulatory patchwork and weaken European standards globally.

The pressure may shape the outcome of the May trilogue. It does not change what is legally in force. A compliance programme planned around an open letter is not a compliance programme.

The realistic timing

Even if the May trilogue produces an agreement, the new dates do not take immediate legal effect. The remaining sequence is roughly:

  1. Trilogue agreement — political deal between negotiators
  2. Final text drafted; lawyer-linguist review; formal versions prepared in 24 EU languages
  3. Parliament plenary vote on the agreed text
  4. Council formal adoption
  5. Signature by the Presidents of Parliament and Council
  6. Publication in the Official Journal
  7. Entry into force on the date specified by the Omnibus, typically with a short transitional window

Compressed legislative schedules can move through this in weeks; more usual schedules take months. The window between an agreed Omnibus and 2 August 2026 is narrow even on the optimistic timeline. If the May trilogue fails again, or if the sectoral disagreement requires a fresh Commission proposal, the original 2 August date will arrive before any delay does.

There is also a practical problem: an Omnibus that retroactively extends the high-risk deadline only works cleanly if it is in force before 2 August 2026. The closer the adoption gets to that date, the messier the regulatory communications get for the days and weeks of overlap. This is solvable, but it is the kind of solvable that produces hedged guidance for compliance teams rather than clean reassurance.

What teams should do now

The position to take is the same as before April:

Plan for 2 August 2026. That is the only date that is legally binding. Move it later only when the Official Journal does.

The Omnibus does not change the work. Risk management under Article 9, logging under Article 12, human oversight under Article 14, conformity assessment, and technical documentation under Annex IV — none of these requirements are simpler in the Parliament’s position than in the current text. A 16-month delay buys time. It does not reduce scope.

Article 50 transparency is the most exposed. The Parliament wanted only a three-month delay (to 2 November 2026) for synthetic-content marking. Even if the Omnibus passes intact, the practical engineering window for Article 50 barely shifts. Channel-level disclosure, machine-readable provenance, and biometric notification need to be design-complete by mid-2026 either way.

Do not restart compliance work. The teams that paused in April will need to recover the cost of paused momentum on top of the underlying engineering. The cleanest path through this period is to assume the original deadline holds, ship against it, and treat any eventual delay as headroom for refinement rather than as a reason to slow down.

The nudifier prohibition is not yet law either. The proposed Article 5 addition banning AI systems that generate non-consensual intimate imagery resembling identifiable real persons sits inside the same Omnibus and shares its fate. If you operate in or near AI image generation, the prohibition is still pending — but the broader prohibited practices regime that has been in force since February 2025 already constrains adjacent uses.

2 Aug 2026 Deadline stands — for now.

The Parliament voted overwhelmingly to delay. The Council was willing to discuss it. The Commission tabled the proposal. And on 29 April, after 12 hours of talks, none of that produced an agreed text.

Until the trilogue concludes, the August 2026 deadline is what the law requires. Penalties for non-compliance accrue against the law as it is currently written.

Free Resource

Free EU AI Act Priority Checklist

The 5 most critical compliance items before the August 2, 2026 deadline. Delivered to your inbox.