On 7 May 2026, the Council of the European Union and the European Parliament reached a provisional political agreement on a Digital Omnibus package amending the AI Act. The headline change is a sixteen-month deferral of the Annex III high-risk obligations that were originally set to bind on 2 August 2026. They will now bind on 2 December 2027, subject to formal adoption of the omnibus by both institutions before the original deadline.
For the security and compliance teams that have been racing toward August, the deferral is real relief. For everyone else, it is also relief — but a more complicated kind. This piece is for both audiences. What was actually agreed, what was not, and what changes about the work you should be doing between now and December 2027.
What was agreed. The omnibus deferral covers two distinct categories of high-risk AI systems, on two distinct timetables.
Annex III standalone high-risk AI systems. The eight categories enumerated in Annex III of the Act — biometrics, critical infrastructure, education and vocational training, employment and workers management, access to essential public and private services and benefits, law enforcement, migration and border control, and administration of justice and democratic processes — were due to begin binding on 2 August 2026. They will now bind on 2 December 2027. This is the deferral that affects most enterprise AI deployments under the Act.
Annex I safety-component high-risk AI in regulated products. AI used as a safety component in machinery, vehicles, medical devices, and other products already subject to EU sector regulation was originally due on 2 August 2027. It will now bind on 2 August 2028. This is the deferral that affects automotive, medical-device, and industrial-machinery use cases.
Both deferrals are subject to formal adoption by the Council and Parliament. The institutions have indicated commitment to adopting the omnibus before 2 August 2026 — without that, the original dates kick in by default. Until the omnibus is published in the Official Journal of the European Union, the legal text in force still references the August dates.
What was not agreed — and what actually moved the wrong way. Two things did not change. One thing accelerated.
Prohibited practices remain in force. The Act’s prohibitions on certain AI uses — manipulative or deceptive techniques exploiting vulnerability, social scoring by public authorities, real-time remote biometric identification in public spaces with narrow exceptions — have been binding since 2 February 2025. None of this is touched by the omnibus.
General-purpose AI obligations remain in force. The obligations on providers of general-purpose AI models — including systemic-risk models above the compute threshold — have been binding since 2 August 2025. The omnibus does not defer them.
Article 50 transparency obligations were accelerated. This is the part most coverage of the omnibus has missed. Providers of AI systems that generate audio, image, video, or text content must implement transparency mechanisms — including watermarking and disclosure to deceived persons — under Article 50. The original schedule gave providers a six-month grace period after the Act’s full application. The omnibus reduces that grace period to three months and brings the binding date forward to 2 December 2026. For organisations whose AI generates media, the watermarking deadline just got tighter, not looser.
The reframe — what sixteen months actually means. There are two narratives competing for the deferral.
The first narrative is the deadline moved, you have time, slow down. This is the one most leaders will hear from vendors trying to relieve sales pressure on their compliance products, and from teams looking for legitimate reasons to defer scope.
The second narrative is the audit window opened wider, the bar will rise to fill it, and the conformity-assessment notified bodies just got busier. This is the one practitioners are quietly telling each other.
Both narratives are true. The honest read is that the deferral changes the urgency curve without changing the architectural requirements. Three substantive things are different in the deferred world.
Notified bodies have more time, which means standards will tighten. The original August 2026 deadline put intense pressure on the small number of accredited conformity-assessment bodies in the EU. With sixteen more months, those bodies have time to consult, publish detailed guidance, observe how organisations actually implement the requirements, and raise the bar accordingly. Conformity assessments approved in 2027 are likely to require evidence richer than what would have passed in mid-2026.
ISO 42001 certification becomes the practical interim bar. Organisations that need a defensible governance posture between now and December 2027 — to satisfy customers, insurers, board members, or their own risk management — increasingly point at ISO 42001 certification as the answer. The standard predates the AI Act, is internationally recognised, and provides a structured AI-specific management system. Expect ISO 42001 certifications to multiply through 2026 and 2027 as the de-facto interim governance proof.
The architectural requirements did not move. The Act’s substantive requirements — risk management under Article 9, data governance under Article 10, technical documentation under Annex IV, record-keeping under Article 12, transparency under Article 13, human oversight under Article 14, accuracy and cybersecurity under Article 15, post-market monitoring under Article 17 — are unchanged. An organisation that had been building toward August 2026 has not been wrong-footed by the deferral; it has been given more runway to execute.
What to actually do with seventeen months. For the security and compliance leaders who have been planning toward August, the next-step question is straightforward. The runway just lengthened. The work is the same.
Build the evidence model, not the documentation pile. Annex IV’s nine documentation elements split into two camps. Three are static documents — system description, design choices, intended purpose. Six need contemporaneous evidence — risk management under Article 9, data governance under Article 10, human oversight under Article 14, accuracy and cybersecurity under Article 15, post-market monitoring under Article 17, plus technical documentation under Article 11 to integrate the others. Most platforms can produce the static three on demand. Few can produce the contemporaneous six without architectural rework. The longer runway is best spent on the architectural work, not on the documentation.
Pursue ISO 42001 certification if you have not already. The standard’s Clause 6.1 documentation, Clause 6.2 measurable AI objectives, Clause 8.4 operational evidence, and Clause 10 continuous improvement give you a defensible interim governance posture that translates cleanly into AI Act conformity assessment when the time comes. Audit programmes for ISO 42001 are becoming materially more operational in 2026 — the accreditation bodies are testing operational evidence rather than just policy documentation. That tests the architecture you are building.
Do not let the watermarking deadline slip past. December 2026 is closer than December 2027. If your AI generates content — text, image, audio, video — Article 50 binds in seven months, not nineteen. The compliance window for transparency mechanisms is now three months rather than six. Organisations using AI for content generation in customer-facing contexts, particularly in marketing, communications, education, and media, should prioritise the watermarking implementation ahead of the high-risk-system work.
Track the formal adoption of the omnibus. Until the Council and Parliament formally adopt the omnibus and the text is published in the Official Journal, the original dates remain the legal position. This means the worst-case scenario — formal adoption stalls and the August 2026 dates kick in by default — is implausible but technically possible. Have a contingency plan that gets you to a compliant posture by the original dates if formal adoption fails.
For non-EU organisations. The omnibus is an EU regulation that affects organisations placing AI systems on the EU market or using them in the EU. For organisations with no EU exposure today, the deferral is informational rather than operational.
But the EU AI Act remains the most-developed comprehensive AI regulation in any jurisdiction. The architectural pattern it establishes — risk classification, documentation requirements, conformity assessment, audit trails, post-market monitoring — is the pattern other jurisdictions are watching and increasingly adopting in their own frameworks. Bill C-27 in Canada died in early 2025 and no replacement has surfaced. The Colorado AI Act was stayed by federal court in April 2026 pending litigation. The UK is taking a sectoral-regulator approach without a comprehensive AI act. None of these jurisdictions has caught up to the EU.
The pragmatic read for non-EU organisations: the EU just told you what AI governance under regulation actually looks like. The architectural choices you make now are likely to be the floor when your jurisdiction’s response arrives.
The position from here. We did not need the August deadline to do the right architectural work. We do not need the December 2027 deadline to do the right architectural work either. The reason to build evidence-rich AI governance is not that a regulator will inspect you on a date — it is that customers, partners, insurers, and your own board increasingly expect to see the evidence at any moment.
The deferral is welcome. The work continues.
Sources referenced in this article: Council of the EU press release, 7 May 2026; European Parliament press release on the AI Act omnibus deal; DLA Piper analysis; Hogan Lovells; IAPP; Travers Smith.