The high-risk deadline moved to December 2027, but August 2, 2026 is still live. Here's what still applies in August, what moved, and why the inventory is the real deadline.
Update, June 30, 2026: Since this article was published, the Council formally adopted the Digital Omnibus on June 29, 2026, following the European Parliament's approval on June 16. The high-risk deferral to December 2027 described below as still pending is now confirmed. It is not yet in force: the amendment awaits publication in the EU's Official Journal and enters into force three days after that publication. The August 2, 2026 transparency obligations under Article 50 are unaffected and remain live.
In May, EU negotiators agreed to push the AI Act's hardest obligations, the ones covering high-risk systems, back to December 2027. The headlines called it a reprieve. I think that reading gets most companies into trouble, because it answers the wrong question. The question was never how much time you have. It is what you can show when someone asks.
Start with what did not move. August 2, 2026 is still a live date. From then, the Article 50 transparency rules apply: you have to tell people when they are dealing with an AI system, label synthetic media, and disclose deepfakes. Enforcement powers over general-purpose AI models switch on too, though the obligations on those models have been in force since last August. The penalties are real and already written down.
What moved is the high-risk tier. Standalone high-risk systems under Annex III, things like AI used in hiring, credit decisions, and worker management, now point to December 2027. High-risk AI built into regulated products lands in August 2028. That is genuine relief, and it is honest relief, because the technical standards needed to comply on the old schedule were never going to be ready in time.
But read the fine print, because almost nobody is. The new dates only become law once the amendment is published in the EU's Official Journal, and that has not happened yet. Until it does, the original timeline is still what is on the books. A company that has already penciled in December 2027 as settled is planning around a proposal, not a regulation. What I tell clients is plain: keep preparing to the original date, and if the extension lands, treat it as margin you did not expect rather than the thing the whole plan rests on.
There is also one deadline almost everyone is ignoring. If you put any generative AI feature into the European market, you need machine-readable labeling of AI-generated content working by December 2, 2026. That one is an engineering job. It sits with your product team, not your lawyers, and it is months away.
Now the part that matters most. The hard part of this law was never the paperwork. It is the inventory. Most organizations genuinely do not know how many AI systems they are running, who owns them, or which ones touch a decision the Act treats as high-risk. Building that picture, and keeping it current as teams ship new things, is the actual work. It is slow and unglamorous, and it does not get easier by waiting. None of it depends on the final standards being published, which means you can do almost all of it today.
The companies that spend the next year and a half on that inventory will walk into 2027 with the deadline already handled. They also get something more useful than compliance out of the exercise. They get evidence. A buyer running a regulated procurement now wants proof that the AI inside your product is governed, not your word that it will be. In an acquisition, AI governance has started showing up in due diligence, and the target that can hand over a clean inventory and real documentation holds its price, while the one that cannot becomes a line the buyer marks down. Readiness turns into something you can sell, not just a cost you carry.
So the deadline softened and the work stayed exactly where it was. The companies treating these months as preparation will be ready. The ones treating them as time off are making a decision they will not feel until it is too late to change it.