For months, the question hanging over every EU AI Act compliance plan was whether the proposed delay to high-risk obligations would actually happen. It has. The European Parliament and Council have now signed off on the AI Act simplification package, and the high-risk deadlines are officially pushed back. If your team quietly stopped worrying about August 2026, it's worth checking exactly what moved and what didn't, because not everything did.
What's been delayed
The headline obligations under Annex III, covering AI used in recruitment, credit scoring, law enforcement, education and border control, now need to comply by 2 December 2027, not 2 August 2026. AI embedded in regulated products under Annex I, think medical devices, machinery and vehicles, moves from August 2027 to 2 August 2028.
That's real breathing room. Sixteen extra months to build the risk management system, technical documentation, data governance and conformity assessment processes that Article 16 and Article 17 require of high-risk system providers.
What hasn't moved
Article 50 transparency obligations are still due on 2 August 2026. If your product involves a chatbot, generates synthetic content, or produces deepfakes, the disclosure requirements land on schedule. Watermarking gets a modest four-month grace period on existing systems, not the wholesale reprieve some teams assumed. Bias detection powers have also been extended to cover all AI systems, not just those classed as high-risk, which quietly raises the stakes for anyone who thought they were out of scope.
The risk here isn't that teams do nothing. It's that they treat "delayed" as "cancelled" and lose the momentum they'd built.
Where the UK sits
The UK still has no standalone AI Act, and as of mid-2026 none is before Parliament. Regulation continues through existing bodies: the ICO on data protection, Ofcom under the Online Safety Act, and sector regulators applying general principles rather than a unified risk framework. It's a lighter-touch, innovation-first posture, deliberately different from Brussels.
But that doesn't mean UK teams are exempt. The AI Act's extraterritorial reach hasn't changed. If your AI system touches EU users or customers, the Act still applies to that activity, delay or no delay. UK businesses assuming distance from Brussels equals distance from the Act are making the same mistake that predates this whole omnibus process.
The practical takeaway
Use the extra time, don't waste it. The delay buys room to build proper testing, validation and monitoring infrastructure for high-risk classification, rather than a reason to shelve it. Keep transparency and watermarking work on the original August timeline. And if you're a UK team with any EU-facing AI product, map your exposure now rather than waiting for a UK equivalent that, on current form, isn't coming any time soon.
If you're mapping AI evaluation evidence against specific Act articles ahead of these deadlines, whichever ones apply to you, EvalSpec is built for exactly that.
Try EvalSpec free - evaluate your AI system across all 5 dimensions. No code required.
Start your free evaluation