TL;DR
Use the extra time to build the documentation and controls you'd want anyway, not to stand still.
What happened
In May 2026, the European Council and Parliament reached a provisional deal on the AI Omnibus, a set of amendments to the EU AI Act. Parliament voted for it in June, and by July 1 law-firm briefings reported it adopted by both institutions, pending publication in the Official Journal.
The key changes:
- Stand-alone high-risk systems, such as AI used in HR and employment, move from August 2, 2026 to December 2, 2027.
- High-risk AI embedded in regulated products, such as AI-enabled medical devices, moves to August 2, 2028.
- Transparency and watermarking rules apply from August 2, 2026, with a grace period to December 2, 2026 for systems already on the market.
- A new ban, from December 2, 2026, on AI systems that generate child sexual abuse material or non-consensual intimate content.
- A narrower test. Features that only assist users or optimize performance won't automatically count as high-risk if their failure poses no health or safety risk.
- AI literacy shifts from a duty on providers and deployers to one the Commission and member states are to support.
So what for your business
A delay is not a repeal. The obligations are the same, with more time to meet them. Teams that treat the new date as the starting line will be rushing in 2027 along with everyone else.
If you build or use AI for recruitment, screening or workplace decisions, you're in the first high-risk group. That means risk management, data quality, logging, human oversight and documentation, all of which take time to build into a system properly.
The narrower scope is worth checking. Some tools you assumed were high-risk may not be, and some you assumed weren't may still be. Check that against the final text.
What to do this month
- Classify every AI system you build or use. Is it high-risk, limited-risk or neither, under the final text?
- Mark the hiring and HR ones. They're the clearest high-risk case and the ones to plan first.
- Start the paper trail. Intended purpose, training and test data, known limits, how a person can override it.
- Set up logging now. Retrofitting it later is far more expensive.
- Check watermarking duties if you generate text, images or audio, and note December 2, 2026.
Source: Sidley Austin LLP, June 22, 2026
Source: Addleshaw Goddard LLP, July 1, 2026. At the time of that briefing the text was awaiting publication in the Official Journal. Check the published text for final wording.
More in Industry
Want help applying this?
We build AI-native software for teams navigating exactly this kind of shift.
Start a conversation