Since late July 2026 the Digital Omnibus has been in force, pushing the deadline for high-risk AI systems from August 2026 to December 2027. This includes the systems HR relies on for recruitment, selection, performance reviews and other employment decisions. So, if you’ve set this aside over summer, you’ve just gained sixteen months.
Part of the regulation already applies. The prohibited practices in article 5 have applied since February 2025. The transparency obligations in article 50 have applied since August 2026, and since then national supervisory authorities have also been able to enforce the rules in practice. The AI literacy obligation remains, but the Omnibus shifts the focus from achieving a specific result to showing that you’re making the necessary efforts.
What matters more is what the delay does not change. The regulation asks for documentation you cannot pull together in the final month.
That last point is the hardest one. You can only demonstrate human oversight if you can show who made which decision at what moment, and on the basis of what information. If that lives in a mailbox or in a recruiter’s head, you have no evidence afterwards and you are left reconstructing everything.
Start with an inventory. Which systems are running, who owns them internally, which decisions do they influence, and where does their data come from? Most organisations find AI in places they never thought of as using AI, such as the screening or matching that ships by default inside their recruitment tool.
After that it becomes a conversation between HR and IT, not just an IT responsibility. HR knows which decisions get made, IT knows where the data comes from. Together you decide which decisions you are willing to hand to a model and which ones you are not.
You are better off making those choices now than in the autumn of 2027, for three reasons. First, an audit trail only starts running on the day you set it up. So if you start late, you have no history to fall back on when the deadline arrives. Also, starting early gives you time to test that trail. You see which fields are missing or which decisions never make it in, and you can still fix that now without any consequences. And finally, the recruitment tools you buy or renew today will still be running in 2027. Whether you can export and log their decisions and the underlying data is something you are best off settling in the contract now.
Sixteen months of delay change the deadline, not the amount of work. The register, the origin of your data and the evidence of human oversight get built while your HR processes run, because reconstructing them afterwards does not work.
This is regulation 3 of 3. Want the pay transparency directive and the sickness contact procedure as well, including the data questions they share?