If you are building anything that puts a learned policy in charge of a safety function, the date you need is 20 January 2027. A lot of secondary sources say 14 January. They are reading the original text. Regulation (EU) 2023/1230 as published did say 14 January. The Corrigendum of 4 July 2023, OJ L 169, corrected the dates in Articles 6, 47 and 50 to 54. The consolidated text applies from 20 January 2027. Six days is not much, but if you are planning a conformity assessment around a date you got from a blog post, it is worth reading the corrigendum yourself rather than taking mine.
What actually changes for software people: Annex I pulls safety components with self-evolving behaviour into conformity assessment. If your keep-out zone is enforced by a model rather than by a relay, that is in scope.
What does not change, and this is the part I wish more vendors said out loud: nothing requires you to red-team a robot policy today. The EU AI Act's Article 15 obligations for embedded AI are deferred to 2028. No insurer I have found asks for it. The market is pre-budget.
So the honest reason to do any of this now is that it is cheaper before the clock runs out than after, and it is still optional. Not because you are non-compliant. You are not.
We publish the clock, with the primary sources for each date and the dates we have got wrong, here: https://www.provael.com/regulatory-clock/
I am not a lawyer, and this is not legal advice. Read the corrigendum.