Written for Heads of Learning, Directors of Education, and Programme Managers at professional certification bodies, regulated training providers, and corporate academies who use — or are planning to use — AI somewhere in assessment or CPD, and who saw the recent AI Act headlines and wondered what they actually change. The short answer: less than the headline suggests, and not in the direction most people assumed.
On 7 May 2026, negotiators for the European Parliament and the Council reached a provisional agreement to amend the EU AI Act. The headline travelled faster than the detail: the high-risk rules, due to apply from 2 August 2026, would be deferred to 2 December 2027 — a sixteen-month slip. Much of the business press read it as "Brussels eases off on AI," and a fair number of boards heard "we have more time" and moved on.
If you run education, qualifications, or CPD at a certification body, a regulated training provider, or a corporate academy, I'd gently push you to read the delay more carefully than the headline did. It is real. It is also narrower, more provisional, and more double-edged than "you've been given breathing room" suggests. The institutions that come out of this well will treat the months ahead as runway, not reprieve.
One thing to say plainly up front: what follows is informed commentary, not legal advice, and the amendment is not yet law. Where it matters, take your own counsel. But the strategic shape is clear enough to act on now.
Sixteen months is not a pause. It is the runway you have been given to get this right before it is no longer optional.
What exactly did the EU change?
The change lives inside the Digital Omnibus on AI, the Commission's package of targeted simplifications to the Act. The most consequential piece is timing. Obligations for stand-alone high-risk systems — the ones listed under Annex III — move from 2 August 2026 to 2 December 2027. High-risk AI embedded in regulated products under Annex I moves to 2 August 2028. The agreement also swapped an earlier, conditional "we'll switch it on when the standards are ready" mechanism for these fixed dates, which is at least a gain in predictability.
The reason for the slip is not a softening of intent. It is that the scaffolding the rules depend on — harmonised standards, common specifications, and the national authorities meant to supervise them — simply isn't in place. The law was set to demand compliance against tools and guidance that don't yet exist. The delay is an admission of that gap, not a change of heart.
And here is the part the headlines skipped: the amendment is provisional. It takes legal effect only once it is formally adopted and published in the Official Journal, which is expected before 2 August 2026. Until that happens, the original timeline stands as written. Planning your 2026 on a delay that has not yet become law is a bet, not a certainty.
What did not get delayed?
A good deal. The prohibitions on unacceptable-risk practices have applied since February 2025. The obligations on general-purpose AI models have applied since August 2025. The transparency duties around AI-generated content remain largely on their original schedule.
Most relevant for our world: Article 4, the AI literacy obligation, has applied since 2 February 2025. It requires providers and deployers to make sure the people who work with their AI systems have a sufficient level of competence to do so responsibly. The same Omnibus proposes to soften how Article 4 is worded — from a duty to ensure literacy toward a duty to support its development — but softened is not switched off. The expectation that you actively build AI competence in your people survives the redraft, and 2 August 2026 remains the date on which national authorities gain formal powers to enforce what is currently on the books. The literacy duty is the deadline that wasn't moved.
Does any of this actually reach a certification body?
More than most realise. Annex III — the high-risk category that has just been deferred — explicitly covers AI used in education and vocational training, including systems that determine access to a programme, evaluate learning outcomes, or monitor and detect prohibited behaviour during an assessment.
Read that against what's happening in practice. If you are using AI to mark, to generate question banks, to adapt a learner's pathway, or to proctor an exam — or you have it on next year's roadmap — you are inside the regime that was delayed, not outside it. The deferral buys you time on a set of obligations that will still land on you. It does not lift them. December 2027 is not "never"; it is the date a body that adopts AI in assessment needs to be ready for, with the documentation, risk management, and human-oversight evidence that high-risk classification demands.
So is the delay good news, or a quiet trap?
Both, and which one it turns out to be depends entirely on what you do between now and then.
It's good news because the original timeline was unrealistic, and a scramble to comply against unfinished standards would have produced box-ticking rather than better assessment. The extra runway is a genuine chance to do the work properly.
It's a trap if you read "provisional, partial, sixteen-month delay" as "stand down." Three things make that reading dangerous. The standards your eventual compliance will be judged against are still being written, so the target is moving. Building AI into assessment in a way you can actually stand over — defensible, documented, fair to candidates — takes far longer than the runway implies once you account for procurement, piloting, and governance. And the literacy duty, the one piece aimed squarely at your staff, was never delayed at all.
The institutions that win here won't bank on a reprieve that is neither final nor complete. They'll use cheap, unhurried time now to avoid expensive, panicked time in 2027.
What should you actually do with the extra time?
Start with visibility, not policy. You cannot govern what you haven't mapped, so the first move is an honest inventory: where does AI already touch your assessment and CPD, who is using it, and on what basis? From there, a short sequence of moves is worth making while the pressure is low.
Four moves for the runway
First, map it. Inventory where AI already touches your assessment and CPD — marking, question generation, adaptive pathways, proctoring — who is using it, and on what basis. You cannot govern what you have not seen.
Second, check literacy now. The people who work with these systems need a sufficient level of AI competence. That obligation is live today, not in 2027, and it is the one that wasn't delayed.
Third, start documenting. Record how your AI-touched assessment makes decisions and how a human stays meaningfully in the loop. That evidence is exactly what the high-risk regime will ask for when it lands.
Fourth, design governance in, not on. Build the oversight into your programmes from the start rather than retrofitting it onto something already built. Compliance designed in is mostly just good programme design.
That last point is the one I'd underline. Compliance bolted on after the fact is brittle and expensive. Compliance designed in is mostly just good programme design: clear purpose, defensible decisions, evidence you can show. The runway is a chance to build the second kind.
Where LearnFrame comes in
If a structured starting point would help, the AI Use Inventory and Literacy Risk Register on our resources page is built for exactly this first mapping step, and our Programme Design Diagnostic examines where AI sits in your assessment and CPD through precisely this lens. Neither is a sales gate; both are there to make the inventory easier. LearnFrame designs and builds digital learning programmes for professional certification bodies, regulated training providers, and corporate academies — strategic direction from Dublin, production capacity through an established Cape Town team — and structures every engagement so the institute owns the outcome, the documentation, and the governance, not a permanent supplier dependency.
The delay is real. For certification bodies, it is also provisional, partial, and pointed at a regime that covers you. Sixteen months is not a pause. It is the time you have been given to get this right before it is no longer optional — and the bodies that treat it that way will be the ones quietly ready while everyone else is reading the next headline.
This article pairs with a shorter Field Notes post on the same pattern. See more insights from LearnFrame.