AI literacy is the part of the Act worth running toward.

For two years, the EU AI Act was sold as a countdown: get ready before it arrives. The countdown has since been reset more than once and as of this spring the obligations everyone braced for sit comfortably in 2027 and 2028. That sounds like a reprieve and in a narrow sense it is. But it also misses the more interesting story, which is that the single part of the Act most worth your attention has been quietly sitting there, in force, the whole time.

01Why the deadlines moved and why it is encouraging.

The high-risk obligations, the ones written about most, were originally due in August 2026. They have now been deferred to December 2027 for standalone systems and August 2028 for AI built into regulated products, as part of a wider simplification package the institutions provisionally agreed in May 2026.¹ It is worth understanding why, because the reasons are genuinely encouraging rather than a sign of retreat.

Part of it is simply practical. The harmonised standards and the national authorities that are meant to make high-risk obligations measurable were not ready. Asking organisations to comply with rules whose measuring stick does not yet exist helps nobody. The rest is strategic and Europe has been open about it. The Draghi report put the productivity gap with the United States largely down to technology. It noted that only a handful of the world's leading tech companies are European. More than forty European chief executives, from ASML to Siemens to Mistral, asked for time to implement the rules well. The conclusion the institutions reached was that the ambition to set the global standard for trustworthy AI and the ambition to build a competitive industry should pull in the same direction, not against each other. A text written against the AI of a few years ago is being tuned to govern the AI of today. That is the system working, not failing.

In force now

• Prohibited usesFeb 2025
• AI literacy · Article 4Feb 2025
• General-purpose modelsAug 2025

Enforceable soon

• Article 4 supervisionAug 2026
• Transparency & markingDec 2026

Moved to later

• High-risk · standaloneDec 2027
• High-risk · in productsAug 2028

02The one obligation that did not move.

One obligation sat out all of this movement. Since February 2025, the Act has expected the people who work with AI on your behalf to have a genuine, working grasp of what it is doing. That is Article 4. It survived almost untouched: the wording was softened a little, from ensuring a level of literacy to taking measures to support its development, but the obligation itself stayed put. From August 2026 national authorities can supervise it.² But the date was never the point.

Here is the genuinely good news. It is also the reason we like this corner of the Act more than any other. Literacy is the one requirement you would invest in even if it had never been written down. Every other obligation is something you do for the regulator. Literacy is something you do for your own people; the regulator simply happens to approve. When the team using AI actually understands where it helps, where it does not and how to tell a good answer from a confident wrong one, the tools stop being a quiet source of anxiety and become a genuine multiplier. The fluency is the point. Meeting Article 4 simply falls out of it.

03Build the part that pays for itself.

This is also the part of the work we care about most. A decade of running AI in regulated production taught us that the difference between a deployment that thrives and one that quietly stalls is almost never the model; it is whether the people around it know how to work with it. A company that builds with AI every day cannot treat literacy as someone else's concern. We take it seriously enough to hold the relevant certifications across AI strategy, governance and fluency. We run that training ourselves. The same logic holds for any organisation putting AI in front of its people: fluency is the foundation, not the finishing touch.

And while the high-risk dates keep settling, literacy is the move available right now. It is the cheapest and highest-leverage thing on the list. It is also the rare obligation that leaves an organisation better at AI rather than merely compliant. Whenever the rest of the Act finally lands, the people using your systems already understand what the rules are about.

The deadlines are weather. Fluency is the thing you actually get to keep.

QUESTIONS ON THIS PIECE

What readers tend to ask.

01What part of the EU AI Act actually applies right now?

Three things are already in force: the ban on prohibited uses and the AI literacy obligation, both since February 2025; the rules for general-purpose models, since August 2025. The big high-risk obligations everyone watched have been pushed well into 2027 and 2028, so a good deal of the Act is live and a good deal is still being settled.

02Why have the high-risk deadlines moved again?

For sound reasons. The harmonised standards and national authorities meant to make those obligations workable were not ready and Europe has openly chosen to align the timeline with its competitiveness ambitions rather than push ahead with rules no one could yet comply with cleanly. A fast-moving field is being given rules it can actually meet. As of spring 2026 the changes are provisionally agreed and expected to be formally adopted before the summer.

03Did Article 4, the AI literacy obligation, change too?

It was kept, not deferred. The wording was softened slightly, from ensuring a level of literacy to taking measures to support its development. From August 2026 national authorities can supervise it. But the date was never the point. Literacy is the one obligation you would invest in even if no law had ever asked for it.

04So what is worth doing now, while the rest settles?

Build AI literacy across the people who use and govern your AI. It is the cheapest, highest-leverage move on the list, it is already required and, unlike everything else, it makes your organisation better at AI rather than simply compliant. When the high-risk obligations finally land, your people already understand the systems the rules are about.