EU AI Act: What Really Changed on 27 July 2026?

On 27 July 2026, the Digital Omnibus Regulation (EU) 2026/1744 took effect and genuinely changed the AI Act (Regulation (EU) 2024/1689) in two places: the training obligation became a measures obligation, and the high-risk deadlines slipped to 2027 and 2028. The Omnibus deliberately left a third building block untouched: the labelling duties under Art. 50 continue to apply unchanged, as they have since 2 August 2026. Anyone telling you the training obligation is gone entirely, or that everything has simply been pushed back, is giving you half the story.

We've sorted out the changes and attached a source to every claim — so you don't have to. Here's what actually changed, and what only sounds like it did.

The training-obligation panic meets a measures obligation

Until 27 July 2026, Art. 4 AI Act required companies to "ensure" that staff had a sufficient level of AI literacy — an obligation of result, with a measurable target. The Digital Omnibus turned this into an obligation to take measures that support the development of that competence (Art. 4 AI Act, as amended by Regulation (EU) 2026/1744). You no longer have to guarantee a particular competence level.

Sounds like relief — but that's only half the story. The obligation itself remains, and without a documented measure tailored to your actual use of AI, you still have no chain of evidence to show authorities, customers, or a court if it comes to that. Anyone who now does nothing at all is mistaking a milder duty for no duty.

The upside for you: you can scale training to how AI is actually used instead of over-training across the board. A sales team drafting quotes with a text assistant typically needs different measures than a development team using a coding assistant — you decide based on risk, document it, and you're done.

“High-risk is already in force” — no, but the run-up time is

The second change concerns high-risk systems such as applicant screening or credit-scoring tools (Annex III AI Act): their obligations now only kick in from 2 December 2027. For AI embedded in products — medical devices or machinery, for example (Annex I AI Act) — the date moves to 2 August 2028. The grandfathering rules for systems already in operation are tied to the same dates (Art. 111, 113 AI Act).

Postponed is not cancelled. Anyone planning or operating a high-risk system today has more time — not less obligation. The 2027 and 2028 deadlines will arrive exactly as surely as 2 August 2026 did.

The benefit of the postponement is real if you use it: you can build a high-risk system properly from the start, instead of retrofitting it at the last minute. That is, as a rule, cheaper and considerably less stressful than any last-minute scramble.

The false all-clear: Art. 50 has applied for a while

Unaffected by the postponement are the transparency obligations in Art. 50(1)–(6) AI Act: labelling chatbots, AI-generated content, and deepfakes. They have applied unchanged since 2 August 2026. The only transitional period: providers who already had an AI system in use before that date have until 2 December 2026 to retrofit machine-readable labelling (Art. 111(4) AI Act).

This transitional period is routinely overstated in the market, so here it is precisely: it applies only to providers, only to machine-readable labelling under Art. 50(2) — and only to systems that were already placed on the market before that date. If you, as a deployer, publish a deepfake or AI-generated text on a matter of public interest, the disclosure duty under Art. 50(4) AI Act applies with no grace period at all.

For context, not as a threat: violations of Art. 50 can be fined up to €15 million or 3% of worldwide annual turnover (Art. 99(4)(g) AI Act), and prohibited practices up to €35 million or 7% (Art. 99(3) AI Act). Almost everywhere, these figures get cited without the sentence that matters most for smaller companies: for SMEs and start-ups, the lower of the two values applies, not the higher one (Art. 99(6) AI Act) — and since the Digital Omnibus, the same rule applies to small mid-cap companies (Art. 99(6a) AI Act). These remain ceiling figures for worst-case scenarios, not a forecast for your company.

We're seeing both extremes in the market right now: providers still advertising the old, stricter version of Art. 4 (“you MUST ensure!”), and providers giving a blanket all-clear because “everything's been postponed”. Neither is true — and following either one means doing too much or too little. Checking what actually applies saves real money either way.

What this means for your company in practice

Four to-dos follow from 27 July 2026: First, check or build your AI inventory — which tools are actually in use at your company. Second, define a suitable, documented competence measure for each use case (Art. 4 AI Act, as amended). Third, check whether your own chatbots or AI content carry the labelling required under Art. 50 — retrofit legacy systems by 2 December 2026. Fourth, if you have high-risk plans: keep the 2027 and 2028 deadlines on your radar instead of filing them away.

The legal position as of 27 July 2026 is nowhere a reason to relax, and nowhere a reason to panic — it's a reason to check exactly what applies to your specific case. Our 30-second check shows you where you stand, with today's legal position — not the one from before the Omnibus amendment.