August 2026: the date everyone mentioned, and what actually changed

Updated 15 August 2026. This article has been revised in light of the Digital Omnibus on AI. It was adopted by the European Parliament on 16 June 2026 and by the Council on 29 June 2026, published in the Official Journal of the EU on 24 July 2026 as Regulation (EU) 2026/1744 and entered into force on 27 July 2026. Through the Omnibus, the application of the high-risk AI requirements moves from 2 August 2026 to 2 December 2027 (stand-alone Annex III systems) and 2 August 2028 (AI in regulated products). The AI literacy obligation of Article 4 has applied since 2 February 2025. The transparency obligations of Article 50 remain applicable from 2 August 2026, with extra time until 2 December 2026 for certain technical marking obligations for systems already on the market. This article explains what that difference means.

The date that popped up everywhere, and was rarely explained well

If you have read anything about the EU AI Act in recent months, one date will have caught your eye: 2 August 2026. In webinars, in newsletters, in LinkedIn posts, everywhere that day was marked as the moment when "the AI Act really takes effect". That was already misleading at the time, and by now it is out of date. The AI Act has been in force for a long time. And the heaviest obligations, those for high-risk AI, have been moved to later by the Digital Omnibus. So what changed around 2 August 2026 was something different and narrower than the story that was going round. Understanding this distinction is vital before basing your strategy on it.

What is already in force (and since when)

For an honest picture, it helps to set the timeline straight:

In other words: the basic obligation around the AI literacy of your staff has already applied for more than a year. What demonstrably meeting Article 4 requires in practice, you can read in a separate insight. Anyone still waiting to start is already behind schedule. What the law itself asks is set out on our pillar page about Article 4.

What the Digital Omnibus has changed

For a long time, 2 August 2026 was known as the day the high-risk AI requirements would start to apply. That is no longer the case. Because the supporting standards, guidelines and national supervisory structures were not ready in time, the EU legislators have moved the application dates for high-risk AI through the Digital Omnibus on AI:

Important: the Omnibus was adopted by Parliament (16 June 2026) and the Council (29 June 2026), published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. The new dates are therefore legally binding. We set out separately what else the Omnibus did and did not change in the AI Act.

What did change around 2 August 2026

The postponement of the high-risk AI requirements does not mean 2 August 2026 was a non-event. A number of things did stay on that date:

There is a lot of confusion about the fines, so to be clear: the highest fine ceiling, up to 35 million euros or 7% of worldwide annual turnover, applies only to the prohibited AI practices of Article 5. For most other obligations, the ceiling is 15 million euros or 3%. Article 4 has no specific EU-level fine ceiling of its own in Article 99. Member states may still attach national enforcement measures to infringements, and a lack of AI literacy can weigh negatively when another AI Act infringement is assessed. For SMEs and startups, moreover, the lower of the two amounts always applies, not the higher. Our free EU AI Act Compliance Checklist takes you through what "demonstrable" means in practice.

What this means if you do not use high-risk AI

A common misconception: "We do nothing with high-risk AI, so the AI Act is not relevant to us." Incorrect. Even if your organisation only uses minimal-risk or limited-risk AI (Copilot, ChatGPT, translation tools, summarising assistants), Article 4 and, where applicable, Article 50 apply. A supervisory authority can ask: show how you support the AI literacy of your people. That obligation is already there and does not disappear with the Omnibus. We calculated separately what training delivers beyond the law. If you already use AI but get too little out of it, Effective AI at Work was made for that.

In concrete terms: if a supervisory authority comes knocking and asks about your approach, you must be able to show who followed which training, when, at what level, and with what assessed result. An internal lunch session or a circulated PDF is weak evidence. A demonstrable, structured path with certification that links the learning outcomes to the relevant articles of the regulation is not.

What this means if you do use high-risk AI

For organisations in recruitment, education, healthcare, local services or critical infrastructure, the high-risk deadline has moved to 2 December 2027 (or 2 August 2028 for AI in regulated products). That sounds like breathing room, but it is a postponement, not a cancellation. On that day you must meet all the high-risk AI obligations: not only AI literacy, but also risk management, data quality checks, human oversight, logging and documentation. You do not build that up in a few weeks. The extra time is precisely meant to do this carefully rather than in a rush. If you work in healthcare, education, public services or critical infrastructure, we have in-depth modules with the regulation and the scenarios of that sector. It is sharpest for recruitment and selection because those tools fall under Annex III, and we set out separately what that means for HR.

The right approach: start with the foundation. AI literacy of all the staff involved (general staff, management, and the specialist roles that work with the high-risk systems) is the base that all the other obligations rest on. Without literate staff, human oversight cannot work, logging cannot be properly interpreted, and a supervisory authority cannot have confidence in the organisational control.

What you can realistically do now

The deadline for the heaviest requirements is further away, the task stays the same. Those who use the time until 2027 rather than wait will be in a strong position later:

In closing

2 August 2026 was never the big bang it was made out to be, and the Digital Omnibus has made that even clearer. The law has been working since 2024. The heaviest requirements have moved to 2027 and 2028. What remains is this: the evidence counts. Until recently you could still say "we are working on it". More and more, the question becomes: "show what you have done." Anyone left empty-handed then risks not only a penalty, but also reputational damage that lasts far longer. We track what changes around the AI Act in the meantime every working day in the EU AI Act Monitor.

The good news: there is more time than thought to get the basics in order. That time is meant to be used, not to be spent waiting.

At AIAdopt, we have developed an audit-ready microtraining programme for this. The base package covers general staff, management, HR and IT, with sector extensions for healthcare, education, local government and critical infrastructure. Every training ends with an assessed certificate that lists the relevant AI Act articles, exactly the kind of evidence a supervisory authority will ask for.

👉 View the base package or build your own package for your organisation. Or simply get in touch.

Written by Rob Ummels in collaboration with Claude (Anthropic). Editorial responsibility: AIAdopt.