An earlier piece on this site ends on 2 December 2027. That date is correct, and it is not the whole answer. It concerns the obligations for high risk systems, meaning Chapter III of the AI Act. One obligation sits in Chapter I, knows no risk class, and has been running since 2 February 2025.
Where the postponement bites, and where it does not
Regulation (EU) 2026/1744 of 8 July 2026, published in the Official Journal on 24 July 2026, amends among other things Article 113 (3) of the AI Act. Point (c) is replaced by the following:
„Kapitel III Abschnitte 1, 2 und 3, mit Ausnahme von Artikel 6 Absatz 5, gelten ab dem: i) 2. Dezember 2027 in Bezug auf KI-Systeme, die gemäß Artikel 6 Absatz 2 und Anhang III als hochriskant eingestuft sind, und ii) 2. August 2028 in Bezug auf KI-Systeme, die gemäß Artikel 6 Absatz 1 und Anhang I als hochriskant eingestuft sind."
The same amending regulation, in the same point, also replaces point (a):
„Die Kapitel I und II gelten ab dem 2. Februar 2025, mit Ausnahme von Artikel 5 Absatz 1 Unterabsatz 1 Buchstaben ba und bb und Artikel 5 Absätze 1a und 1b, die ab dem 2. Dezember 2026 gelten."
That is the decisive finding. The legislator opened point (a) in the same amendment, carved something out of it, and left Chapter I sitting on 2 February 2025. The exceptions concern only new prohibitions in Article 5, which is Chapter II. Chapter I is untouched.
Article 4 is the last article of Chapter I. It has applied since 2 February 2025.
What Article 4 says today
This too has changed, and the change is newer than most summaries. The original 2024 wording was:
„Die Anbieter und Betreiber von KI-Systemen ergreifen Maßnahmen, um nach besten Kräften sicherzustellen, dass ihr Personal und andere Personen, die in ihrem Auftrag mit dem Betrieb und der Nutzung von KI-Systemen befasst sind, über ein ausreichendes Maß an KI-Kompetenz verfügen [...]"
Regulation (EU) 2026/1744 replaces Article 4 in full. It entered into force on the third day following publication, meaning 27 July 2026. Paragraph 1 of the version now in force reads:
„Die Anbieter und Betreiber von KI-Systemen ergreifen Maßnahmen, um die Entwicklung der KI-Kompetenz ihres Personals und anderer Personen zu unterstützen, die in ihrem Auftrag mit dem Betrieb und der Nutzung von KI-Systemen befasst sind [...] Diese Verpflichtung verpflichtet Anbieter oder Betreiber nicht, für irgendeine Person ein bestimmtes Niveau an KI-Kompetenz zu garantieren."
In English: providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, taking into account their technical knowledge, experience, education and training and the context the systems are to be used in, as well as the persons or groups on whom they are to be used. This obligation does not require providers or deployers to guarantee any specific level of AI literacy for any individual.
The difference lies in two points. „Take measures to ensure, to their best extent, that they have a sufficient level" has become „take measures to support the development". And a second sentence now states expressly that no particular level is owed for any single person. The amendment lowers what has to be delivered. It does not remove it.
The test you can run in an afternoon
Article 4 has no threshold, no company size and no risk class. It is three questions, and you answer all three from your own records.
- Do you operate an AI system, meaning are you a deployer within the meaning of the Regulation? The duty falls expressly on providers and deployers.
- Are there staff or other persons dealing with the operation and use of that system on your behalf? The wording reaches beyond your own workforce.
- Have you taken measures that support the development of their AI literacy, and can that be evidenced?
If the first two answers are yes and the third is no, an obligation is running, and not since yesterday.
What attaches to it, precisely and no more
Precision is worth it here, because the debate tends to reach for the Regulation's large numbers. Article 4 does not appear in the list in Article 99 (4), nor in paragraphs 3 and 5. The fine ranges set out there apply to Article 5, to the obligations under Articles 16, 22, 23, 24, 25, 26 and 50, to requirements on notified bodies, and to supplying incorrect information.
For breaches of Article 4, Article 99 (1) governs, and the same amending regulation replaced that paragraph too:
„Entsprechend den Vorgaben dieser Verordnung erlassen die Mitgliedstaaten Vorschriften für Sanktionen und andere Durchsetzungsmaßnahmen, zu denen auch Geldbußen, Verwarnungen und nichtmonetäre Maßnahmen gehören können, die bei jeglichem Verstoß gegen diese Verordnung durch Akteure Anwendung finden [...] Die vorgesehenen Sanktionen müssen wirksam, verhältnismäßig und abschreckend sein."
In English: Member States lay down rules on penalties and other enforcement measures, which may include fines, warnings and non monetary measures, applicable to any infringement of the Regulation by operators, and those penalties must be effective, proportionate and dissuasive. Two formulations there are new and matter here: fines are named expressly among the national measures, and they are available for any infringement. So anyone looking for a figure will not find one for Article 4 in the European text. That does not make the duty smaller. It moves who puts a number on it.
What this has to do with qualification, and what it does not
A measure that develops employees' AI literacy is a training measure like any other. It therefore falls into the same works constitution machinery: where measures by the employer change the work and existing occupational knowledge no longer suffices, Section 97 (2) BetrVG applies, and Section 98 BetrVG governs how the measure is carried out. Answering Article 4 with a mandatory course planned around codetermination trades one problem for a second.
Equally plainly in the other direction: QualiShift does not discharge this obligation. We are not a training provider, we produce no learning content and we do not teach. This piece states that a duty is running. It does not sell the remedy.
What this piece does not do
This piece is not legal advice. It sets out the wording in force and places the application dates. Whether a given system is an AI system within the meaning of Article 3, and who is the deployer in a specific case, turns on the individual facts and belongs with the people who own that classification in house.
Sources
- Regulation (EU) 2024/1689 (AI Act), Articles 4, 99 and 113.
- Regulation (EU) 2026/1744 of 8 July 2026, Official Journal of the European Union, L series 2026/1744 of 24 July 2026, points 5 and 40 of Article 1 and Article 4 of the amending regulation.
- Sections 97 (2) and 98 Works Constitution Act.