The deadline that did not come — and the obligations that apply anyway

2 August 2026 was long regarded as the big deadline of the European AI Act. On that day, the obligations for high-risk systems were due to take effect. Six days before, on 27 July 2026, an amending regulation entered into force instead, postponing those obligations.

The basis is Regulation (EU) 2026/1744, the so-called «Digital Omnibus on AI». It was adopted on 8 July 2026 and published in the Official Journal on 24 July 2026: Regulation (EU) 2026/1744 in the Official Journal.

What was postponed

• Stand-alone high-risk systems (Annex III): from 2 August 2026 to 2 December 2027.
• AI in regulated products (Annex I): from 2 August 2027 to 2 August 2028.

The reason given is a practical one: the technical standards and testing tools companies were supposed to align themselves with were not finished, and in several member states the competent authorities were missing. An obligation without a yardstick cannot be met.

What was not postponed

Here lies the part the headlines usually leave out. Three obligations are untouched by the postponement:

• Prohibited practices (Article 5) — apply unchanged since 2 February 2025.
• AI literacy (Article 4) — likewise applies since 2 February 2025. Anyone deploying AI must ensure that the people working with it understand it.
• Transparency (Article 50) — stayed at 2 August 2026 and has therefore been in force for barely two weeks.

What was postponed is the high-risk assessment. The obligation to disclose that a machine is involved has applied since 2 August.

Article 50 requires three things: anyone speaking to an AI system must be told. Artificially generated content must be recognisable as such. And the outputs of generative systems must be marked in machine-readable form. For the last point, systems that already existed beforehand have a transition period until 2 December 2026.

What this means for Swiss businesses

The AI Act does not stop at the EU's external border. It also applies to providers outside the Union if the output of a system is used within the EU. A Swiss business operating an AI-supported service for customers in Germany or France falls under it — regardless of where the server stands.

For Switzerland itself, the revised Data Protection Act continues to apply alongside. In January 2026, the FDPIC recalled that it is already applicable to AI applications today — without any need for a dedicated AI law.

What makes sense now

The postponement is extra time, not an all-clear. It still makes sense to do the classification first: which systems do we deploy, and which of them would be high-risk systems? That work is the precondition for everything else and is not changed by the new date — it simply has more room.

In the short term, Article 50 is what counts, because it already applies. The question that goes with it is simple: can anyone dealing with our systems tell that an AI was involved? Where the answer is no, that is this month's building site — not December 2027's.