Antefacts

Note · 31 August 2026

The high-risk deadline moved. The notified body timetable did not.

Regulation (EU) 2026/1744 defers the EU AI Act’s entire high-risk regime by sixteen months and more. It leaves the machinery for designating notified bodies exactly where it was. The distance between those two facts is where the next two years of this market will be decided.

What happened

A deadline was removed three weeks before it arrived

Regulation (EU) 2026/1744, the Digital Omnibus on AI, was adopted on 8 July 2026 and entered into force on 27 July 2026. It amends Regulation (EU) 2024/1689 “as regards the simplification of the implementation of harmonised rules on artificial intelligence”.

Its practical effect is a change to Article 113, the application-dates provision. Chapter III Sections 1 to 3 — classification under Article 6, the requirements in Articles 8 to 15, and the provider obligations in Articles 16 to 27 — no longer apply from 2 August 2026.

Anyone who spent two years building towards that date was building towards a wall that was taken down days before impact.

The dates

What applies when

ObligationApplies fromDeferred?
Prohibited practices (Art 5)2 February 2025No
GPAI model obligations (Ch V)2 August 2025No
Notified body designation framework (Ch III Sec 4)2 August 2025No
Governance and penalties2 August 2025No
Transparency obligations (Art 50) and general application2 August 2026No
High-risk requirements and provider obligations, Annex III systems2 December 2027Yes
High-risk requirements and provider obligations, Annex I embedded systems2 August 2028Yes

Annex III systems are the stand-alone high-risk categories: biometrics, critical infrastructure, education, employment, essential services and creditworthiness, law enforcement, migration, and the administration of justice. Annex I systems are those embedded as safety components in products already regulated under Union harmonisation legislation.

The gap

Bodies are being designated for a regime that has not started

Chapter III Section 4 governs notifying authorities and the designation of notified bodies. It was not touched by the Omnibus. It has applied since 2 August 2025, and it continues to apply now.

So the position as of today is this. The machinery for appointing the organisations that will assess high-risk AI systems is live. The obligations those organisations will assess against do not bite for another fifteen months, and for embedded systems, another two years.

That is not a dead period. It is the only period in which a body can choose its technical method deliberately rather than improvise one under deadline pressure. Once December 2027 arrives, whatever a body is doing at that moment is what it will keep doing, because there will be no time to change it.

Consequence

Two readings, and both are correct

For providers, urgency has genuinely gone. Anyone who was going to buy assurance tooling this year because of the August 2026 date now has sixteen months they did not have in July. Budgets that were being defended on the grounds of a hard deadline no longer have one to point at.

For assessment bodies, the opposite is true. The designation clock did not stop. A body designated in 2026 will be asked, in December 2027, how it evaluates whether an AI system meets Article 15. Article 15(3) requires accuracy levels and the relevant accuracy metrics to be declared in the instructions for use. It does not say those declared figures must be reproducible by anyone other than the provider who declared them.

That gap between a declared number and a checkable one is not closed by the regulation. It has to be closed by method.

Our position

Why we are writing this down

We build evaluation infrastructure, so we have an obvious interest in people concluding that evaluation method matters. Stating that plainly is cheaper than pretending otherwise.

What we would say regardless of commercial interest is narrower. The deferral changes when the obligations bite. It changes nothing about whether the underlying problem is solved, and the underlying problem is that a conformity assessment resting on a test nobody outside the room can rerun is an attestation rather than evidence. Sixteen extra months is time to fix that properly. It is also, for anyone who chooses to read it that way, sixteen months of not having to.

Corrections. This is our reading of a regulation that entered force five weeks ago, and the consolidated text is still settling. If any date or article reference here is wrong, tell us and we will correct it on the page and say what changed. We would rather be corrected in public than be quietly wrong.
Sources
  • Regulation (EU) 2024/1689, consolidated text as of 27 July 2026 — EUR-Lex
  • Regulation (EU) 2026/1744, Digital Omnibus on AI, adopted 8 July 2026, in force 27 July 2026 — EUR-Lex
  • Article 113, entry into force and application, as amended — artificialintelligenceact.eu

Nothing on this page is legal advice. Antefacts is not a law firm. Anyone with an obligation under this Regulation should take advice on their own position.