IMPERIALArticles

The AI Act and the omnibus

Europe was right to wait, once

Delaying the AI Act’s high-risk rules was sensible, because the standards were missing, but a second delay would turn caution into habit.

· Policy

Curved glass facade of the Paul-Henri Spaak building, part of the European Parliament complex in Brussels, under a grey sky
The Paul-Henri Spaak building in Brussels, part of the European Parliament complex. Photo: Fred Romero, CC BY 2.0, via Wikimedia Commons.

Sixteen months is how much longer Europe has given itself. Most of the AI Act, including its rules for high-risk systems used in hiring, schools and border control, was due to apply on August 2nd 2026.1 Under the “digital omnibus” that Parliament approved on June 16th and the Council adopted on June 29th, those rules now start on December 2nd 2027.2 Rules for AI built into products such as lifts and toys move to August 2nd 2028, a year later than the 36-month schedule set in 2024.37 Anyone applying for a graduate job in 2027 will do so before the rules for CV-screening software apply.

The Commission proposed the omnibus on November 19th 2025, and Parliament and the Council reached a political agreement on May 7th 2026.2 We think the delay it contains was right, and that it should be the last. The Act asks firms to prove compliance, and the yardstick is meant to be harmonised standards, which give a firm a presumption of conformity if it follows them. The Commission’s standardisation page lists one finished standard, on quality management systems, out of ten areas of work.5 Insisting on compliance in August 2026 would have been like scheduling the final before anyone had painted lines on the pitch. Postponing the match does not make football optional.

The omnibus is also narrower than the headlines about Brussels blinking suggest. The bans on prohibited practices have applied since February 2025.1 The Commission’s enforcement powers over general-purpose models, including fines, began on August 2nd 2026, and nothing in the omnibus moved them.6 The package adds a ban on systems that generate non-consensual intimate imagery, which applies from December 2nd, 15 days from now, and gives the AI Office wider oversight of systems built on general-purpose models.14 A regulator that delays one chapter and tightens two is not in retreat. Those who say Europe has gutted its own law should name the chapter.

The strongest objection is that delay has victims. Parliament’s vote was 423 to 57, with 174 abstentions, so more than a quarter of those voting declined to bless it.2 Critics can fairly say that employers and universities may now run unaudited screening software for another 16 months, and that the omnibus extends simplified duties from small firms to small mid-caps, which are not small.4 That is a real cost.

It is smaller than the alternative. A duty that nobody can verify protects nobody; it gives auditors something to do and applicants nothing to hold. Compliance on paper in August 2026, with no agreed test to mark it against, would have rewarded the firms best at writing their own marking schemes, who tend to be the largest. The Commission’s own account is that the staggered timing is meant to ensure that technical standards and support tools are in place first.3 That is a reason to wait for standards, not to stop writing them.

Which is why the dates matter more than the delay. The Commission’s original proposal tied the high-risk timing to the availability of standards or support tools.2 A deadline pegged to readiness recedes whenever readiness slips. The final text instead fixes calendar dates, and the standards bodies now have the one thing they lacked, which is a day on which somebody will ask for the work.2 If December 2027 arrives with one standard of ten, that will be a failure of the people writing them, and Europe should say so rather than reach for another omnibus.

Europe has spent its one reprieve. A second would teach every firm that the dates in the Official Journal are merely suggestions.

The Imperial Review is an independent publication. Not affiliated with Imperial College London.