Three hours before this article published, Reuters reported that EU member states and the European Parliament had failed to reach agreement on the Commission’s proposed simplification of the AI Act. The trilogue, the closed-door negotiating session that turns Commission proposals into actual law, ended with no deal.

That sentence sounds boring. It is not.

The AI Act’s biggest enforcement deadline, the one that turns the high-risk system rules into law you can be fined 35 million euros for breaking, lands on August 2, 2026. Ninety-five days from today. And as of this morning, Brussels does not actually agree on what the rule book is supposed to say.

What the omnibus was supposed to do

When the Commission announced the Digital Omnibus in November 2025, the pitch was straightforward. The AI Act, in its enacted form, is the most restrictive AI law in the world. It treats general-purpose models as systemic risks. It treats anything used in employment, credit, education, or law enforcement as a high-risk system. It carries fines up to 7 percent of global revenue. American AI companies hated it. European AI companies hated it more.

The omnibus was supposed to soften the edges. Push back deadlines. Simplify documentation. Let SMEs off the worst of the procedural burden. Roll the AI Act, the Data Act, and parts of GDPR into a single coordinated reform.

The Commission needed Council and Parliament to ratify the package together. Wednesday’s session was supposed to deliver that. It did not.

Why it broke

According to the Reuters readout, member states pushed for harder rollback. Parliament refused to gut the high-risk obligations. The civil society coalition (over 100 NGOs and a separate coordinated US-based campaign) had already publicly called for the omnibus to be rejected outright as a deregulatory backdoor. Apple separately criticized parts of the package as bad for privacy and security.

The result: no compromise text. Trilogue suspended. Negotiations continue, but with the August deadline now a real cliff.

The August 2 problem

Here is what happens at midnight on August 2 if the omnibus does not pass.

Every general-purpose AI model placed on the EU market falls under the systemic-risk transparency regime. Every high-risk Annex III system (employment screening, credit scoring, biometric ID, education sorting, anything law-enforcement adjacent) needs CE marking, a conformity assessment, technical documentation, and a quality management system. Fines up to 35 million euros or 7 percent of worldwide revenue, whichever is higher.

There is no pause button. The AI Office in Brussels has the legal authority to start opening cases the day the deadline passes. Whether they will is a political question, but the legal posture flips on August 2 regardless.

The exits have started

OpenEvidence, the Miami-based AI medical decision support tool now valued at $12 billion and used daily by 40 percent of US doctors, pulled its app from the EU and UK app stores this week. The company’s public statement cited regulatory uncertainty including the AI Act. They are the first US AI unicorn to publicly geofence Europe out of their product over the law.

They will not be the last. OpenAI paused its UK data center plans earlier this month over energy costs but the regulatory overhang is doing similar work on product side. Anthropic’s European rollout has been deliberately slow. Meta and Google have spent the last six months in continuous lobbying meetings that, until Wednesday, looked like they were going to deliver a softer regime.

Wednesday made the regime less predictable, not softer. That is the worst possible outcome for product launch teams.

What it means for the global rule book

The AI Act was supposed to be the model. Brazil, India, the UK, and several US states wrote AI bills that explicitly cited Brussels as the template. The Biden executive orders borrowed AI Act risk categories almost verbatim. Trump killed those executive orders, but the state laws (Colorado SB 24-205, California SB 53, the New York RAISE Act, Texas TRAIGA) are all variants of the same European architecture.

If Brussels cannot make its own architecture work, the rest of the world has to ask whether the architecture is the problem.

Meanwhile the United States is moving the other direction. The DOJ intervened in Colorado last Friday to block SB 24-205 on federal preemption grounds (a story this publication covered earlier today). The Trump administration’s position is that there will be one federal AI policy and it will be permissive. Brussels’ position is that there will be 27 enforcement authorities and they will be strict.

Right now the Trump position is more coherent than the Brussels position. That is going to drive product decisions for the next twelve months.

What to watch

Three things over the next two weeks.

One: whether the Council schedules another trilogue before the May parliamentary recess. If they do not, the AI Office will start guidance documents that effectively freeze the law in its current form.

Two: how many other US AI products follow OpenEvidence in pulling their EU rollouts. Watch the medical, fintech, and HR vertical first. Those are the cleanest fits for high-risk classification.

Three: whether any European Member of Parliament publicly proposes a legislative pause on the August 2 deadline. That would be the cleanest political signal that the cliff is real and Brussels knows it.

Europe spent four years writing the world’s most restrictive AI law. It has 95 days to figure out what the law actually requires. Right now nobody in the room knows.