Two deadlines. One agreement. Don’t confuse them.
The EU AI Act Omnibus political agreement, reached on May 7, 2026, is being read in many compliance circles as a general delay. That reading is incomplete. The political agreement creates two structurally different compliance tracks for two different populations, with different obligations due roughly a year apart.
Track one lands December 2, 2026. It covers GenAI providers and two specific obligations. First, synthetic content watermarking: AI-generated audio, image, video, and text must carry a machine-readable label. The Omnibus compressed the grace period for this obligation from six months to three, according to Latham & Watkins’ analysis. No harmonized EU technical standard for watermarking currently exists, which means providers are building toward a specification that hasn’t been finalized. Second, the Article 5 prohibition on AI-generated intimate imagery, now explicitly named in the prohibited practices list per Baker Botts’ reading of the agreement. This doesn’t require building anything. It requires stopping something, and three months isn’t much time to audit and restrict a model’s generative boundaries. Fines for prohibited practice violations may reach up to €35 million or 7% of global annual turnover, according to the agreement’s terms, though this figure hasn’t been verified against the official text.
Who This Affects
Track two lands December 2, 2027. Annex III high-risk system compliance — covering biometrics, employment, critical infrastructure, education, and law enforcement AI — was extended from the prior August 2, 2026 deadline. That’s eighteen months of additional runway for organizations with active compliance programs. The risk: compliance teams that read “delay” into the headlines and paused their programs will find themselves rebuilding under pressure in 2027. Risk assessments completed toward August 2026 remain valid. Technical documentation in progress should continue. The conformity assessment requirements didn’t change. Only the deadline did.
The catch on both tracks: the Omnibus is a political agreement, not enacted law. Formal Council adoption and publication in the EU Official Journal are required before these deadlines become legally binding. Build toward the agreement terms now. Set a review checkpoint for Official Journal publication. Don’t treat the pending formal step as a reason to wait.
The real question is whether the EU AI Office publishes the watermarking technical specification early enough for GenAI providers to build against it before December 2. Three months from specification to compliant implementation is tight for any organization without a head start on watermarking architecture. Organizations that haven’t started are already behind a reasonable build schedule.
Don’t expect the two tracks to converge. A system that triggers both GenAI transparency obligations and Annex III high-risk classification carries December 2026 obligations on its transparency layer and December 2027 obligations on its risk management framework — two programs, two timelines, one AI system.