EU Lawmakers Reach Provisional Agreement to Delay Key EU AI Act Obligations

A blog post from the law firm Sidley Austin summarizes a provisional agreement between the European Council and European Parliament on the EU Digital Omnibus on AI, which makes targeted changes to the EU AI Act. The deal was reached on 7 May 2026, and Parliament voted to adopt it on 16 June 2026. Formal approval by the Council is still outstanding. The Commission proposed the Omnibus on 19 November 2025 after concerns about late standards and compliance costs.

The AI Act was due to apply most of its obligations from 2 August 2026, including core duties for high-risk systems. Under the agreement, rules for stand-alone high-risk systems, such as those used in employment, move to 2 December 2027, and rules for high-risk systems embedded in regulated products, such as medical devices, move to 2 August 2028. Watermarking duties for AI-generated content are postponed to 2 December 2026 for providers whose systems were already on the market before 2 August 2026.

Several other points are covered. The Commission's idea of dropping registration for systems relying on the Article 6(3) exemption was abandoned, though a simplified procedure will apply. The meaning of safety component is narrowed, so assistive or performance-optimizing AI functions that cannot cause health or safety risks are not automatically high-risk. A new ban on AI that generates child sexual abuse material or non-consensual intimate content applies from 2 December 2026. The AI literacy duty shifts so that the Commission and Member States support providers and deployers. Sandbox deadlines slip to 2 August 2027, and some SME exemptions extend to small mid-cap firms.

After Council adoption, the text goes into the Official Journal and applies three days on. Both institutions aim to finish before 2 August 2026.