The Digital Omnibus on AI has rewritten the compliance calendar for Europe’s flagship artificial intelligence law, and the first of its new deadlines, 2 December 2026, is less than four weeks away. Parliament and the Council reached a provisional agreement on 7 May 2026, after talks had failed to close a month earlier, and the co-legislators aimed to adopt it before 2 August 2026.
The Commission proposed the change on 19 November 2025 as part of its wider digital simplification package. The political deal keeps that logic but adds a sharper edge: it delays the heavy obligations and introduces a new prohibition at the same time.
What the Digital Omnibus on AI changes for high-risk systems
The central change in the Digital Omnibus on AI is timing. Stand-alone high-risk systems in the Annex III areas, which include biometrics, critical infrastructure, education, employment, migration, asylum and border control, now face obligations from 2 December 2027. High-risk systems that serve as safety components of products, such as lifts and toys, follow on 2 August 2028.
For providers, that gives roughly a year more than the original August 2026 start for the first group. The Digital Omnibus on AI therefore works as a delay for the most demanding rules, and the extra months matter because harmonised standards and guidance have lagged the legal text.
Not everything slides. Article 50 transparency obligations still start on 2 August 2026 as originally scheduled, so chatbots and systems that interact with people have had no reprieve.
The new ban and the watermarking deadline
The agreement adds a prohibition on AI systems that generate child sexual abuse material, or that create sexually explicit images, video or audio of an identifiable person without consent. It covers placing such systems on the market, placing them on the market without reasonable safeguards, and deployers using them for those purposes. Compliance is due by 2 December 2026.
The same date applies to watermarking of AI-generated content, which moves to 2 December 2026 as a delay from the earlier schedule. Providers of generative tools should treat that date as the live constraint, not the 2027 deadlines, because the deadline falls this year.
Other changes that matter to companies
Several smaller amendments in the Digital Omnibus on AI reduce the burden on smaller firms. SME exemptions extend to small mid-cap enterprises, which widens the group that benefits from lighter documentation. The definition of a safety component narrows, so functions that only assist users or optimise performance avoid high-risk status when their failure creates no health or safety risk.
Machinery makers need only meet sectoral safety rules, with safeguards for equivalent protection. Providers may process personal data where strictly necessary to detect and correct bias, under safeguards, a change that responds to a long-standing tension between data protection and fairness testing. The AI Office also gains streamlined enforcement over certain general-purpose systems, and providers must still register in the EU database any system they consider exempt under Article 6(3).
What happens next
After formal adoption of the Digital Omnibus on AI by both institutions, the text goes to the Official Journal and enters into force three days later. Companies should confirm publication status before relying on the new dates, since this analysis draws on the May agreement.
The practical advice is simple. Prioritise the December 2026 deliverables, meaning the watermarking mechanism and the new prohibition screening, then use the extra year to build conformity processes for Annex III. The Digital Omnibus on AI buys time, but it does not remove the work.
