Paris: The political deal that European negotiators reached on 7 May reopens the AI Act not by rewriting its core obligations but by stretching the timeline against which compliance is judged. For French firms that have spent the past eighteen months building documentation pipelines for high-risk systems under Annex III, the agreement effectively pushes the binding application date out to 2 December 2027. Annex I products, those embedded in regulated goods such as medical devices, machinery and toys, gain even more breathing room, with their obligations now sliding to 2 August 2028. The compromise was sold in Strasbourg as a simplification omnibus, but for businesses the more honest description is a calibrated retreat from a calendar that industry, several member states and the Commission’s own competitiveness desk had all flagged as unworkable.
The negotiating dynamic that produced the deal had been building since the Draghi competitiveness arguments took hold inside the Berlaymont. Council majorities led by France, the Netherlands and the Nordics pressed for delay on the grounds that supporting technical standards from CEN-CENELEC and ETSI were unlikely to be finalised on the original schedule, and that conformity assessment bodies were not yet accredited in sufficient numbers. Parliament’s Internal Market and Civil Liberties co-rapporteurs accepted timeline relief in exchange for tighter language on AI-generated intimate content. From 2 December 2026 the Act will explicitly prohibit nudifier applications that produce non-consensual sexualised imagery and any AI tool that generates child sexual abuse material. That carve-out was the price the Parliament extracted for moving on the timetable.
The SME provisions are the second layer of the deal, and they will have the broadest practical effect on the French AI ecosystem. The agreement extends the Act’s simplified compliance framework, reduced fines, regulatory sandbox access and standardised documentation templates, to companies with up to 750 employees and 150 million euros in annual revenue. That is well beyond the EU’s traditional SME ceiling and catches a long tail of mid-cap firms that had been lobbying through Hexagone numerique and similar national bodies. For Mistral and the broader cohort of French model providers, the change removes a competitive distortion that had threatened to penalise scale-up firms without giving them the resources to absorb full-tier compliance.
Watermarking obligations under Article 50(2) move with the rest of the calendar but on a shorter leash. Providers of generative systems will have until 2 December 2026 rather than the six-month delay originally floated. The shorter postponement reflects pressure from the European External Action Service, which has argued that detection capabilities matter more in an information environment hardened by 2024 election interference cases. The compromise leaves model providers little time to align with technical standards that are themselves still being drafted, but it preserves the policy signal that synthetic media identification is a near-term obligation rather than an aspirational target.
The provisional agreement still requires formal adoption, expected by July 2026 ahead of the parliamentary recess. The political coalition behind it is fragile in two places. Greens and some Renew members are uneasy about the high-risk delays without compensating enforcement reinforcement. And the Council’s German bloc has indicated it will press for clarifications on the technical-standards trigger that allows the Commission to bring the new deadlines forward if standards land early. Brussels will need both questions answered before the file lands cleanly on the legislative pipeline, and the rapporteurs have signalled that any reopening risks unravelling the careful balance between deferred application and the new prohibition perimeter.
