Brussels has granted European companies a delay that many were calling for: the most onerous obligations of the AI Act for high-risk systems, initially scheduled for August 2, 2026, are pushed back to late 2027. According to experts consulted by The AI / news, this reprieve should not lead to pausing projects, as the problems to be solved would remain the same by the deadline.
- The AI Act’s high-risk obligations are postponed to late 2027.
- Transparency obligations still apply from December 2026.
- The delay does not justify pausing AI projects, since the underlying problems remain.
- Companies must inventory their AI systems, train human supervisors, and document their AI use now.
Don’t confuse a pause with regulatory silence
The postponement only concerns part of the regulation. The full obligations for high-risk systems in Annex III, which cover recruitment, credit scoring, biometrics, education, and judicial systems, are shifted. The rest of the obligations remain, with no reprieve. The absolute bans (manipulation, social scoring, emotion recognition at work) have been in force since February 2025, and the obligations for general-purpose AI models have applied since August 2025.
The transparency obligations, which require being informed when speaking to a chatbot or labeling AI-generated content, remain due from December 2026, with no delay. The first step, therefore, is to establish a precise mapping of these risks: identify what is still immediately required, what is slipping, and for how long.
Turn the documentation burden into an asset
The European Parliament speaks of “simplification,” but this does not remove the substantive requirements. It shifts the focus to proof. Data governance, bias testing, decision traceability, and truly effective human supervision will be demanded by regulators, investors, or clients questioning their suppliers about their AI use.
Companies that dedicate the next eighteen months to compiling this documentation will have a ready compliance file, but also a common language between legal, human resources, IT, and senior management. They will know what their AI actually does, for whom, and with what limits.
Prioritize rather than tackling everything at once
The delay allows for proceeding in a reasonable order. Three priorities emerge from The AI / news. The first is to inventory before classifying: many organizations use high-risk systems without knowing it, whether it’s a scoring module integrated into an ATS (applicant tracking system), a feature of a third-party HR tool, or a plugin added without central validation.
The second is to treat human supervision as a skill. The regulation requires real human intervention, what is called human in the loop, in high-impact decisions like recruitment, credit, or access to essential services. This requires training people to understand algorithmic recommendations, to question them, and to document cases where they deviate from them. The AI / news notes that teams will soon include AI agents among their members, making their management (setting rules, controlling results) all the more necessary.
The third priority is not to focus solely on Annex III. The transparency obligations are not moving on the calendar, and chatbots, generated content, and interview analysis systems constitute, according to the text, this year’s project.
Why act now
The political consensus that allowed this postponement remains fragile. The vote in the European Parliament tightened significantly between the initial discussions and the final ballot. Nothing guarantees that a new period of tensions won’t result, by 2027, in a tightening rather than a relaxation. Building compliance while counting on future leniency is, according to The AI / news, more of a gamble than a strategy.
