The EU AI Act needs more than one deadline.

I’d start an AI compliance discussion with a list of systems, rather than a date. What does each system do? Who uses it? What does the organisation provide, and what does it deploy? Those are the questions I would want answered before accepting a reassuring green box on a dashboard.

The European Commission’s Article 50 guidance separates several transparency duties. Providers of systems that interact directly with people must inform them that they are dealing with AI, unless that is obvious. Providers’ machine-readable marking duties are distinct from deployers’ disclosure duties for covered deepfakes and certain public-interest text. The conditions and exceptions matter. It is too broad to say that every piece of AI-assisted content needs the same label.

The general Article 50 application date is 2 August 2026. The Commission also describes a limited transition until 2 December 2026 for marking duties concerning systems already on the market before that general application date. That is not a postponement of every transparency obligation.

The Commission’s announcement that the AI Omnibus entered into force gives different dates for high-risk systems: 2 December 2027 for Annex III and 2 August 2028 for high-risk AI embedded in regulated products. Those dates should travel with their categories. “The deadline moved” leaves too much unsaid.

My practical suggestion is to give each system a short record: its purpose, the relevant role, the obligation being assessed, the evidence behind that assessment and the person responsible. If an exception is being relied on, write down why. If the answer is still uncertain, leave it visibly unresolved rather than letting it inherit the status of the rest of the programme.

I would also separate a notice being present from a notice being appropriate. A screenshot can show what a customer sees. It cannot, on its own, settle which legal provision applies. That is where operational evidence and legal interpretation need to meet.

For a senior leadership team, my question is simple: could you pick one customer-facing AI system and follow that record from the product through to its scope decision? If you cannot, start there. I would rather resolve one concrete uncertainty than build a calendar which hides it.