August 2 was not the deadline most slide decks said it was
On 2 August 2026 the EU AI Act crossed a real threshold, but a narrower one than many compliance presentations predicted a year ago. What entered into application that day are the transparency obligations of Article 50: telling users when they are talking to a chatbot, marking AI-generated content so that other systems can recognize it as such, and labeling deepfakes. The enforcement tools for general-purpose AI models became active the same day. Digital Applied's checklist is a useful walkthrough of what this means for teams that publish AI-assisted work.
What did not happen on 2 August is the part worth internalizing: the heavy high-risk regime did not switch on.
What the Digital Omnibus changed
The Digital Omnibus package was published in the EU Official Journal on 24 July 2026 and entered into force on 27 July. As Gibson Dunn explains, it postponed the obligations for high-risk systems listed in Annex III (hiring, credit scoring, biometric use cases and the like) to 2 December 2027, and the obligations for AI embedded in products covered by Annex I to 2 August 2028.
One further date matters for anyone shipping generated media: solutions for marking artificial content, watermarking included, have until 2 December 2026 to reach compliance.
The working calendar now looks like this:
- 2 August 2026: Article 50 transparency duties apply; GPAI enforcement tools are live.
- 2 December 2026: marking solutions for artificial content must be compliant.
- 2 December 2027: Annex III high-risk obligations apply.
- 2 August 2028: Annex I embedded systems follow.
The fines are designed to be taken seriously
The penalty structure has three tiers. Prohibited practices can cost up to 35 million EUR or 7% of global turnover. Violations of high-risk obligations reach 15 million EUR or 3%. Even supplying incorrect information to authorities carries up to 7.5 million EUR or 1%. These are not symbolic numbers; they are calibrated to make transparency cheaper than the alternative.
No EU office does not mean no EU exposure
The Act reaches beyond the Union's borders. If the output of your system meaningfully touches the EU, your company can fall within scope. A support chatbot operated from Istanbul that serves customers in Munich is part of this conversation, whether or not anyone in the org chart sits in Europe.
What to do this quarter
Working daily with Turkish and German clients from İzmir and Berlin, we at Clodron keep the practical list short:
- Disclose your chatbots. Anywhere an AI talks to a customer, say so plainly in the interface. This is live law now, not a future obligation.
- Mark generated media. If your marketing pipeline produces AI images or video, make sure the output carries machine-readable marking, and know which tool applies it.
- Put it in supplier contracts. Ask every AI vendor how their output is marked and what documentation they provide. Their December 2026 deadline is your dependency.
- Map the calendar to your roadmap. If anything you are building could land in Annex III, 2 December 2027 is your date. Postponed is not cancelled.
Transparency is cheap to build in now and expensive to retrofit under enforcement. That is the real message of August 2026.