- → On August 2, 2026, most of theAI ActThe 2024 European regulation which regulates the development and use of AI according to risk levels. becomes applicable : transparency, supervision of GPAI models, sandboxes and sanctions regime.
- → THE fines can reach €35 million or 7% of global turnover for prohibited practices.
- → THE Digital Omnibus has postponed certain “high risk” deadlines to 2027-2028, but not the date of August 2 for the rest.
- → Concretely: if you use or provide AItransparency towards users becomes mandatory this summer.
This is the deadline that all legal departments in Europe have circled in red: the August 2, 2026the European regulation on artificial intelligence, the AI Act, is entering its broadest phase of application. Two years to the day after its entry into force, most of the text ceases to be theoretical and becomes binding. And despite rumors of postponement, this date stands firm.
For an overview of the regulation, its risk categories and its history, see our comprehensive guide to the AI Act. Here, we answer only one question, the most concrete: what really becomes binding on August 2, 2026?
There was indeed a relaxation, the “Digital Omnibus”, which postponed certain obligations. But above all it has sowed confusion about what really applies, and when. Here’s the sort, without jargon.
August 2, 2026 · full application
7% of turnover · maximum fine
1 sandbox · per Member State
GPAI · under direct supervision
01, The seesaw
August 2, 2026, why this date
The AI Act entered into force on August 1, 2024. The text provided for application in stages: prohibited practices from February 2025, the obligations of general purpose AI models (GPAI) in August 2025, then the bulk of the regulation two years after entry into force, i.e. August 2, 2026.
This date therefore marks the full application of the heart of the text. The European Commission has confirmed: with the exception of one technical article, all the rest of the law begins to apply on that day. It’s no longer a distant prospect, it’s just a few weeks away.
02, What applies
What becomes obligatory on this day
Four blocks actually come into force on August 2, 2026, and they affect a majority of companies, not just tech giants.
Transparency (Article 50) you must warn the user when they interact with an AI, and report content generated or manipulated by AI (text, images, audio, video).
Supervision of GPAI models the obligations of suppliers of general-purpose models (Claude, GPT, Mistral, Gemini, Llama) have existed since August 2025, but the Commission can only sanction from August 2, 2026.
Regulatory sandboxes each Member State must have set up at least one “sandbox” allowing the testing of AI under supervision, operational on that date.
The sanctions regime the fines become applicable in full. Without a sanction system, the obligations would remain theoretical.
03, Reports
What the Digital Omnibus pushed back
This is the source of all the confusion. At the end of 2025, the Commission proposed a simplification package, the Digital Omnibus, to give businesses breathing space and take into account technical standards that are not yet ready. But be careful: it does not postpone the date of August 2 en bloc, it mainly targets “high risk”.
Systems classified as high risk (biometrics, recruitment, justice, credit, etc.) see their obligations postponed, depending on the case, towards the end of 2027 or 2028, while the harmonized standards are finalized. A short grace period was also granted for the automatic tagging of synthetic content already online. The rest falls on August 2, 2026.
04, Fines
Up to 7% of global turnover
This is what gives weightThe numerical parameters of a neural network, adjusted during training to learn patterns in the data. to the text. The scale of sanctions is modeled, but more severe, on that of GDPRGeneral Data Protection Regulation, the European framework on digital privacy.and it is progressive according to severity.
| Type of violation | Maximum fine |
|---|---|
| Prohibited practices (Article 5) | €35 million or 7% of global turnover |
| Shortcomings of GPAI suppliers | €15 million or 3% of global turnover |
| Other obligations (transparency, documentation) | €15 million or 3% of global turnover |
For an SME, the fixed ceiling (15 or 35 M€) often takes precedence; for a multinational, it’s the percentage that hurts. In either case, the message is clear: compliance is no longer optional.
05, Who is concerned
Are you affected, and what should you do?
Many companies believe that the AI Act only targets model makers. Fake. As soon as you deploy an AI in front of customers or employees, you are a “deployer” and transparency is your responsibility. Here is a simple roadmap for a peaceful summer.
1
Map your AIs
Identify all the AI systems you use or provide: chatbots, HR tools, content generation, scoring. We only protect what we know.
2
Identify your role
Are you a provider (you develop the model) or a deployer (you use it)? The bonds are not the same, and some AIs are high risk.
3
Enable transparency
Clearly inform users that they are interacting with an AI, and mark the content you generate. This is the most immediate obligation.
4
Document
Keep logs, keep your risk assessments and, for models, a summary of training data. Proof of conformity is built upstream.
5
Follow the schedule
High risk is postponed, but transparency and fines arrive from August 2. Do not postpone what is already due.
06, For France
What this changes for French companies
For French players, the challenge is twofold. On the supplier side, Mistral and other European models must prove their GPAI compliance, an area where they have a cultural and regulatory advantage over the Americans. On the user side, thousands of SMEs and ETIs who have integrated ChatGPT, Claude or Copilot into their processes are discovering that they too have obligations.
The good news is that most of the effort comes down to a simple discipline: knowing where AI is used, saying it, and documenting it. Those who do it now will approach August 2 without stress. Others will discover that AI, in Europe, is now a subject of governance as much as technology.
To go further: The complete guide to the AI Act · AI Act: the list of systemic risk models · USA-China: joint AI safeguards protocol · The Mistral model sheet · Compare AI models
Frequently asked questions
- When does the AI Act come into force?
- August 2, 2026 marks the full application of most of the AI Act, two years after its entry into force on August 1, 2024. The prohibited practices had already applied since February 2025 and the GPAI obligations since August 2025.
- What becomes mandatory on August 2, 2026?
- Transparency (Article 50), supervision and possible sanctions against GPAI model providers, regulatory sandboxes in each member state, and comprehensive fines regime become applicable on that day.
- Did the Digital Omnibus beat back the AI Act?
- No, not as a whole. The Digital Omnibus has shifted certain deadlines for high-risk systems (towards 2027-2028) and granted a short grace period for the marking of synthetic content, but the date of August 2, 2026 remains valid for the rest of the text.
- What are the fines under the AI Act?
- Up to €35 million or 7% of global turnover for prohibited practices, and up to €15 million or 3% of turnover for GPAI supplier breaches and other obligations. The higher of the two applies.
- Is my business affected by the AI Act?
- Probably yes. As soon as you deploy an AI in front of customers or employees (chatbot, HR tool, content generation), you are a “deployer” and the obligation of transparency concerns you, even if you did not develop the model.
- What is a GPAI model?
- A General-Purpose AI model, such as Claude, GPT, Mistral, Gemini or Llama, capable of many tasks. Their suppliers have obligations of documentation, transparency and, for the most powerful, assessment of systemic risks.
- What should a French company do before August 2, 2026?
- Map your uses of AI, identify your role (supplier or deployer), activate transparency towards users, document your systems and follow the schedule of obligations. The main thing is a discipline: knowing where AI is used, saying it and documenting it.

