From August 2026, a European company using artificial intelligence will not only have to explain what it gains from it: it will have to be able to demonstrate how it controls it. Regulation (EU) 2024/1689—the AI Act—has entered its effective implementation phase, and with it, management discourse has incorporated three words that previously existed only in the legal department: traceability, transparency, and accountability.
It's worth clarifying something from the outset, because there's a lot of commercial confusion surrounding this date: the calendar changed mid-year. Anyone selling a sense of urgency without explaining the change either doesn't know or prefers you don't.
The AI Act came into force on August 1, 2024, with a phased implementation schedule. The prohibitions on unacceptable risk practices and AI literacy obligations applied from February 2, 2025. The obligations for general-purpose AI models (GPAI) and the governance structure applied from August 2, 2025.
August 2, 2026 marks two specific things:
Now for the part that changes the message. In 2026, the Union approved the package known as Digital Omnibus, which postponed the obligations of high-risk systems: the uses of Annex III —personnel selection, credit scoring, biometrics, education, essential services— are moved to December 2027, and the systems of Annex I, AI embedded in already regulated products, to August 2028.
In other words: if your system is high-risk, you have more time than you're being told. If your system interacts with people or generates content, you have less leeway than you think.
Date | What applies | Who is affected? |
February 2, 2025 | Prohibitions and AI Literacy | All organizations |
August 2, 2025 | GPAI obligations and governance | Providers of general-purpose models |
August 2, 2026 | GPAI sanctions and transparency (art. 50) | Who deploys AI that interacts with people or generates content |
December 2, 2027 | Annex III High-Risk Systems | HR, credit, biometrics, education, essential services |
August 2, 2028 | High-risk systems of Annex I | AI embedded in already regulated products |
The maximum penalties reach €35 million or 71% of annual global revenue, whichever is higher. That's the figure that brought this discussion to the board.
Here's the real problem, and it's not a legal one. It's an inventory issue.
To comply with any regulatory obligation, you must first understand what AI systems exist within the organization, what they do, what data they use, and who is responsible for each one. Most companies cannot answer this question because a significant portion of their AI usage never went through the IT department.
IBM's 2025 Cost of a Data Breach report quantified this phenomenon: 201% of the breaches analyzed involved unauthorized AI within the organization, and these incidents added an average of $670,000 to the average cost of a breach, which that year was $4.44 million. Even more revealing: 97% of the organizations that experienced an AI-related incident lacked adequate access controls, and 63% had no AI governance policy at all.
This connects to a phenomenon we have analyzed in detail: Shadow AI, the risk that comes through the corporate card [internal link]. Tools contracted by a department, containing customer data, without visibility to the CISO or DPO. They do not appear in any inventory because there was never an onboarding process.
Without inventory, any compliance plan is a document disconnected from operational reality.
The tactical error would be to treat the postponement as an excuse for inaction. The elements required by the regulations are not bureaucracy: they are the very practices that allow an AI system to operate safely. Those who implement them out of obligation discover that they needed them all the same.
Inventory of AI systems. What models and tools are used, in what process, with what data, who is responsible for the business, and in what risk category does each one fall?.
Risk classification by use case, not by tool. The same model can be used for low-risk summarizing minutes and high-risk screening resumes. The classification is based on usage.
Decision traceability. A record of what came in, what came out, which version of the model was involved, and who validated it. When a system executes actions and doesn't just suggest them, the audit log ceases to be a good practice and becomes the only available defense mechanism.
Defined human supervision. It's not "there is a person reviewing," but rather: what decisions require validation, with what criteria, within what timeframe, and what happens if that person is unavailable.
Transparency towards the user. If someone is interacting with a system, they should be aware of it. If content was artificially generated, it should be flagged.
Technical and data documentation. Where does the training or context data come from, what known limitations does the system have, and what evaluations have been done?.
To structure all of this, there's no need to invent a framework of your own. NIST AI Risk Management Framework It offers the operational risk management model, with a specific profile for generative AI that identifies twelve risk categories. The standard ISO/IEC 42001:2023 It provides what NIST does not: a certifiable management system, which is what a corporate client or auditor will ask you to show.
This is where the regulatory debate ceases to be legal and becomes technical, which is where we work.
A system not designed to record actions cannot add traceability as a later layer without modifying its architecture. A system without a granular permissions model cannot demonstrate who accessed what. A system tied to a single model vendor cannot document version changes because it lacks control over when they occur.
These three things—traceability, permissions, and the ability to replace the model—are decided during the design phase. Adding them later is expensive. Designing them from the beginning is not.
This is the difference between security by design and guardrails slapped on at the end, and it's why at The Cloud Group, governance and compliance aren't sold as add-ons: they're included in every project, with isolated environments, encryption in transit, granular access control, and factory-built audit logs. We developed it in Our work on AI risks, hallucinations, and governance
If a decision needs to be made at the next meeting about what to do about this, this is the order that produces results:
What's not advisable is the opposite: drafting a forty-page AI governance policy before even knowing what systems the company has. It's a lot of work, it reassures the board for a quarter, and it doesn't change anything operationally.
For three years, the corporate question about AI was "what can we do with this?". From now on, it coexists with another: "can we demonstrate how we do it?".
The second question doesn't stop anyone. It stops those who built without architectural plans, permits, or registration, and are now discovering that traceability can't be added with a band-aid. For those who designed well, compliance is simply documenting something that already exists.
From August 2, 2026, the sanctioning powers over providers of general purpose models and the transparency obligations of Article 50 are effective: to report when a person interacts with an AI system, to mark artificially generated content and to declare the use of emotion recognition or biometric categorization.
Yes. The Digital Omnibus package postponed the obligations for high-risk systems: Annex III uses (personnel selection, credit scoring, biometrics, education, essential services) to December 2027, and Annex I uses (AI embedded in already regulated products) to August 2028. Transparency and GPAI obligations were not postponed.
The maximum penalties reach 35 million euros or 7% of the previous year's annual worldwide turnover, whichever is higher, for the most serious infringements related to prohibited practices.
The inventory assesses which AI systems actually exist within the organization, in which processes, with what data, and who is responsible for each. Risk is then categorized by use case, not by tool, and systems that handle personal data, decisions about people, or money are audited first.
The NIST AI Risk Management Framework is a voluntary risk management framework with a specific focus on generative AI; it is not certifiable. ISO/IEC 42001:2023 defines an AI management system that is certifiable, and therefore is what a corporate client or auditor will require proof of.
Not if it's designed from the beginning. Traceability, permission control, and the ability to replace the model are architectural decisions: incorporating them into the design has a marginal cost, adding them after the system is in production requires redoing integrations.
Do you know what AI systems are actually in place in your organization? Our AI audit diagnoses the model, data, traceability, and cost: what to save, what to rewrite, and what to stop. Fixed price and written deliverable. Request a diagnosis → |
Español Español 
English
Français
العربية
Português
Deutsch
Italiano
