In short
- The AI Act is not on hold. Its general rules have applied since 2 August 2026.
- On 2 December 2026, existing AI systems must mark their generated content in a machine-readable way, and new bans take effect.
- Prohibited practices, AI literacy and transparency rules already apply today.
- Only the heavy obligations for high-risk AI were postponed: to 2 December 2027 (Annex III) and 2 August 2028 (Annex I).
- If you use AI in HR, recruitment or education, you have just over a year to prepare.
When the EU agreed to delay parts of the AI Act this summer, many companies breathed a sigh of relief. Some concluded the AI Act was on hold. It isn’t.
The Digital Omnibus on AI entered into force on 27 July 2026. It gave companies more time for the heaviest obligations, but it did not pause the law. And on 2 December 2026, less than three weeks from now, two more obligations kick in.
In this article, you’ll learn exactly what changes, what already applies, what was postponed, and which steps to take now if your organisation builds, buys or uses AI in its processes.
What is the AI Act?
The AI Act (officially Regulation (EU) 2024/1689) is the European law that sets rules for anyone who develops, sells or uses AI systems. It’s a regulation, not a directive. That means it applies directly in every member state, including Belgium, without the need for a national law first.
The AI Act takes a risk-based approach. The greater the risk an AI application poses to people, the stricter the rules:
- Unacceptable risk: banned. Think of social scoring or manipulative AI.
- High risk: allowed, but with strict requirements for documentation, testing, human oversight and registration. For example, AI used in recruitment or credit scoring.
- Limited risk: transparency obligations. People must know they’re talking to AI or looking at AI-generated content.
- Minimal risk: no specific obligations. Think of a spam filter.
Most companies use AI mainly in the last two categories. But as soon as AI supports decisions about people, that can change quickly.
What is the Digital Omnibus on AI?
The Digital Omnibus on AI (Regulation (EU) 2026/1744) is an amendment to the AI Act that the European Commission proposed in November 2025. Its goal is to simplify the rules and give companies more time, because the technical standards for high-risk AI weren’t ready yet.
After a political agreement in May 2026, the text was formally adopted in July. It was published on 24 July 2026 and entered into force on 27 July 2026, just before the original 2 August deadline.
The key thing to remember: the Omnibus is a targeted delay, not a general moratorium.
What changes on 2 December 2026?
1. AI-generated content must be machine-readable. Since 2 August 2026, providers of AI systems that generate synthetic audio, images, video or text must mark their output in a machine-readable way. That way it can be detected as AI-generated, for example through a watermark or metadata. Systems that were already on the market before 2 August got a grace period. That grace period ends on 2 December 2026.
Have you built a tool that generates content with AI, such as product descriptions, reports, images or customer emails? Or added AI generation to an existing application? Then check now whether the output is marked. If you use a tool from an external vendor, ask how the vendor handles this.
2. New prohibited practices. The Omnibus adds new bans to the AI Act: AI systems that generate non-consensual intimate content (so-called nudifier apps) and child sexual abuse material. These bans apply from 2 December 2026, with no exemption for systems already on the market.
Most businesses won’t be affected directly. But if you offer image or video generation, your safeguards need to demonstrably hold up.
What already applies today?
The delay did not touch these obligations:
- Prohibited AI practices have applied since 2 February 2025.
- AI literacy has applied since 2 February 2025. The Omnibus softened the wording: the Commission and member states now support AI literacy, rather than every organisation having to ensure it alone. The obligation still exists, though. Your people need to understand the AI tools they work with, including their limitations.
- Transparency (Article 50) has applied to new systems since 2 August 2026. A chatbot must tell people they’re talking to AI. Deepfakes must be labelled.
- Fines can be imposed since 2 August 2026.
What was actually postponed?
The real delay concerns high-risk AI systems: the technical requirements, documentation, risk management, quality management and conformity assessments.
Obligation | Applies from |
|---|---|
Prohibited practices and AI literacy | 2 February 2025 |
Transparency for new systems (Art. 50) | 2 August 2026 |
Machine-readable marking for existing systems (Art. 50(2)) | 2 December 2026 |
New bans (nudifier apps, CSAM) | 2 December 2026 |
High-risk systems, Annex III (e.g. HR, education) | 2 December 2027 |
High-risk AI in regulated products, Annex I (e.g. machinery, medical devices) | 2 August 2028 |
High-risk systems used by public authorities | 2 August 2030 |
Annex III covers standalone AI systems in sensitive areas. For businesses, these are the most relevant: recruitment and selection, decisions on promotion, dismissal or task allocation, employee evaluation, access to education and creditworthiness assessment.
The Omnibus also extended the relief for SMEs to small mid-caps, with simpler technical documentation and more flexible post-market monitoring. If you rely on the Article 6(3) exemption because you believe your system isn’t high-risk after all, you still have to register it in the EU database, although a simplified procedure now applies.
Provider or deployer: which role do you have?
The AI Act makes an important distinction between two roles, and your obligations depend on it:
- Provider: you develop an AI system, or have it developed, and place it on the market or put it into service under your own name. Providers carry the heaviest obligations: documentation, testing, conformity assessment and marking generated content.
- Deployer: you use an AI system within your own organisation, for example an AI tool to screen CVs. You must use the system according to its instructions, organise human oversight, keep logs and inform the people affected.
Watch out: many companies become providers without realising it. If you integrate an AI model into your own application and offer it to customers, or substantially modify an existing system, you may be considered a provider yourself. That happens more often than you’d think with custom software and automation.
What about Belgium?
Because the AI Act is a regulation, its rules apply in Belgium without any additional law. Supervision, however, has to be organised nationally. A government declaration in January 2025 designated the BIPT (the Belgian Institute for Postal Services and Telecommunications) as the main supervisory authority, with the FPS Economy coordinating implementation.
At the end of July 2026, however, that appointment had not yet been made official, Belgian broadcaster VRT reported. That doesn’t mean you can ignore the rules. The obligations apply directly, and a supervisor can still act later on past infringements. Keep an eye on the BIPT website for the latest status.
What are the penalties for non-compliance?
The AI Act sets three tiers of fines:
- Up to €35 million or 7% of worldwide annual turnover for prohibited AI practices.
- Up to €15 million or 3% for most other obligations, including the transparency rules and high-risk obligations.
- Up to €7.5 million or 1% for supplying incorrect or misleading information to authorities.
For large companies, the higher of the two amounts applies. For SMEs and start-ups, the lower one applies. But even for an SME, a fine isn’t the only cost. An AI system that has to be modified or taken out of use after the fact often costs more than the fine itself.
Why “we have until 2027” is the wrong takeaway
A year sounds like a lot. In practice, it isn’t.
Compliance with the high-risk rules is not a document you write the week before the deadline. It means knowing which AI runs where in your organisation, how it was tested, what data it uses, who oversees it and how you log its decisions. If your AI is woven into processes that were never designed with that in mind, the work is in the processes, not the paperwork.
We often see the same pattern in practice. A team starts using an AI tool. An existing automation gets an AI step added. A chatbot goes live on the website. A year later, nobody knows exactly where AI influences decisions. You can’t classify, test or document what you can’t see.
That’s exactly where companies that build quality in from the start have the advantage. Traceability, testing and human oversight aren’t extra steps you bolt on at the end. They’re part of a well-designed process. And that process pays off beyond compliance too: fewer errors, more first-time-right and more trust from your customers.
Five steps to take before the end of the year
- Map your AI. List every AI tool and automation in your organisation, including the ones teams started using on their own. Note for each tool what it does, what data it uses and who is responsible.
- Determine your role. For each system, check whether you’re the provider or the deployer. With custom software or your own integrations, you’re often the provider.
- Check your generated content. Does any of your tooling produce AI-generated text, images or audio? Make sure the output is marked before 2 December. If you work with a vendor, ask in writing how they handle it.
- Classify your risk. Determine which systems could fall under Annex III. Those need a plan for 2027 now, with time for testing, documentation and any changes.
- Build quality into the process. Testing, logging and human oversight are easier to design in than to retrofit. Include them in every new automation, and make sure your people know how to work with the AI tools.
How Recce can help
At Recce, we optimise processes with intelligent automation, and we bring quality in from the very first idea. That makes us a natural partner when AI has to be both effective and controllable.
In practice, we help you:
- map where AI and automation sit in your processes;
- design automation that’s traceable, testable and well documented;
- build quality checks and human oversight into your processes;
- make sure the AI you build today won’t need rebuilding in 2027.
Want to know where your organisation stands? Get in touch and we’ll take a look together.
Frequently asked questions about the AI Act
Has the AI Act been postponed?
No, not entirely. The AI Act has applied since 2 August 2026. Only the obligations for high-risk AI systems were postponed, to 2 December 2027 (Annex III) and 2 August 2028 (Annex I).
What changes on 2 December 2026?
AI systems that were already on the market before 2 August 2026 must then mark their generated content in a machine-readable way. New bans also take effect on AI that generates non-consensual intimate content or child sexual abuse material.
Does the AI Act apply to SMEs?
Yes. The AI Act applies to every organisation that develops or uses AI, regardless of size. SMEs and small mid-caps do get some relief, such as simpler documentation and lower maximum fines.
Do companies need to ensure AI literacy?
Yes. The obligation has applied since 2 February 2025. The Omnibus softened the wording, but your people still need to understand enough about the AI tools they work with.
When is my AI system high-risk?
Among other cases, when it’s used for recruitment, employee evaluation, education, credit scoring or as a safety component in regulated products such as machinery or medical devices. If in doubt, have it assessed.
Who supervises the AI Act in Belgium?
The BIPT has been designated as the main supervisory authority, with the FPS Economy coordinating. At the end of July 2026, the formal appointment had not yet been completed. The obligations apply in the meantime.
How high are the fines?
Up to €35 million or 7% of worldwide annual turnover for prohibited practices, and up to €15 million or 3% for most other infringements. For SMEs, the lower of the two amounts applies.
This article is general information, not legal advice. For a legal assessment of your specific situation, consult a specialised lawyer.