EU AI Act explained: timeline, risk categories and what to do now

Or, how a delay became everyone’s favorite excuse
It’s safe to assume that right now, somewhere in your organization, someone is switching on an AI feature. Maybe it’s the shiny “summarize this” button that appeared in your CRM overnight. Maybe it’s a recruiter who just found out the hiring platform can rank candidates for them. One thing’s for certain, though: nobody consulted you. They didn’t fill out a form… And they sure as hell didn’t put it in your register either.
That, in a nutshell, is the problem the EU AI Act lands on.
The short version: the EU AI Act (Regulation (EU) 2024/1689) is the EU’s risk-based law for artificial intelligence. It entered into force on August 1, 2024, and applies in phases until August 2, 2028. The bans, AI literacy, the rules for general-purpose AI and the transparency duties already apply, but the heavy high-risk obligations follow in December 2027 and August 2028.
We’re Responsum, a privacy management software company from Zaventem, Belgium. We wrote this guide for privacy, legal and compliance teams who’d like a clear view of the AI Act without the breathless LinkedIn takes. It reflects the timeline as amended by the Digital Omnibus on AI, and it’s based on the official text on EUR-Lex and the European Commission’s AI Act Service Desk.
When does the EU AI Act apply? Spoiler alert: it already does
Unlike the universe, the AI Act didn’t arrive with one big bang. It arrived in installments, like a series that keeps getting renewed for one more season… and is currently still being renewed. Some rules have applied since February 2025, but the heaviest ones, for high-risk AI, don’t land until December 2027 and August 2028.
Just to paint a clear picture:
| Date | What applies |
|---|---|
| Aug 1, 2024 | The AI Act enters into force. |
| Feb 2, 2025 | Definitions, AI literacy (since softened by the Digital Omnibus) and the ban on prohibited AI practices apply. |
| Aug 2, 2025 | Obligations for providers of general-purpose AI (GPAI) models apply. Member States must designate national authorities and set penalties. |
| Aug 2, 2026 | Transparency rules (Article 50) apply, and the European AI Office can start fining providers of GPAI models. |
| Dec 2, 2026 | New bans on AI that generates non-consensual sexual deepfakes or child sexual abuse material apply. |
| Dec 2, 2026 | The marking grace period ends: generative AI systems already on the market before August 2, 2026, must now meet the Article 50(2) marking rules. |
| Aug 2, 2027 | Every Member State should run at least one AI regulatory sandbox. |
| Dec 2, 2027 | High-risk rules for Annex III systems apply. |
| Aug 2, 2028 | High-risk rules for AI in products covered by Annex I apply. |
Take another look at those December 2026 rows… That’s next month!
What changed with the Digital Omnibus?
The original Act set August 2, 2026, as the date for most high-risk obligations. Then the Digital Omnibus on AI, Regulation (EU) 2026/1744, came along and moved them. It was published in the Official Journal on July 24, 2026, and entered into force on July 27, 2026, a mere week before the old deadline. Cutting it fine doesn’t begin to describe it.
The EU’s stated reason: harmonized standards, guidance and national authorities simply weren’t ready in time. And fair enough, it’s hard to comply with a standard that doesn’t exist yet.
The thing is, a delay is not a rollback. The risk-based structure and core obligations stay. The transparency rules kept their August 2026 date. So if you read “AI Act delayed” somewhere and quietly closed the project folder… you might want to reopen it.
That said, the high-risk deadlines aren’t quite the cliff edge they look like either. High-risk systems already on the market before their deadline only fall under the new rules if they’re substantially modified afterward. And providers and deployers of AI systems intended for use by public authorities have until August 2, 2030, to comply.
The four risk categories, or: not all AI is created equal
The AI Act sorts AI systems by the harm they can cause. The higher the risk, the heavier the obligations. It’s an important distinction, because a spam filter and a CV-screening tool are both “AI” in the same way a garden hose and a fire hose are both “water,” which is technically true but practically useless.
| Risk category | Examples | What it means | Applies from |
|---|---|---|---|
| Unacceptable risk | Social scoring, manipulative AI, emotion recognition at work or school, untargeted scraping of facial images | Banned | Feb 2, 2025 (new bans Dec 2, 2026) |
| High risk | AI in hiring, credit scoring, education, biometrics, critical infrastructure | Risk management, data governance, documentation, human oversight, logging, conformity assessment | Dec 2, 2027 (Annex III), Aug 2, 2028 (Annex I) |
| Limited risk | Chatbots, deepfakes, AI-generated text and images | Tell people they’re dealing with AI; label synthetic content | Aug 2, 2026 |
| Minimal risk | Spam filters, AI in video games | No new obligations; voluntary codes | Not applicable |
What counts as high-risk AI?
There are two roads into the high-risk category. The first: your AI system is a safety component of a product that EU product law already covers, such as machinery or medical devices (Annex I). The second: it’s used in one of the sensitive areas listed in Annex III, like employment, education, access to essential services, law enforcement or migration.
There’s a filter, though. An Annex III system isn’t high-risk if it poses no significant risk, for example because it only performs a narrow procedural task. Tempting to use as a shortcut, right? Not so fast. Providers who rely on that exception must document their assessment, and any system that profiles people stays high-risk, no matter how modest it claims to be.
So it’s an exception. Not a loophole.
Where does general-purpose AI fit?
General-purpose AI models, the large language models behind the chatbots your colleagues have been “just trying out” for a while now, follow their own track. Their providers have had obligations since August 2, 2025: technical documentation, a copyright policy and a public summary of the training data. Models with systemic risk get extra homework on top, such as evaluations and incident reporting.
The good news for most of you: if you only use these tools, you’re a deployer, not a GPAI provider. That particular backpack belongs to whoever built the model. So unless your side project is the next ChatGPT, that one’s not your problem. Mostly.*
* The “mostly” is there for a reason. Modify a general-purpose model substantially enough and you become its provider, at least for the part you changed. The Commission’s guidelines put that bar at using more than a third of the compute it took to train the original model. And if you build your own AI system on top of someone else’s model and release it under your name, you’re the provider of that system, with the obligations that come with it.
Does the EU AI Act apply to my company?
If you build or use AI in the EU, it probably does. How much depends on your role, because the Act hands out obligations accordingly:
- Providers develop an AI system and place it on the market under their own name, so they carry most of the weight.
- Deployers use an AI system in a professional context: they follow the instructions for use, make sure a human keeps a hand on the wheel and, for some high-risk uses, run a fundamental rights impact assessment.
- Importers and distributors check that what they sell actually complies.
And no, being based outside the EU isn’t a get-out-of-jail-free card. The Act also reaches non-EU companies when their AI output is used in the EU.
Just to paint a clear picture, take a CV-screening tool. The software company that builds it and sells it under its own name is the provider. The bank that uses it to sort through job applications is the deployer. If the tool comes from an American vendor, the EU company that brings it onto the European market under that vendor’s name is the importer, and the reseller who sells it on to the bank is a distributor.
As for 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 breaches, whichever is higher. For SMEs and startups, and for the new category of “small mid-caps” the Digital Omnibus added, the cap is whichever of the two amounts is lower. Needless to say, that’s enough to make even a GDPR fine look modest. Well, almost.
What should you do about the EU AI Act now?
December 2027 sounds far away. So did May 25, 2018, once upon a time… and we all remember how relaxed the spring of 2018 was.
Once you count the time it takes for procurement, vendor questionnaires, contract updates and documentation, 2027 isn’t far at all. It’s roughly one budget cycle and a couple of very long email threads.
Your 90-day AI Act checklist
- ☐Build an inventory of every AI system you develop, buy or use. Yes, including the AI features quietly hiding inside your SaaS tools.
- ☐Classify each system by risk category and record your role (provider or deployer).
- ☐Confirm that nothing in your inventory is a prohibited practice.
- ☐Check transparency: are chatbots disclosed, and is synthetic content labeled?
- ☐Train the people who use or oversee AI (AI literacy).
- ☐Ask vendors for their AI Act documentation and add AI clauses to your contracts.
- ☐Give every high-risk candidate an owner and a review date.
How does the AI Act connect to the GDPR?
The AI Act doesn’t replace the GDPR. It moves in next door. When an AI system processes personal data, both apply, side by side.
A GDPR data protection impact assessment and an AI Act fundamental rights impact assessment overlap quite a bit, so run them together rather than twice. Once is plenty. The Digital Omnibus also clarifies the legal basis for processing special categories of personal data to detect and correct bias in AI systems, but only where that’s strictly necessary.
And here’s the part privacy teams tend to like: you’re not starting from zero. Your record of processing activities already maps where personal data flows, and most AI features sit right on top of those flows. That makes your register the natural starting point for your AI inventory.
Frequently asked questions about the EU AI Act
Is the EU AI Act already in force?
Yes. It entered into force on August 1, 2024. Prohibitions and a (since softened) AI literacy duty have applied since February 2025, GPAI rules since August 2025 and transparency rules since August 2026.
Has the EU AI Act been delayed?
Partly. The Digital Omnibus on AI moved high-risk obligations from August 2, 2026, to December 2, 2027 (Annex III) and August 2, 2028 (Annex I). The other deadlines didn’t move.
Who enforces the AI Act in Belgium?
National market surveillance authorities, with the European AI Office supervising general-purpose AI models. At the time of writing, Belgium hasn’t formally designated its authority in law yet. The federal government agreement points to the BIPT, so that’s the name to watch.
Does the AI Act apply if we only use ChatGPT or Copilot?
Usually as a deployer, which means AI literacy, transparency where relevant and sound governance. High-risk obligations only kick in if you use AI for a high-risk purpose, such as screening job candidates.
So, about that button in your CRM…
Nobody expects you to have every AI system classified, assessed and documented by Friday. But you should know they exist, what they do with personal data and who’s responsible for them. That’s not a 2027 problem. That’s a this-quarter problem.
Start your AI Act readiness with Responsum. Your privacy register already knows where personal data flows. Time to introduce it to your AI.