1. Why the timeline matters more than any single rule
The AI Act, Regulation (EU) 2024/1689, does not come into force all at once. Its provisions apply in phases: some prohibitions and basic obligations are already operational, while others concern general-purpose models, generated content and high-risk systems and arrive later. For a company the risk is not so much ignoring the Regulation as moving forward without knowing which obligation is already active and which is not.
The timeline must be read together with another question: which AI systems do you actually use? Without an inventory, the deadlines remain theory. Many companies discover that they use AI functions inside software they already pay for, such as writing tools, candidate screening or document analysis. The first step is not to read the Regulation line by line, but to understand which systems fall within scope and which people they affect.
2. The phases of application, in order
The Regulation follows a cascading logic. The rules on prohibited practices and the obligation to train staff on the use of AI come first. Then the rules for general-purpose models come into play, meaning the foundation models on which many products are built. Transparency obligations follow, along with the broader body of compliance requirements, including the high-risk systems listed in the Regulation. Finally, there are high-risk systems embedded in products already regulated by other safety legislation, which have longer timelines.
The timeline may change. Subsequent legislative measures may shift some dates or alter the scope of certain obligations. Before setting internal commitments, check the up-to-date status of the official text and ask your adviser which provisions are operational on the date you are reading. For this reason the plan must be reviewed at regular intervals, not filed once it is approved. Whoever updates it must record the date and source of each change.
| Area | What it covers | What to do now |
|---|---|---|
| Prohibited practices | Uses of AI considered unacceptable | Verify that no system falls within these uses |
| AI literacy | Understanding of the tool by those who use it | Prepare a training plan by role |
| General-purpose models | Obligations of those who develop foundation models | Request documentation and conditions from suppliers |
| Transparency | Inform people who interact with an AI and label synthetic content | Map chatbots, emails and generated content |
| High risk | Systems on the Regulation's list or in regulated products | Classify systems and understand your own role |
3. What is already operational: prohibitions and training
The prohibitions on unacceptable practices and the obligation of AI literacy are among the parts that apply first. The former covers uses such as manipulation that exploits a person's vulnerabilities, or emotion recognition in the workplace, except for medical or safety reasons. The latter covers anyone using AI systems in the company, regardless of risk level: the organisation must ensure that staff have an adequate understanding of the tool.
For most small and medium-sized enterprises, therefore, the first concrete piece of work is this: verify that no use falls under the prohibitions and start targeted training. A generic course that is the same for everyone is not needed. A salesperson who uses a text generator needs to know which data they may enter; someone using a staff-selection tool must understand the limits of its output and when a person needs to review it.
4. General-purpose models and the supply chain
General-purpose models are models trained on large amounts of data and suited to very different tasks. The main obligations fall on those who develop them and place them on the market: technical documentation, information for those who integrate them, and a policy on respecting copyright. Anyone who buys a service built on these models does not automatically become responsible for everything, but must know what they have received from the supplier and under which conditions they can use it.
For your company the useful question is: are you simply using a tool, or are you building a product on top of a model? In the second case your role may change, especially if you substantially modify the system or put your own name on it. Ask the supplier for the available documentation and the conditions of use, and check with your adviser whether the modification you are planning moves you into a different role.
5. Transparency and generated content
Transparency obligations concern those who interact directly with an AI system, such as a chatbot on the website or a voice assistant for customers: the person must know they are speaking with a machine, unless this is obvious from the context. They also concern synthetic content, such as images, audio or video that are generated or altered, which must be recognisable as such when they could be mistaken for real.
For marketing and communications the most sensitive point is labelling. A realistic image of a person who does not exist, or a cloned voice, must be clearly flagged. Prepare a simple internal rule to apply across all channels, rather than deciding case by case in every department. The message should be short and identical across all channels, so that the public does not receive different versions of the same notice.
6. High-risk systems: what is asked of those who use them
A system is high-risk when it falls within the list of use cases set out in the Regulation or when it is a safety component of an already regulated product. Typical examples include staff selection, creditworthiness assessment and certain decisions on access to essential services. Those who use these systems in the company have obligations of their own: use in line with the instructions, human oversight entrusted to competent people, control of the input data and monitoring of how the system operates.
For some deployers a fundamental rights impact assessment is also required. It depends on the type of entity and the use, so it must be checked case by case. If it applies to you, it must be prepared before the system starts, not afterwards. We cover it in a dedicated guide. Ask your adviser whether your organisation falls within these entities, and record the answer together with the date on which you received it.
7. How to order the work in phases
A sensible order avoids spending time and budget on the less urgent parts. The steps below start from what can be done immediately and move towards continuous management, without requiring everything to be ready on the first day. A shared order of priority is needed, with an owner for each step and a review date. This order should be written down and shared with those who manage the systems, because otherwise everyone will interpret it in their own way.
- Inventory the AI systems in use or purchased, including functions built into software.
- Immediately verify that no use falls under the prohibitions, and record the decision with its reasons.
- Classify each system by risk level and by your role: developer, integrator or simple user.
- Start staff training, differentiated by role.
- Map the transparency information for chatbots, generated content and communications.
- For high-risk systems, plan human oversight, log retention and information for the people involved.
- Review the plan whenever the applicable timeline, the use of the systems or the supplier changes.
8. What to read first and what to explore further
The cluster covers several questions that deserve dedicated guides. The first is the classification of systems by risk level, because almost everything else depends on it. Next come the prohibitions, which should be read with concrete examples, staff training, the obligations for those who use high-risk systems and the fundamental rights impact assessment. These are topics to explore separately, each with an operational checklist.
The relationship between the AI Act and the GDPR deserves a separate comparison, because the two regimes overlap on personal data. If you use AI to make decisions about people, also read the guide on human resources. The first topic to explore, however, is classification: without it you do not know which guide you really need. Once classification is clear, the rest of the timeline becomes much easier to read.
9. How to start
The first concrete step this week is a list of the AI systems in use, with three columns: who uses it, for which activity and on which data. A perfect document is not needed. You need a list that your adviser and IT supplier can complete together, starting with the tools that touch customers and staff. Also ask your IT supplier which AI functions are active today and which are planned in upcoming updates.
Then book a meeting with the compliance adviser to confirm the timeline applicable to your situation. Bring the list and your questions about your role in relation to each system. A timeline read in light of your systems is worth far more than a generic one, because it shows where to start and what can wait. Also ask them to name a person who remains responsible for the timeline and its review.
Frequently Asked Questions
Do I already have to comply with the AI Act if I only use third-party tools?
In part, yes. Even those who only use third-party tools have obligations, in particular on staff training and compliance with the prohibitions. The detail depends on the type of system and how you use it. If the system is high-risk or interacts with customers, the transparency and proper-use obligations become more concrete. The supplier must provide you with information: ask for it in writing.
Is the AI Act timeline already fixed in a final form?
It must be checked case by case. The Regulation sets out phases of application, but subsequent legislative measures may change the dates or the scope of some obligations. For this reason you should not build a plan on a date read in a newspaper article. Check the up-to-date official text and get confirmation from an adviser before making internal commitments.
Who should handle AI Act compliance in the company?
It is usually a shared responsibility. Management sets the priorities, the IT lead maintains the inventory of systems, the HR or legal function checks uses that affect people, and the adviser helps interpret the Regulation. Each system must have a clear internal owner, because a decision that belongs to nobody tends not to be taken.
How long does it take to prepare a compliance plan?
It depends on the number of systems, their impact on people and the level of documentation already available. A company with few tools and a recent inventory can complete the first phase in a short time. Those with many software packages containing built-in AI functions need more time, especially to verify with suppliers what the systems actually do.
Does the software supplier take care of everything?
No. The supplier has its own obligations for the system it develops, but you remain responsible for how you use it in your organisation. Ask for documentation, instructions for use, information on known limitations and risks, and a clear statement of how the system handles data. A contract that provides this information helps you demonstrate due diligence.
Is a system used only internally excluded from the Regulation?
Not necessarily. The fact that a system is used only inside the company does not exclude it from scope. What matters is what it does, whom it affects and in which context. An internal system that evaluates candidates or employees may be subject to specific obligations. For this reason classification must be carried out on the actual use, not only on the internal nature of the tool.