AI in recruitment: what the AI Act already requires
Last updated: August 2026
You use a tool that ranks CVs by relevance. Or a chatbot that asks three questions before passing a candidate on to you. Or a matching score between a profile and an assignment. If so, you fall under the European regulation on artificial intelligence, better known as the AI Act.
The calendar has just changed. 2 August 2026 was supposed to be the date on which recruitment tools moved into a strict regime. That is no longer the case: the heaviest obligations have been postponed.
The trap would be to conclude there is nothing to do. Two obligations have applied in full since February 2025, and they have largely stayed under the radar. One of them carries the highest fine ceiling in the entire regulation: 35 million euros or 7% of worldwide turnover.
In this article: what changed this summer in the European calendar, and the two obligations already in force. Then what awaits you in December 2027, and a six-point checklist for using AI without exposing yourself.
Is a recruitment tool a high-risk AI?
Under the logic of the AI Act, yes. The regulation classifies systems by their use, and employment sits among the most sensitive areas. A system used to publish vacancies, filter applications, rank candidates or evaluate workers falls into the "high-risk" category, listed in Annex III of the regulation.
What counts is not the product's commercial label. It makes no difference whether a vendor talks about an algorithm, a matching engine, intelligent automation or assisted sorting. The criterion is the function performed: does this system contribute to deciding who moves to the next stage?
Three examples to show where the line falls.
- A dashboard showing your applications by date of arrival: that is not AI, it is a display.
- A tool ranking your candidates by matching score against the job description: high risk.
- A chatbot asking eliminating questions and screening out profiles that do not tick the boxes: high risk.
The same reasoning applies to managing workers already in post: task allocation, performance evaluation or decisions on promotion and dismissal fall within the same framework.
What changed on 27 July 2026
An essential and very recent calendar update.
European regulation 2026/1744, known as the Digital Omnibus, was published in the Official Journal of the European Union on 24 July 2026 and entered into force three days later, on 27 July 2026. It amends the AI Act's application calendar. Obligations for high-risk systems under Annex III, including recruitment, move from 2 August 2026 to 2 December 2027. For high-risk systems embedded in regulated products, the deadline moves to 2 August 2028.
Two ways to read this postponement, and both matter.
It is a delay in application, not a repeal. The obligations exist, their enforceability shifts. The text you will have to comply with in December 2027 is already written and known. Nothing improves by waiting.
The delay does not touch everything. The chapters of the regulation that already apply are unaffected. And those are precisely the ones containing immediate prohibitions.
Two obligations in force since February 2025
Here is the essential part. Since 2 February 2025, two blocks of the regulation have applied in full, regardless of any postponement.
1. Prohibited practices
Some uses of AI are not regulated: they are simply banned. Two of them bear directly on HR.
Emotion recognition in the workplace. A system that claims to infer a person's emotional state from their voice, face, gaze or posture is prohibited in the work context. The regulation provides strict exceptions for medical or safety reasons, for example to detect driver drowsiness. Those exceptions do not cover the evaluation of a candidate or a worker.
So this targets video interview tools that promise to measure a candidate's confidence, motivation or sincerity. Such features were sold for years. They no longer have a place in a selection process in Europe.
Biometric categorisation. A system that infers from biometric data a person's ethnic origin, religious or philosophical beliefs, political opinions, sex life or trade union membership is prohibited. In an HR context, the mention of trade union membership is not a theoretical detail.
These prohibitions apply directly: they do not depend on a Belgian transposition law, nor on the full designation of supervisory authorities.
2. AI literacy across your teams
This is the least visible obligation in the regulation, and the broadest. Any organisation deploying AI systems must ensure that the people using them understand the tool. What it does, what it does not do, its limits and the risks it carries. The expected level depends on each person's role.
Translated into your daily work: whoever uses your CV screening software must know on what basis that software ranks profiles, and why they cannot follow its ranking blindly. This does not call for a university course. It calls for real, tailored and traceable training.
Which penalties, for which breach?
The regulation organises fines in tiers. Here is what concerns an employer.
| Breach | Fine ceiling | Applicable since |
|---|---|---|
| Use of a prohibited practice (emotion recognition at work, biometric categorisation) | 35 million euros or 7% of worldwide annual turnover | 2 February 2025 |
| Failure to meet high-risk system obligations | 15 million euros or 3% of worldwide annual turnover | 2 December 2027 |
For companies, the higher of the two amounts applies. For SMEs the regulation provides for proportionality, but the principle stands: the prohibited practice is the most heavily penalised breach in the whole text, and it is the one already enforceable.
What awaits you in December 2027, as a user
Watch out for a common misunderstanding: you are not the system's provider, you are its user, what the regulation calls the deployer. You have your own obligations, distinct from those of the software vendor.
The main ones, to prepare now:
Follow the instructions for use. Use the system within the conditions set by the provider, and not beyond them.
Assign human oversight to competent people. Naming someone is not enough. That person must have the competence, the training, the authority and the means to contradict the machine. A recruiter with neither the time nor the mandate to depart from the automatic ranking does not constitute human oversight.
Ensure input data is relevant. If your historical recruitment data carries a bias, your tool will learn it.
Monitor operation and report serious incidents.
Keep the system logs for at least six months.
Inform your workers before deployment. Before putting a high-risk system to work in the workplace, you must inform the workers concerned and their representatives. In Belgium this dovetails with informing and consulting the works council or the trade union delegation, and with the place of such tools in your work regulations.
One point not to miss: these obligations cannot be transferred to the provider by contract. A clause claiming to place compliance on the vendor does not protect you. You remain responsible for the use you make of the tool.
AI does not exempt you from Belgian employment law
The AI Act adds to existing rules, it does not replace them. Three intersections to keep in mind.
The GDPR still applies in full. An application is a set of personal data. Legal basis, retention period, information to candidates, right of access: nothing has gone away. And a decision taken entirely by a machine, with a significant effect on the person, is already strictly regulated.
Anti-discrimination legislation still applies. An algorithm that statistically screens out candidates of a certain age or gender produces discrimination, even without intent. And you will not be able to hide behind the tool.
Pay transparency is arriving in parallel. The ban on asking a candidate about previous pay will also apply to an automated form or a screening chatbot. We cover this in our article on what changes in your job ads.
Using AI without exposing yourself: 6 points
1. Inventory your tools and their real function
List everything that touches your applications: recruitment platform, scoring module, chatbot, video interview tool, CV analysis extension. For each, one question: does this tool contribute to screening out or ranking a person?
If you are comparing solutions, our article on the best software for registering temp workers sets out which features to expect from a tool, and which you can do without.
2. Remove emotion analysis
Check the features of your video interview tools. If one of them offers a measure of engagement, enthusiasm or reliability based on face or voice, switch it off. This is not a grey area: it has been prohibited since February 2025.
3. Keep a documented human decision
Automatic ranking can help you set priorities. It must not decide a rejection on its own. Keep a record of who validated what.
4. Train the people who screen
A short session is often enough: what the tool does, what data it relies on, what errors it typically makes, how to contradict it. Keep a record of the training, it is part of the obligation.
5. Inform candidates and workers
One clear sentence in your job ad and in your internal procedure beats a discovery in the middle of a dispute. For workers, that information belongs in the induction itself: see our onboarding step plan for a new employee.
6. Require your provider's documentation
Ask for the instructions for use, the description of training data, the known limits and the results of bias testing. A provider who cannot supply that today will not be ready in December 2027, and you are the one using their product.
What employers should remember
The delay changes the date, not the work. Obligations for high-risk recruitment systems take effect on 2 December 2027 instead of 2 August 2026. The content itself is already written.
Two things are already enforceable. The ban on analysing emotions at work and the training of people who use AI have applied since 2 February 2025, with a fine ceiling of 35 million euros or 7% of worldwide turnover for prohibited practices.
You are responsible for use, not only the vendor. A system user's obligations cannot be contracted away. Real, documented human oversight is your best protection.
How Recruit helps you
Hiring temporary staff without administrative worries? That is why we built Recruit.
With Recruit:
- Contracts in under 60 seconds - Create trial and agency contracts without the hassle
- Automatic salary calculation - Correct pay under current legislation and your joint committee
- Dimona and social documents - We handle every declaration, you focus on your business
- The decision stays yours - You choose who works and when, with no imposed automatic screening
- Personal support 24/7 - Always a dedicated contact who knows your company
Try Recruit or contact our HR experts for tailored advice.
Frequently asked questions
Is my recruitment software covered by the AI Act?
If it filters, ranks or recommends candidates, it falls into the high-risk category of Annex III. What counts is the function performed, not the product's commercial name. A simple display of applications by date is not covered.
Does the delay to December 2027 exempt me from everything?
No. It shifts the obligations for high-risk systems. Prohibited practices and the AI literacy obligation have applied since 2 February 2025 and are not affected by that delay.
Can I analyse a candidate's voice or face?
Not if the aim is to infer their emotional state. Emotion recognition in the workplace is a prohibited practice, with strict exceptions limited to medical or safety grounds, which do not cover evaluating a candidate.
Must I tell my workers that I use AI?
Yes. Informing workers and their representatives before a high-risk system is put into service is part of the user's obligations. In Belgium this dovetails with your internal information and consultation duties.
Who is responsible, me or the software vendor?
Both, on different fronts. The vendor answers for the system's compliance, you answer for the use you make of it. Your obligations as a user cannot be transferred to the vendor by a contractual clause.
The information in this article is purely indicative and does not replace professional legal or accounting advice. Legislation changes regularly. Always consult the applicable legislation or contact an HR expert for advice tailored to your situation.