Blog

What changes for HR?

The AI Act and assessments

"Nancy, will we still be allowed to use those tests?" That was the question from an HR manager, dropped in somewhere between two other points, as if it were a detail. She had read something about the AI Act and high-risk, and she was not comfortable with it.

I completely understand. A lot gets written about it and very little gets explained. And if you run an SME with five vacancies and no legal department, you just want to know one thing: do I need to do something now, yes or no?

The short answer: a psychometric test is not automatically high-risk AI in itself. But AI that interprets test results, scores candidates or gives selection advice can be. So the difference is not so much the fact that you use a test, but how that test works technically and above all what the system does with the results.

And the main obligations have meanwhile been pushed back to 2 December 2027. So you have time. Which does not mean you have nothing to do if you let AI interpret your tests.

What the AI Act actually says about recruitment and selection

The AI Act classifies AI systems by risk. At the top are the prohibited practices, below that the high-risk category, and that last one is the interesting one for us.

Annex III of the regulation lists the applications that count as high-risk. Point 4 of that list covers employment and workforce management. It includes AI systems used to analyse and filter job applications and to evaluate candidates.

So AI systems that evaluate, filter, rank or select candidates in recruitment fall, in principle, under the high-risk category of Annex III. The AI Act does provide a few limited exceptions.

That is not new and it is not unreasonable either. Who gets to work and who does not touches someone's life.

So what matters to the legislator is not simply that you use a test. It is how the system works, what it does with the information, and what role the output plays in your selection process.

What changed in July 2026

On 27 July 2026 the so-called AI Omnibus entered into force: Regulation (EU) 2026/1744. It moved the main high-risk obligations for Annex III systems from 2 August 2026 to 2 December 2027.

The reason was fairly simple: the technical standards were not developed far enough, and supervision and conformity assessment were not fully ready either. It is hard to require companies to comply in full when part of the practical framework is still missing.

What did stay in force, and this is often forgotten:

The prohibited AI practices have applied since 2 February 2025. Think of emotion recognition in a work context. AI that tries to infer from facial expressions or tone of voice whether someone is nervous, enthusiastic or insecure is in principle not allowed, apart from a few narrow exceptions for medical or safety reasons.
The rules for general-purpose AI models have applied since 2 August 2025. Think of broad AI models that can be used for all kinds of applications. Those rules are mainly aimed at organisations that develop such models or place them on the European market.
The transparency obligations under Article 50 have applied since 2 August 2026 and were not postponed. In certain situations people must know that they are communicating directly with an AI system, for example a chatbot or an AI agent.

In short

Prohibited AI practices, 2 February 2025, already in force
General-purpose AI models, 2 August 2025, already in force
Transparency obligations under Article 50, 2 August 2026, not postponed
High-risk Annex III, including certain applications in recruitment and selection, 2 December 2027
High-risk in products, Annex I, 2 August 2028. So that does not cover the classic recruitment applications from Annex III.

Where the line runs between a test and an AI system

A classic psychometric test works with a fixed calculation. The candidate fills in a questionnaire, the answers go through a norm based on research, or through a predetermined calculation, and a score comes out. The software follows fixed rules that people set in advance.

An AI system goes further than software that only executes rules laid down by people beforehand. A key characteristic in the AI Act is that the system infers from the input how it generates certain output, such as a prediction, a recommendation or a decision.

In recruitment it becomes important as soon as AI is used to evaluate, filter or rank candidates, or to support decisions in the selection process. Those applications are listed explicitly in Annex III of the AI Act.

How we do it at TalentTester

Our tests are scored without AI. The calculation behind the reports is a fixed psychometric norm or a fixed calculation. There is no AI model analysing or ranking candidates, and the system does not decide who you should or should not hire.

That is not a marketing choice, it is simply how these instruments are built. And let us be honest: it makes your life a good deal simpler, because you remain the one who decides. I give you a report, you place it next to your interview and the other information you have about the candidate.

Will I ever add AI to the interpretation? Not for now.

That goes for all our tests: the Big Five personality test, the 4 colours personality test, the aptitude test and the sales assessment.

What you can already do now, even though it is 2027

December 2027 seems far away. But the question you will need to answer is this: which systems do you use, what do they do, and what role do they play in your decision? Working that out takes time, and you would rather not be doing it in November 2027.

Make a list of every tool in your hiring process. Including your ATS, your job site, your CV screener and the AI features that quietly crept into your software.
Ask each supplier whether there is AI in it, what that AI does exactly, and whether the system evaluates, filters or ranks candidates. Get the answer in writing, not just over the phone.
Establish who takes the decision. For high-risk AI, human oversight is mandatory. The person working with it must be able to assess the output, question it and depart from it.
Check when you have to inform candidates. Transparency rules apply to systems that communicate directly with candidates, such as certain chatbots or AI agents. And if you use a high-risk AI system from Annex III that takes or supports decisions about candidates, a specific duty to inform those candidates applies on top of that.
Keep track of which systems you use, what you use them for and what role they play in your selection process. If you later have to explain how a decision came about, you do not want to be reconstructing it after the fact.
You can start on points four and five today. There is really no need to wait until December 2027.

What I think of it

I am not a lawyer and I am not going to pretend otherwise. For the finer points you call a lawyer, and for larger organisations that is genuinely necessary.

But I do not find the thinking behind this law unreasonable. If AI helps determine who moves forward in a selection, then you should know what that system does, there should be human oversight, and you should inform candidates where that is needed. For much of that we did not really need a European regulation. It should simply be normal.

What does bother me is the way the word AI now gets stuck on everything. I see providers selling their assessment as AI-powered because it sounds better. But if there really is AI in it, and that AI evaluates, filters or ranks candidates, that can have genuine legal consequences. So that client may be buying a whole set of extra obligations without realising it.

So always keep asking.

Frequently asked questions about the AI Act and assessments

Does a personality test fall under the AI Act?

Not automatically. A test with a fixed psychometric calculation is not an AI system in itself. If AI is used to interpret results, or to evaluate, filter or rank candidates, then the system can fall under the high-risk category for recruitment and selection.

At TalentTester, the scoring and interpretation of the tests happen without AI.

From when do I need to be compliant?

The main high-risk obligations for Annex III systems, which cover certain AI applications in recruitment and selection, apply from 2 December 2027.

Some transparency obligations have applied since 2 August 2026, for example when a candidate communicates directly with an AI system.

Can an AI system decide who I hire?

For high-risk AI, human oversight is mandatory. The person working with the system must be able to understand the output, assess it and depart from it where necessary. So you cannot simply follow the output blindly.

Separately from the AI Act, Article 22 of the GDPR has applied for much longer. That article restricts decisions taken solely by automated means which produce legal effects or similarly significantly affect someone. There are specific exceptions and conditions to that.

What if my supplier says their tool uses AI?

Then ask in writing what the system actually does. Does it evaluate, filter or rank candidates? Does it give predictions or recommendations that you use in the selection?

Also ask what documentation the supplier provides, so that you know which obligations apply to you.