Regulation in PracticeThe AI Act in HR does not start with software, but with decision-making
In many HR teams, the European AI Act is read as a technology issue. In practice, it is less about tools than about how recommendations influence decisions concerning people.
Björn Slegers ·
AI in HR has entered a new phase. While the conversation previously centered on experimentation, productivity, and handy applications, attention is now shifting toward accountability. The European AI Act accelerates that shift. Not because every HR department must suddenly become legal specialists, but because it is becoming clear that technology in workforce processes is never purely an IT choice.
In many organizations, the response begins predictably. An inventory of tools is drawn up, conversations with vendors take place, and perhaps a risk assessment template is introduced. That is understandable, but it only scratches the surface of reality. The relevant question is rarely just which software is being used. The far more relevant question is where recommendations, scores, or generated text begin to influence decisions about people.
That is precisely where AI in HR is most often underestimated. Not in spectacular use cases, but in the small accelerations of everyday work. A job posting that is automatically refined. An initial ranking of candidates. A summary of performance reviews. A suggestion for internal mobility. A prediction of employee turnover. On paper, the human remains the decision-maker. In practice, judgment often shifts much earlier, leaning toward the system that provides the initial interpretation.
This makes the AI Act relevant even for organizations that believe they do very little with AI. After all, the debate is not solely about fully automated decisions. It is also about the extent to which a recommendation guides human judgment, how transparently that happens, and whether an individual can still genuinely deviate from it. Anyone who only checks whether a button is formally pressed by a manager misses the real risk.
Therefore, the first mature question is not whether this tool is permitted. The first mature question is: in which process does this tool intervene? Only then do the practical questions that truly matter follow. What data feeds the model, for what purpose, in what context, and who evaluates the outcome? What remains visible to the employee or candidate, what remains contestable for the manager, and what can still be explained in retrospect when requested?
In HR, the most difficult problems rarely stem from overtly bad intentions. They arise from frictionless efficiency. A manager receives a plausible-looking shortlist and follows it under time pressure. A recruiter adopts a generated summary because it sounds professional. A team lead uses a score as a starting point and ends up treating it as the conclusion. Formally, there is human oversight. In substance, the decision is already biased. Data and AI provide signals. Humans make decisions.
This also means that AI literacy in HR is something different from a one-time training session on prompts or policy. It is a shared understanding of the limitations of any recommendation. HR, managers, procurement, legal, and where relevant the works council or social partners, must understand what an outcome does and does not convey. Not everyone needs to build the model, but everyone who acts on its output must be able to recognize its uncertainty.
Good governance therefore does not begin with the model, but with the process map. In which steps of recruitment, evaluation, planning, mobility, or absence management can AI guide a decision? Where is speed beneficial, and where does speed destroy the very nuance required? At what moments should deviating from advice not be viewed as inefficiency, but as sound professional judgment? Such choices are matters of organizational design, not technical configuration.
Here lies an uncomfortable truth for executive leadership and boards. A vendor can provide documentation, clarify risk categories, and describe technical measures. What a vendor cannot determine is which evaluation framework is legitimate within your own organization. That responsibility remains internal. What counts as good performance, reasonable comparability, proportionate follow-up, or an acceptable margin of error are not software settings, but governance choices.
For leadership, this demands a shift in focus. Not away from innovation, but away from the assumption that consistency automatically creates quality. In people processes, a completely homogenized decision is rarely a strong decision. Effective leaders actively create space for contestability, for substantiated deviation from a recommendation, and for a second look whenever the context demands it. That may sometimes feel slower, but it elevates the quality of judgment.
This also touches on the broader discussion surrounding HR data. Many organizations want to deploy AI while data definitions, ownership, and provenance remain vague. In such cases, speed is purchased at the cost of false certainty. At the same time, simply connecting more data is not a solution in itself. When weak assumptions circulate through more systems, they only become more scalable. The question is therefore not just whether data is available, but whether the organization understands what that data represents and what it is or is not suitable for.
In the coming years, the most mature organizations will not necessarily be those with the longest AI policies or the highest number of pilots. The difference will show in the quality of their decision-making. Can they explain how a recommendation was used, who was responsible for weighing the decision, and how an employee or candidate can contest an outcome? That is not only relevant for compliance or regulation. It is the cornerstone of internal trust.
In that sense, the AI Act is less a technology assessment than a maturity test for HR and leadership. Those who reduce AI to software management will constantly lag behind. Those who treat AI as an opportunity to organize processes, responsibilities, and decision rights with greater precision gain something far more valuable than speed: not only reduced risk, but also higher quality in how organizations treat people. And that is precisely where the conversation belongs.
How Almuruna One handles AI and the AI Act