AI in HR delivers the biggest gains in administration and the smallest in assessing people. HR is also the area where the EU AI Act bites hardest: one use is banned outright and almost everything around recruitment and assessment counts as high risk.
In recruitment and selection, AI mainly does the groundwork: job adverts, pre-selection, scheduling and candidate communication. The steadiest gains lie in personnel administration, because changes to records, documents and deadline tracking do not require a judgement about a person.
On top of that sits a legal layer that hardly any article on this topic covers. The ban on emotion recognition has applied since February 2025, the high-risk obligations take effect later than originally planned, and there are three things you must already have in place as an employer today.
Further down you will find which processes are best suited to AI, how to keep a decision about an employee defensible, and in what order to start.
What does AI do in recruitment and selection?
In recruitment and selection, AI mainly does the groundwork around the decision, not the decision itself. Writing and sharpening job adverts, structuring incoming applications, scheduling interviews and keeping candidates informed are the most common uses.
Recruiters also use AI to pre-select CVs on experience, skills and level of education. That is exactly where the legal part of this article begins, because pre-selecting candidates is not an administrative act but an assessment.
Something else is going on too. Candidates now use AI just as heavily as recruiters, so applications have become polished and interchangeable. Selecting on presentation therefore works less and less well, and that is a reason to review your selection criteria rather than automate screening further.
Which tool an HR team uses for this partly determines whether candidates’ personal data ends up in a training set.
What does AI do in personnel administration?
In personnel administration, AI delivers the most gain with the least risk. Changes such as new addresses and contract hours are processed from a single notification, standard documents are generated from file data, and contract and probation deadlines are tracked rather than discovered after the fact.
An internal assistant can also take over part of the flow of questions. Questions about leave balances, the travel expenses scheme or requesting an employer’s statement have an answer that is already in the handbook or the HR system, and that answer can be given straight away.
The difference from the previous section is that nobody is being assessed here. A wrongly generated document is a mistake you can correct; a wrongly rejected candidate is not. If your organisation wants to make a start, this is the place to do it.
Which AI uses are banned in HR?
One AI use affects HR directly through a ban: systems that infer people’s emotions in the workplace. Article 5(1)(f) of the AI Act prohibits placing on the market and using AI systems that infer emotions in the context of work and education. The prohibitions have applied since 2 February 2025.
The exception is narrow: the ban does not apply when the system serves a medical or safety purpose. Analysing a job interview for facial expressions, voice or speech rate to say something about behaviour or motivation does not fall under that exception.
The ban is not a theoretical point. In HR content, including Dutch pages that rank well on this topic, assessing facial expressions and speech patterns is still presented as a benefit of AI. If you buy such a tool because an article recommends it, you are buying a prohibited practice.
Which HR uses count as high risk?
Six HR uses count as high risk under Annex III, point 4 of the AI Act: recruitment and selection, targeted job adverts, evaluating candidates, decisions on promotion, terms of employment or termination, allocating tasks based on individual behaviour or personal traits, and monitoring or evaluating performance and behaviour.
That puts almost everything an HR department might want to do with AI around people in the heaviest category after the ban. The administrative uses from the previous section fall outside it, which makes the dividing line usable in practice.
The supplier does not carry those obligations alone. An employer that chooses, configures or uses a system to support personnel decisions is a deployer and has its own obligations. You cannot hand those over to your software supplier by contract.
There is a lot of confusion about the start date. The high-risk obligations were due to apply from 2 August 2026. With the Digital Omnibus, Regulation (EU) 2026/1744, which entered into force on 27 July 2026, they have moved to 2 December 2027, and to 2 August 2028 for safety components. A delay is not a cancellation, and the preparation time it gives you is exactly the reason to take stock now.
Some high-risk uses also require an assessment of their impact on the rights of the people affected.
What must you already have in place for AI in HR?
Three things already apply to an employer using AI in HR, regardless of the delay for high risk.
The first is the prohibitions. They have applied since 2 February 2025 with no transition period, so a tool that infers emotions in the workplace has to go today.
The second is AI literacy. Article 4 of the AI Act has applied since 2 February 2025 and, since 27 July 2026, under Regulation (EU) 2026/1744 it has been an obligation of effort: you take measures that support AI literacy, without a guaranteed level being required for each employee. The obligation itself has not lapsed. What it covers is explained in the article on AI literacy.
The third is the GDPR. Personnel files contain salary details, sickness reports and appraisals, and every processing operation needs a legal basis and a retention period. Large-scale or systematic processing also requires a data protection impact assessment.
Which HR processes are best suited to AI?
The HR processes best suited to AI are the ones where nobody passes judgement on a person. One question separates the two groups: does an individual employee or candidate bear the consequences of the outcome?
If the answer is no, as with document generation, deadline tracking, leave processing and answering questions about schemes, automation is defensible. If the answer is yes, as with selection, appraisal, contract renewal and dismissal, AI stays limited to preparation and the decision stays with a person.
Expectations about the return need adjusting. The figures circulating in HR articles are adoption figures: how many recruiters use AI and how positive they are about it. Adoption is not impact.
Impact figures look different. In a survey of 25,000 workers in eleven occupations in Denmark, Humlum and Vestergaard report time savings of 2.8 percent of working time, about an hour a week, and that figure is self-reported. In a survey by Yotzov and colleagues, published by the NBER in 2026, 89 percent of executives in four countries say they have seen no effect on labour productivity over the past three years, again self-reported.
That gap between perceived and measured gains is the strongest argument for starting small and measuring. If you skip that, you also get AI use that develops out of the organisation’s sight, known as shadow AI.
For teams that want to get the basics demonstrably in order before tackling processes, AI training for teams is the route that ties in fastest with the literacy obligation.
How do you keep a decision about an employee defensible?
A decision about an employee remains defensible when you can show four things afterwards, even if AI was involved. None of the four calls for expensive software.
The first is who decided. A name, not a system. The second is on what basis: which data was used, which tool, and which question was put to that tool.
The third is that someone else reviewed it. From my years at major banks I bring the four-eyes principle into AI work, and for personnel decisions it is not a luxury but the cheapest insurance there is: one second reader who puts the reasoning next to the outcome.
The fourth is that the person concerned knows. Article 26(7) of the AI Act requires an employer to inform employees and their representatives before a high-risk system is put into use, and Article 86 gives affected persons the right to know the main factors behind a decision. A mention in the privacy notice is the minimum for this, not the end point.
Compliance thinking from the banking sector translates here into three specific questions: which data goes in, which claim do you make based on the outcome, and who is liable if that claim is wrong. The last question requires an awareness that a language model can produce plausible-looking reasoning that has no basis in fact, a phenomenon known as AI hallucination.
How do you get started with AI in HR?
Getting started with AI in HR takes five steps, and the first is taking stock rather than introducing tools. Map which tools are already in use within HR and by whom, because in most organisations the answer is not zero.
Then remove anything that falls under the ban, classify what remains as high risk or not, and start building with the administrative work, where nobody is assessed.
Finally, record what has been agreed and make sure the team knows it. An AI policy nobody has read is not a measure, and the literacy obligation in Article 4 is about exactly that last part.
AI in HR therefore follows the same dividing line this article opened with: administration is fine, judgement about people stays with people, and the regulation draws that line more strictly than most HR content suggests. If you want to set down that line for your organisation in a way of working you can show to a regulator or a works council, the AI Literacy Course covers how to set out such a way of working in an internal AI policy.
Frequently asked questions about AI in HR
If you want to use AI in HR, besides the bans and the high-risk list you will run into questions about the reasoning behind that classification and about penalties.
Why does AI in recruitment and selection count as high risk?
AI in recruitment and selection counts as high risk because, according to the AI Act, such systems can have a significant impact on a person’s career prospects and livelihood. The recitals of the regulation point to the danger of perpetuating historical patterns of discrimination, for example against women, certain age groups or people with disabilities. According to the same recital, systems that monitor performance and behaviour can affect the rights to privacy and data protection.
Is every AI tool in recruitment automatically high risk?
No, not every AI tool in recruitment is automatically high risk, because Article 6(3) of the AI Act provides an exception for systems that pose no significant risk. Examples are a system for a narrow procedural task or for a preparatory task in an assessment. The exception no longer applies as soon as the system profiles people, so a system that builds candidate profiles always counts as high risk.
What fines apply to AI in HR?
AI in HR is subject to the fines in Article 99 of the AI Act, with the highest category for prohibited uses such as emotion recognition in the workplace: up to 35 million euros or 7 percent of worldwide annual turnover, if that is higher. If an employer breaches its obligations as a deployer under Article 26, the maximum is 15 million euros or 3 percent. For SMEs, the lower of the two amounts applies.
