AI in HR: Where It Helps, Where It Gets Risky, and What Applies From 2027

Few areas seem as well suited to AI support at first glance as HR work. Job postings, application processes, references, onboarding plans, works agreements, answering recurring employee questions – all of this is text work with a high degree of repetition.

And few areas are as legally sensitive. It involves personal data, often special categories, co-determination rights, the General Equal Treatment Act – and, from December 2027, the highest regulatory tier of the EU AI Act. Starting here without thinking it through creates a problem for you.

The good news: the line between “unproblematic and immediately useful” and “strictly regulated” is fairly clear. It doesn’t run through the technology, but through the question of whether the system influences a decision about a person.

The uncritical area: drafting text and making knowledge accessible

Job postings

Drafting an appealing ad from a task description and requirements profile is a pure text task. It also helps to check for wording that could be unnecessarily off-putting or discriminatory – age references, gender-coded phrasing, or inflated requirement lists, for example. This tends to lower the AGG risk rather than raise it, because a second pair of eyes reviews it.

Responsibility for the content of the ad naturally stays with HR.

A distinction that’s often overlooked: Writing the ad text itself is uncritical. But Annex III, point 4(a) explicitly also names AI systems intended to target the placement of job ads. Algorithmic targeting in ad placement therefore falls into the high-risk category – regardless of who wrote the text. In practice, this is the most common blind spot, because targeting usually sits with the media agency or job platform, and nobody in the company thinks about it. Ask your providers about this before booking the next campaign.

Answering employee questions

An assistant connected to your internal policies that answers questions about vacation entitlement, travel expense rules, parental leave, or training budgets noticeably relieves HR – especially in smaller teams where the same questions keep coming up.

Two conditions: answers must come from the actually current documents and name their source. And there must be a clear note that HR decides on individual cases.

Onboarding and knowledge transfer

Generating a structured onboarding plan from existing materials, or giving new colleagues an assistant that explains internal processes, is uncritical and works quickly. With skilled-labor shortages, time to productivity is a meaningful metric.

References

A tricky classic with a clear solution: the draft may be generated from the performance review and job description, while responsibility for content and assessment stays with the manager. Since references follow their own coded language, review by an experienced person is indispensable anyway.

The critical area: selecting and assessing people

As soon as a system contributes to a decision about a person, the assessment changes fundamentally.

What the EU AI Act says

The AI Act classifies AI systems in the employment context as high-risk applications under Annex III. This includes, in particular, systems used for personnel selection – for example, to filter or assess applications – as well as systems that influence promotions, terminations, task assignment, or performance evaluation.

Following the timeline change via the AI Omnibus, the full requirements for Annex III systems apply from December 2, 2027. These include, among other things, risk management, data quality requirements, technical documentation, logging, human oversight, and a conformity assessment.

Separately, a ban has already been in force since February 2025: emotion recognition in the workplace is prohibited. That rules out software that tries to infer personality traits from facial expression, voice, or speech patterns during an interview.

What the GDPR says

Independent of the AI Act, Article 22 GDPR applies: data subjects have the right not to be subject to a decision based solely on automated processing that produces legal effects concerning them or similarly significantly affects them. An automated rejection in an application process falls under this.

What matters is that human involvement is genuine. Anyone who adopts an AI-generated ranking unchecked and merely signs off on it formally hasn’t made a human decision – they’ve documented an automated one.

What the AGG (German equal treatment law) means

Systems that learn from historical data adopt the patterns in that data. If a particular group was predominantly hired in the past, a system trained on that history will perpetuate that selection – without any discriminatory criterion appearing anywhere. The result can be indirect discrimination, for which the company is liable.

The practical recommendation

For mid-sized companies, the sensible path is: No AI-supported pre-selection or assessment of applicants, unless there’s a solid need for it. At typical applicant volumes, the benefit is small, the effort for legally sound implementation is high, and the liability risk is real.

What is unproblematic, on the other hand: having application documents summarized so the manager gets a quicker overview, without generating an assessment or ranking. Having interview questions suggested from the job description. Drafting rejection letters – after the decision has been made.

Co-determination: involve early, not after the fact

Where a works council exists, introducing systems capable of monitoring behavior or performance is subject to co-determination. What matters is the system’s objective capability, not the intent behind it. An assistant that logs who asked which question and when can already meet that threshold.

In practice, it works well to involve the works council in shaping the approach from the start, rather than asking for approval at the end. The points typically covered in a works agreement are the ones you should clarify anyway: the purpose of use, which data is processed, who may view logs, an explicit exclusion of performance monitoring, retention periods, training, and a complaints procedure.

Data protection: the mandatory pieces

• Legal basis: determine one for each processing activity – in the employment context, typically via necessity for the employment relationship rather than consent, since the voluntariness of consent in an employment relationship is questionable.
• Data processing agreement with the provider, including clarity on storage locations and subprocessors.
• Record of processing activities: update it accordingly.
• Data protection impact assessment: check whether one is required – usually necessary for applicant selection and performance evaluation.
• A deletion concept, especially for applicant data, whose retention periods are limited.
• Transparency toward employees and applicants regarding the use of these systems.

A sensible starting point

Start in the uncritical area. An assistant for employee questions about internal policies can be implemented within a few weeks, provides noticeable relief, and creates none of the risks described above – provided it only accesses general policies, not personnel files.

Once this case is running, trust, a works agreement, and data protection documentation are all in place. On that basis, the next step can be discussed properly – with the knowledge that anything moving toward selection and assessment falls under significantly stricter requirements from December 2027.

This article does not replace legal advice. For the specific design in your case, please consult your data protection officer and, where available, your employment law counsel.

A works agreement: the usual points to cover

Where an employee representation body exists, an agreement is the most practical path – not as a hurdle, but because it clarifies exactly the questions you need to clarify anyway. Proven points to cover:

• Purpose and scope: Which systems, which departments, which tasks.
• Explicit exclusion: No performance and behavior monitoring, no automated assessment of individuals.
• Data types: Which data is processed, and which is explicitly excluded.
• Logs: Purpose, who may access them, retention period, a ban on using them to assess behavior.
• Training: Scope and frequency of training.
• Complaints channel: Who employees can turn to.
• Review: A fixed date on which the agreement is jointly reviewed.

The mistake that happens most often

It’s describing the process as “human-supervised” when it factually isn’t. If a system produces a ranking of applicants and the responsible person adopts that order without assessing the documents themselves, there’s no genuine human decision – regardless of who signs off.

Genuine human involvement requires that the deciding person has both the authority and the actual ability to decide differently, and has the information needed to do so. Anyone who can’t establish these conditions should leave the use case alone.

Frequently asked questions

Are we allowed to have application documents summarized?

A pure summary without assessment or ranking is significantly less critical than pre-selection. Make sure no suitability statement is generated, and that applicants are informed about the use. As always, the processing itself needs a legal basis and a data processing agreement.

What about AI-supported voice or video analysis in interviews?

Emotion recognition in the workplace has been banned since February 2025. That rules out systems that infer personality traits from facial expression, voice, or speech patterns.

Can employees consent to a processing activity?

In principle yes, but in practice consent in an employment relationship is often problematic because its voluntariness can be questioned. Typically, you rely on necessity for the employment relationship instead – discuss this with your data protection officer.

What’s a sensible way to start?

With an assistant for questions about internal policies. No personal data involved beyond the query itself, immediate relief, and it builds the groundwork – agreement, training, documentation – for everything that follows.

Permitted, sensitive, prohibited – an overview

Because the line often blurs in discussion, here are the common applications sorted by risk level.

Unproblematic, ready to use immediately: Drafting job postings and checking them for off-putting or AGG-relevant wording. Deriving interview questions from the job description. Drafting rejection letters after the decision has been made. Answering questions about internal policies. Generating onboarding plans from existing materials. Drafting reference letters based on the manager’s assessment. Making works agreements and policies easier to understand.

Sensitive, only with careful design: Summarizing application documents – permitted as long as no assessment or ranking is created and applicants are informed. Analyzing employee surveys – only in aggregate, never at the individual level. Preparing for personnel meetings – as a structuring aid, not as an assessment of the person.

High-risk under Annex III, full requirements from December 2, 2027: Targeted placement of job ads. Pre-selection or assessment of applicants. Decisions on promotion, transfer, or termination. Task assignment based on behavior or personality traits. Performance and behavior evaluation.

Prohibited since February 2025: Emotion recognition in the workplace. Social scoring of individuals based on their behavior.

For mid-sized companies, this leads to a simple recommendation: stay in the first group. It covers most of the actual relief potential and creates none of the risks mentioned.

Transparency toward applicants

A point that’s easy to meet and easy to forget: if AI is used in the application process – even just for summarization – a notice belongs in the privacy information for applicants.

A single paragraph is enough: which systems are used, for what purpose, that decisions are made by humans, and who applicants can turn to with questions. This openness costs nothing and heads off exactly the discussion that arises when someone else finds out.

Handling applicant data in practice

Application documents regularly contain especially sensitive information – health data in CVs, disability status, occasionally indications of religious or union affiliation. This has two practical consequences.

First, application documents don’t belong in a generally connected knowledge system. They sit in a separate, tightly access-controlled area, and that area is explicitly excluded from an assistant’s scope.

Second, handling them requires a deletion routine. Retention after a process concludes is time-limited; if documents were also processed in an AI system, deletion has to take effect there too – including any caches and logs. Ask your provider about this precisely before you start.