# AI in HR: Opportunities and Legal Boundaries

Author: Marin Niehues, Niehues Consulting · Published: 2025-07-28 · Updated: 2026-11-16 · AI Adoption & Operations
Source: https://www.niehuesconsulting.com/blog/ai-in-hr/

AI in HR speeds up recruiting and employee services noticeably, but it requires clear legal guardrails on data protection and co-determination. For the executive team, the question is less whether HR uses AI than how the use is organized so that it reaches daily operations and withstands scrutiny. Organizations that rank applications by value and depth of intervention and assign responsibilities early gain speed in the process and reliability towards employees and applicants. This article provides an overview and does not constitute legal advice.

## Where does AI create value in HR today?

AI creates value in HR wherever recurring, text-heavy tasks arise and no one is taking a decision about an individual. Language models deliver measurable effects quickly here, because they shorten cycle times without taking over the decision about a person. The applications can be arranged along the employment lifecycle:

- Attraction: drafts of job adverts, interview scheduling with applicants, summaries of documents for the hiring department.
- Entry: an onboarding assistant that answers new employees' questions from the existing knowledge base.
- Operations: a service desk for standard requests, plus the analysis of free-text answers in employee surveys.
- Development: learning recommendations based on role profiles.
- Planning: scenarios for headcount demand and attrition at the level of business units.

Which of these applications pays off is decided by a business case for each one, and the sequence follows the criteria in the article on [prioritizing AI use cases](/blog/ai-use-case-prioritisation/). HR adds one further variable: the effect on individual people. It determines how much legal and organizational preparation an application needs. The overall route from prototype to operations is set out in the article [From pilot to scale](/blog/from-pilot-to-scale-ai-operations/).

In practice, the tools come from three directions. The large HR suites such as SAP SuccessFactors or Workday have built AI features into the entire employee lifecycle. General assistants such as Microsoft Copilot or enterprise editions of ChatGPT handle text work across all HR processes. Specialist providers cover individual steps such as pre-selection. Whether a tool evaluates people or merely prepares work determines its legal classification.

## Which legal boundaries apply to AI in HR?

Four areas of regulation apply to AI in HR at the same time, from data protection and the AI Act to equal treatment and co-determination.
**Areas of regulation at a glance**
- Data protection: application and personnel data are personal data. Every AI application needs a defined purpose with a robust legal basis, and the individuals concerned are informed. The GDPR also restricts decisions taken solely by automated means that have legal or similarly significant effects on people.
- AI Act, prohibitions: emotion recognition in the workplace is prohibited, except for medical or safety purposes. According to the European Commission's guidelines, this also covers candidates in a recruitment process.
- AI Act, high risk: Annex III lists systems for recruitment and selection as high-risk, as well as systems that influence decisions on promotion or termination, allocate tasks or evaluate performance and behavior. This results in requirements for risk management and documentation as well as for human oversight.
- AI Act, timeline: the obligations apply in stages. The prohibitions and the duty to build AI literacy among staff already apply, while the requirements for high-risk systems follow at a later stage.
- Equal treatment: the German General Equal Treatment Act applies even when an algorithm pre-sorts candidates. Bias in training or reference data can lead to discrimination for which the company is answerable.
- Co-determination: in Germany, the works council has a say as soon as a system is capable of monitoring the behavior or performance of employees.

Responsibility stays with the company, even when an external provider supplies the system. The AI Act distinguishes providers, who develop a system and place it on the market, from deployers, who use it. An employer is normally a deployer. It uses the system in line with its instructions, assigns human oversight to competent people and retains the logs. Before putting a high-risk system into service, it informs the workers' representatives and the employees affected. This makes provider selection and contract design management topics. An overview of the regulation is given in the article on the [EU AI Act](/blog/eu-ai-act/), and the classification in detail in the article on [high-risk AI under the AI Act](/blog/high-risk-ai-under-the-ai-act/).

**The more a system co-determines outcomes for people, the earlier legal counsel and the works council belong at the table. This sequence shortens the introduction, because late objections cost more than early clarification.**

## Which HR applications are prohibited, and which are high-risk?

In HR, the analysis of emotional states is the main prohibited use, and any system that classifies people for personnel decisions is high-risk. Classification follows four questions that you should answer for every planned application before a provider is selected.

**Checklist for classifying an HR application**
- Does the system infer emotions or intentions from voice or facial expression? Then its use in the workplace and in recruitment is prohibited.
- Does the system evaluate or rank applicants and employees, steer the targeted placement of job adverts, or feed into personnel decisions such as promotion and dismissal? Then it is a high-risk system.
- Does the system perform only a narrow procedural task or a preparatory task without profiling individuals? Then the provider may exclude it from the high-risk category but must document that assessment. As soon as profiling takes place, it remains high-risk.
- Can the employer view usage or performance data of individual employees through the system? Then the works council must be involved.

For video analysis in job interviews, this produces a clear line. Systems that infer nervousness or enthusiasm from facial expression or tone of voice are not permitted. What remains permitted is the transcription of an interview recorded with consent, together with a summary of its content for the selection decision, provided data protection and high-risk requirements are met and a human decides.

Dismissals have an equally clear boundary. The decision is taken by a person with authority to represent the employer, it is declared in writing, and any existing works council must be consulted beforehand. AI can organize documents or compile social selection data. Its output is a draft whose derivation the decision-maker understands and reviews.

## A decision framework: depth of intervention and sequence

A robust ordering criterion for HR applications is the depth of intervention: how strongly does the output of the system influence a decision about a specific person? This produces an introduction in three waves.

- Wave 1, assistance and service: applications that do not touch decisions about individuals, such as interview scheduling or a service desk for policy questions.
- Wave 2, preparation: the AI summarizes and structures, the human decides in a documented and traceable way. This is where the most careful demarcation is needed, because the line to profiling is quickly crossed.
- Wave 3, evaluation: systems that classify or rank people. They follow only once legal review and a works agreement are complete and quality has been measured.

A company with its own works council and more than one planned application negotiates a framework agreement before the first pilot, setting out purpose and analysis limits and adding later applications by annex. For a single Wave 1 application, a targeted individual arrangement is sufficient.

**Introduction routes compared**
Criterion | Broad rollout | Individual pilots | With Niehues Consulting
Legal clarification | Retrospective | New for each pilot | Once up front as a framework for all applications
AI Act classification | Unchecked | Inconsistent across pilots | Checklist for each application before provider selection
Works council | High risk of conflict | Repeated negotiation | Framework agreement before the first pilot
Evidence of value | Hard to attribute | Different for each pilot | Uniform metrics for each application
Handover to operations | Ownership unclear | Often open | Ownership and enablement before launch

## Co-determination and employees are part of delivery

In companies with a works council, an AI application subject to co-determination is delivered only once the works agreement is in place and employees use it. In practice, the agreement determines the timeline. Co-determination over technical systems applies as soon as the employer can obtain usage or performance data through the system, and that is the case for every enterprise license with logging. Where employees use a tool only through private accounts without the employer having access to data, this condition falls away, as the Hamburg Labor Court ruled for ChatGPT in 2024.
Companies that approach the works council only after the pilot lose months.

Employees who want to know which data an assistant processes and who sees the results receive a clear answer before go-live. The introduction also includes a written rule on which decisions a human always takes, and a point of contact that receives and analyses error cases. You meet the AI Act's AI literacy duty most effectively with training built around the specific tool and your own workflow.

## A partner for HR programs: Niehues Consulting

[Niehues Consulting](/) supports AI initiatives in HR from the business case into regular operations. In AI management, we clarify requirements, calculate the value of each application and set the sequence, including the allocation to the waves described above. Prototypes prove the value in the real HR workflow, and the transfer into operations takes place with named responsibilities and enabled users.

The approach follows the Niehues Delivery Model with its phases of understanding, building and introducing. Three of its review questions matter most for HR programs. Who decides, who delivers, who uses, and where is the human who takes the decision about a person? Which roles, decision paths and interfaces does operations need so that oversight and error handling work? Are users enabled and processes adapted before the system goes live? The entry point is a free initial consultation of 30 minutes in which the target picture and first applications are clarified: [book an initial consultation](/#kontakt).

The legal assessment in individual cases is carried out by specialist lawyers and data protection officers. Niehues Consulting ensures that their findings arrive as requirements in workflows and roles.

**AI in HR: the sequence decides**
Start with applications that involve no decision about individuals and clarify law and co-determination in parallel. Evaluating systems follow once the framework agreement is in place and quality has been measured. This is how AI reaches everyday HR work and withstands any review.

**Sources**
- European Union: Regulation (EU) 2024/1689 (AI Act), regulation, 2024
- European Commission: Guidelines on prohibited AI practices, guidelines, 2025
- European Union: Regulation (EU) 2016/679 (GDPR), regulation, 2016
- Germany: Betriebsverfassungsgesetz, § 87, statute, 1972
- Arbeitsgericht Hamburg: 24 BVGa 1/24, court decision, 2024

## Frequently asked questions

### Can AI screen job applications?

AI may screen job applications if there is a legal basis, applicants are informed, the system has been checked for bias and a human takes the final decision. Systems that filter or evaluate applications are classed as high-risk under the AI Act. Assessing the specific application is a matter for legal advice.

### Does the works council have to approve AI in HR?

In Germany, the works council has co-determination rights as soon as the employer can obtain usage or performance data on individual employees through the system, which applies to every enterprise license with logging. Without agreement, the system cannot be introduced, and in a dispute the conciliation committee decides. One exception is use through private accounts without employer access to data, which the Hamburg Labor Court classed as free of co-determination in 2024.

### When do the AI Act obligations apply to HR software?

The AI Act's obligations apply in stages: the ban on emotion recognition in the workplace and the AI literacy duty already apply, while the requirements for high-risk HR systems follow later. Plan new recruitment and evaluation systems so that they meet high-risk requirements from the point of procurement. An applicant tracking system usually runs longer than any transition period.

### Is AI video interview analysis legal?

Video analysis that infers emotions or personality traits from facial expression or voice is prohibited under the AI Act. Transcribing an interview recorded with consent and summarizing its content remain permitted, provided data protection and high-risk requirements are met. The selection decision is taken by a human.

### Can AI decide on a dismissal?

AI may not decide on a dismissal. The GDPR prohibits decisions with legal effects taken solely by automated means, and in Germany the dismissal is declared in writing by an authorized person after consulting any existing works council. AI may organize documents and compile data, while the weighing of interests stays with a human.

### What are examples of AI tools for HR?

AI tools for HR come from three groups: HR suites with built-in AI features such as SAP SuccessFactors or Workday, general assistants such as Microsoft Copilot, and specialist providers for individual steps such as pre-selection. What determines effort and obligations is whether a tool evaluates people or only prepares work. Make this classification before comparing providers.

### Can language models be used in HR in compliance with the GDPR?

Language models can be used in HR in compliance with the GDPR when purpose and legal basis are settled and a data processing agreement with the provider is in place. Personnel data is among the most sensitive data sets, so companies define in advance which data an assistant may see. Applications with a high risk to employees additionally require a data protection impact assessment.

### Which AI applications in HR carry the lowest risk?

Applications that involve no decision about individuals carry the lowest risk, such as advert drafts or interview scheduling with applicants. They deliver value quickly and suit the first wave of introduction. They can still be subject to co-determination as soon as usage data is logged.
