# AI in HR: Practical HR Agents for Mid-Sized Companies

> An HR agent can answer employee questions, draft documents, organise recruiting and track training. Once it ranks candidates or assesses performance, the EU AI Act treats it as high-risk, and the GDPR and, in Germany, works council rights also apply.

Canonical URL: https://www.ruebstudio.de/en/guides/ai-agents-by-department/ai-agents-for-hr  
Author: Dr.-Ing. Marcus Rüb, CEO  
Published: 2026-11-14

HR teams in mid-sized companies are usually small and carry a lot of routine work: questions about leave and working hours, employment letters, job adverts, interview scheduling, training records. This is where AI agents can help. HR is also the area that regulators watch most closely. This guide sets out which HR tasks suit an agent and which rules frame what you can do, with a focus on the EU. The legal notes are for orientation only and are not legal advice.

## What is an HR agent, and what is it used for?

An HR agent is a language model connected to your HR documents and a defined set of tools, which completes HR tasks in several steps. It answers questions, prepares drafts and triggers workflows. It does not make decisions about people; those stay with HR or the line manager.

The difference from a plain chatbot is action. A chatbot returns text. An agent can look up the relevant section of the employee handbook, prepare an employment verification letter from a template, send an interview slot to a candidate, or enrol someone in a mandatory course. Each of those steps is a tool that you explicitly allow.

Common uses in companies with a few hundred to a few thousand staff:

- **Employee self-service:** answers to recurring questions on leave, parental leave, expenses and internal policies, always pointing to the source document.
- **Documents:** drafts of letters, contract amendments, references and offer letters that a person reviews and signs.
- **Recruiting logistics:** writing job adverts, acknowledging applications, scheduling interviews, chasing missing documents.
- **Learning and compliance training:** tracking mandatory courses, matching development offers to roles, sending reminders.
- **Onboarding:** first-day checklists, access requests, naming the right contacts.

For the general architecture of department-level agents and where they pay off, see the overview of [AI agents for business teams](https://www.ruebstudio.de/en/ai-agents).

## How does an HR agent work? The building blocks

An HR agent has four parts: a language model, a knowledge base of your HR documents, approved tools that reach into your systems, and a permissions and logging layer. In HR the last layer matters most, because personnel data is among the most sensitive data a company holds.

**Language model.** It interprets questions, writes answers and plans the next step. Whether you use a cloud model or run one yourself is a real decision in HR: health information, salaries and disciplinary records raise hard questions about where data is processed. Self-hosted options are covered on the page about [sovereign AI](https://www.ruebstudio.de/en/sovereign-ai).

**Knowledge base.** Employee handbook, collective agreements, travel policy, working-time rules. The agent retrieves the relevant passage and cites it. That only works if the documents are current and consistent. How to build such a knowledge base is shown in the [knowledge management agent](https://www.ruebstudio.de/en/ai-agents/knowledge-management).

**Tools.** Read access to master data in the HR system, write access to a drafts folder, access to calendars or the learning platform. Each tool is approved separately and limited to what is needed.

**Permissions and logging.** An employee sees only their own data; a manager sees only their team. Every action is logged. For high-risk systems the AI Act requires deployers to keep automatically generated logs for at least six months ([Article 26 AI Act](https://artificialintelligenceact.eu/article/26/)).

## Use cases: where does high-risk start?

The key question for any HR agent is simple: does it assess people, or influence decisions about them? If not, the requirements are manageable. If it does, it very likely falls under Annex III of the AI Act and becomes a high-risk system with substantial obligations.

[Annex III, point 4 of the AI Act](https://artificialintelligenceact.eu/annex/3/) covers employment in two groups. First, AI systems for recruitment or selection, in particular placing targeted job adverts, analysing and filtering applications, and evaluating candidates. Second, systems intended to make decisions on terms of employment, promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance and behaviour.

That gives a practical split:

| Use case | Typical task | Classification (orientation) |
|---|---|---|
| Self-service | Questions on leave, expenses, policies | not in Annex III |
| Documents | Drafting letters and amendments | not in Annex III if nothing is assessed |
| Recruiting logistics | Scheduling, acknowledgements, completeness checks | borderline, depends on design |
| Candidate screening | Filtering applications, ranking candidates | Annex III, point 4(a) |
| Performance and allocation | Rating performance, assigning work by behaviour | Annex III, point 4(b) |

[Article 6(3)](https://artificialintelligenceact.eu/article/6/) allows an exception: a system listed in Annex III is not high-risk if, for example, it only performs a narrow procedural task or a preparatory task without materially influencing the outcome. The exception never applies where the system profiles natural persons. An agent that builds a suitability profile from CVs therefore cannot rely on it.

### Employee self-service

The lowest-risk starting point. The agent answers only from approved documents and hands personal matters to the right person. It assesses no one and decides nothing. Draw a clear line: questions about illness, conflict or dismissal go to a human rather than being answered by the agent.

### Documents and templates

Employment letters, contract changes, draft references. The agent fills templates with master data and drafts text; a person checks and signs. Take care with references: readers weigh every phrase, and a draft must not quietly contain a judgement that nobody actually made.

### Recruiting support

The line runs through the middle of the process. Writing adverts, acknowledging applications and scheduling interviews is logistics. Ranking applications by fit or scoring candidates is selection, and that is Annex III. A separate article on screening applications with AI, which looks at this boundary in detail, will appear in this section.

### Learning and onboarding

Tracking training records, suggesting courses by role, working through onboarding checklists. This stays low-risk as long as the agent matches offers to roles and mandatory requirements. Once it uses behaviour or performance data to decide who gets which work, it touches Annex III, point 4(b) again.

## Which rules apply? The AI Act, the GDPR and works councils

Three sets of rules interact for HR agents in Europe: the AI Act with its prohibitions, high-risk duties and AI literacy duty; the GDPR with its protection against purely automated decisions; and, in Germany and elsewhere, employee representation rights. None of them bans HR agents. All of them require you to know in advance what the agent does.

### The AI Act: what already applies and what is coming

Some duties already apply. Under [Article 113](https://artificialintelligenceact.eu/article/113/), Chapters I and II have applied since 2 February 2025. That includes the duty in [Article 4](https://artificialintelligenceact.eu/article/4/) to support AI literacy among staff who operate or use AI systems. It also includes the ban in [Article 5(1)(f)](https://artificialintelligenceact.eu/article/5/) on inferring the emotions of people in the workplace, except for medical or safety reasons. An agent that reads candidate or employee sentiment from video interviews or chat messages is therefore not permitted.

Under the text as amended by the Digital Omnibus, the obligations for Annex III high-risk systems apply from 2 December 2027 ([Article 113](https://artificialintelligenceact.eu/article/113/)). As a deployer you will then need, among other things, to use the system according to the provider's instructions, assign human oversight to people with the right competence and authority, monitor operation, keep logs, and inform workers' representatives and affected workers before putting the system into service at the workplace ([Article 26](https://artificialintelligenceact.eu/article/26/)). The full list of deployer duties is covered in [the EU AI Act for mid-sized companies](https://www.ruebstudio.de/en/guides/ai-law-and-compliance/eu-ai-act-obligations-deployers).

The later date is not a reason to wait. A system you introduce today will still be running in 2027. If it is built from the start so that a person decides and everything is logged, you will not need to rebuild it. Patterns for approval and review steps are covered in [human-in-the-loop for AI agents](https://www.ruebstudio.de/en/guides/ai-agent-architecture/human-in-the-loop-ai-agents).

### The GDPR: no automatic decisions about people

[Article 22 GDPR](https://gdpr-info.eu/art-22-gdpr/) gives people the right not to be subject to a decision based solely on automated processing, including profiling, that produces legal effects or similarly significantly affects them. Where such decisions are allowed by contract or explicit consent, the controller must at least offer human intervention, a chance to state one's view, and a way to contest the decision. A rejection email sent by an agent without human review is exactly the case the article targets.

The usual principles also apply: purpose limitation, data minimisation, a lawful basis, retention limits. For systems that systematically assess candidates or staff, a data protection impact assessment is the obvious step, and the AI Act expects deployers to use the provider's information for it ([Article 26](https://artificialintelligenceact.eu/article/26/)). An HR agent can be operated in line with the GDPR when these questions are answered before launch, not afterwards.

### Works councils: information and co-determination

Many European countries give employee representatives a say in workplace technology. Germany's Works Constitution Act is a clear example, and it now names AI explicitly. Under [section 90(1) no. 3](https://dejure.org/gesetze/BetrVG/90.html), the works council must be informed in good time about planned work processes, including the use of artificial intelligence. Where the council has to assess the introduction of AI, bringing in an external expert counts as necessary ([section 80(3)](https://dejure.org/gesetze/BetrVG/80.html)).

The strongest right is co-determination over technical systems designed to monitor employee behaviour or performance ([section 87(1) no. 6](https://dejure.org/gesetze/BetrVG/87.html)). An HR agent that logs users and queries or processes performance data may fall into this category, so have the classification checked by an employment lawyer. In practice, in a German site with a works council, no HR agent should go live without a works agreement. A dedicated article on works councils and co-determination will appear in this section. If you operate outside Germany, check the equivalent consultation duties in each country where you have staff.

## What does an HR agent need to work?

Three things: well-organised documents, an HR system with fine-grained permissions, and an organisation that knows who is accountable for the answers. If one is missing, the project starts with tidying up, which is no bad thing. That order helps even without AI.

**Data and documents.** Check that the handbook, policies and agreements exist in one current version in one place. Conflicting versions produce conflicting answers. Decide which documents the agent may read and which it may not, such as personnel files or medical data.

**Systems.** Many HR platforms offer APIs for master data, absences and documents. Check whether yours supports read access with granular permissions. An agent running on an administrator account is a data protection problem, even if it only answers questions.

**Organisation.** Who maintains the knowledge base? Who reviews drafts? Who picks up a question the agent escalates? Who would provide human oversight under Article 26 if you later run a high-risk system? Name these roles before the agent starts.

**Competence.** The AI literacy duty under Article 4 already applies. Anyone who works with the agent or checks its output should understand how it reaches answers, where it can be wrong and when not to rely on it. A short, documented session for HR and line managers is the practical evidence. A compact [workshop](https://www.ruebstudio.de/en/workshop) is one way to cover it.

## A step-by-step approach

Introduce HR agents from low-risk to demanding: first self-service from fixed sources, then document drafting, then recruiting logistics. Assessment features come last, if at all, and only with employee representation agreed, an impact assessment done and a clear human decision in every case.

1. **Write down the purpose.** One sentence per task: what the agent does and what it explicitly does not do. This list is also the basis for the Annex III check.
2. **Classify each task.** Does it assess people or influence hiring, promotion, termination or task allocation? If yes, treat it as high-risk.
3. **Involve employee representatives early.** Inform them at the start, not just before launch. Draft the agreement together: purpose, data, analyses that are ruled out, log access, retention.
4. **Settle data protection.** Lawful basis, records of processing, a data processing agreement with the model provider, data location, deletion rules, and an impact assessment where needed.
5. **Clean up the knowledge base.** Current versions, clear owners, restricted documents marked.
6. **Pilot with a small group.** One site or department for a few weeks, every answer with a source. Collect wrong answers and fix the cause; it is usually in the documents.
7. **Scale and monitor.** Only extend to more groups and functions once the pilot is stable. Sample answers regularly and review the logs within the limits of your agreement.

In our own operations, starting with an agent that only reads and drafts has worked well. Write actions in other systems are added only once answers have been traceably correct for several weeks.

## Limits and risks of AI in HR

The biggest risks are not technical but in how the agent is used: biased assessments, answers without a basis in policy, unclear accountability and lost trust among staff. An HR agent that employees do not trust will not be used, however good it is technically.

**Bias.** Language models pick up patterns from training data and from your own data. An agent asked to rate applications against past hires will reproduce past patterns. Anti-discrimination law applies to the hiring decision whatever tool prepared it. That is why assessing people does not belong in a first pilot.

**Wrong answers.** An agent can misstate a rule, such as remaining leave or notice periods. Countermeasures: answers only from approved documents, always with a source, and escalation to a person for personal or legally significant questions.

**Monitoring by the back door.** Logs of who asked what and when are valuable for quality and sensitive for trust. Agree in writing who may see logs and what they will not be used for.

**Prohibited features.** Emotion recognition of staff or candidates is banned under Article 5. Check purchased HR software for such features, even when they appear under another name.

**When this approach does not fit.** In very small teams with few questions a week, building an agent rarely pays off. If policies are scattered or out of date, the agent will expose the mess but not fix it. And where staff fundamentally reject AI in HR, an open conversation comes before any pilot.

## How do I get started?

Start with a list of the ten most frequent questions your HR team receives and the documents that answer them. That gives you a self-service agent with no assessment function, which normally sits outside Annex III and is straightforward to agree with employee representatives and your data protection officer.

Three practical first steps:

- **Question log:** for four weeks, record the questions that reach HR by email, phone and ticket. The top ten are your pilot scope.
- **Document check:** for each question, confirm that a current, unambiguous answer exists in a document. Where it does not, fix the document first.
- **Early conversation with employee representatives:** present the plan with its purpose, the data involved and what the agent will explicitly not do.

Once that pilot runs reliably, add document drafting and recruiting logistics. Deeper articles on works councils and on screening applications under the AI Act will follow in this section. For agents in other departments, see the overview of [AI agents by department](https://www.ruebstudio.de/en/ai-agents).

## Frequently asked questions

### Does the EU AI Act classify every HR use of AI as high-risk?

No. Annex III lists specific purposes such as filtering applications, evaluating candidates, deciding on promotion or termination, and monitoring performance. An assistant that explains the holiday policy is not on that list.

### When do the high-risk rules for HR systems start to apply?

Under the text as amended by the Digital Omnibus, obligations for Annex III high-risk systems apply from 2 December 2027. The ban on emotion recognition at work and the AI literacy duty have applied since 2 February 2025.

### What is a works council and why does it matter for HR agents?

In Germany, a works council is an elected employee body with information and co-determination rights. Systems that can monitor employee behaviour or performance generally need its agreement before going live.

### Can an AI agent reject a job applicant on its own?

Article 22 GDPR gives people the right not to be subject to decisions based solely on automated processing that significantly affect them. In practice the agent prepares, and a person decides and can explain the decision.

## Sources

- [EU AI Act, Annex III: High-risk AI systems (point 4, employment)](https://artificialintelligenceact.eu/annex/3/)
- [EU AI Act, Article 6: Classification rules for high-risk AI systems](https://artificialintelligenceact.eu/article/6/)
- [EU AI Act, Article 26: Obligations of deployers](https://artificialintelligenceact.eu/article/26/)
- [EU AI Act, Article 5: Prohibited AI practices](https://artificialintelligenceact.eu/article/5/)
- [EU AI Act, Article 4: AI literacy](https://artificialintelligenceact.eu/article/4/)
- [EU AI Act, Article 113: Entry into force and application](https://artificialintelligenceact.eu/article/113/)
- [GDPR Article 22: Automated individual decision-making](https://gdpr-info.eu/art-22-gdpr/)
- [German Works Constitution Act, section 87: Co-determination](https://dejure.org/gesetze/BetrVG/87.html)
- [German Works Constitution Act, section 90: Information and consultation](https://dejure.org/gesetze/BetrVG/90.html)
- [German Works Constitution Act, section 80: General duties of the works council](https://dejure.org/gesetze/BetrVG/80.html)
