An AI chatbot that answers employee questions about leave, pay and policies may be introduced once four things are in place: a clear notice that the answer comes from an AI (Art. 50 of the EU AI Act, in force since 2 August 2026), a permission concept that shows only what the asking person may see, a data protection review including a data processing agreement and, where a works council (Betriebsrat) exists, its consent under section 87(1) no. 6 of the Works Constitution Act (BetrVG). A pure information chatbot is not a high-risk system by the wording of the law, but as soon as it evaluates performance or allocates tasks, it is one. You are liable for wrong answers yourself.
At a glance
- The notice “You are talking to an AI” has been mandatory since 2 August 2026 (Art. 50(1) AI Act), unless it is obvious. By its wording the duty falls on the provider, but set the notice yourself anyway.
- A chatbot that only explains policies is not a high-risk system by the wording of Annex III. If it evaluates performance or allocates tasks based on behaviour, it is.
- The language model itself knows no access rights. The permission concept must sit in the knowledge base and in the connected systems (guidance of the German Data Protection Conference, DSK, on RAG).
- Questions about illness, pregnancy or severe disability are health data. Plan for them before the chatbot goes live, not after.
- Wrong answers are on your account: the Federal Labour Court (BAG) recognises liability for incorrect information, and a Canadian tribunal in 2024 did not accept a chatbot as a legal person of its own.
Monday morning, 7:40: the HR mailbox holds 31 questions. How many days of leave are left, when does the payslip arrive, what applies when someone moves house, how does parental leave work. Three quarters of it is written down somewhere on the intranet, and someone on the team asks whether an AI chatbot could not answer it. Technically that is an afternoon. Legally it is a project involving data protection, the works council and liability, and it falls between the cracks because it is neither pure recruiting AI nor pure shadow AI.
This article is written for HR leads and managing directors who want to introduce an assistant for employees or who already have one in the house as a feature of their HR software. It describes the framework in general terms, names no products and does not replace legal advice in an individual case. Applicants fall under a different framework, which I described in the article on AI in candidate selection, and the same goes for performance reviews.
I work as an interim and fractional HR manager and, as an HR manager, have myself been responsible for mailboxes like the one above. That is a self-interest, and I disclose it: the last section covers when you do not need me for this. I checked the legal statements on 5 November 2026 against statutory text and court and authority websites. The vendor figures come from my own survey and rest solely on public information.
May an AI chatbot answer employee questions, and how common is it?
Yes, and it is more common than most people assume. In the Bitkom Research survey (602 companies with 3 or more employees, calendar weeks 18 to 25 of 2026), 13 percent of companies already use AI to answer internal HR enquiries automatically, and 53 percent can imagine doing so. The law does not forbid the use. It attaches duties to which data the assistant sees, what it does and who can monitor its behaviour.
For context: general AI use stands at 57 percent in the autumn Bitkom survey (603 companies with 20 or more employees, calendar weeks 28 to 33 of 2026), against 36 percent the year before. If your team is planning an assistant, you are not early, but the rules are young. The notice about AI interaction has applied only since 2 August 2026, and many rollout projects from the spring did not plan for it.
Behind the word “chatbot” stand two very different things. One is an information system that answers from policies, works agreements and manuals. The other is an assistant that looks into connected systems, such as the leave account of the person asking, or triggers transactions. This article covers both, because the legal questions shift at exactly that boundary.
For reference
AI in companies and in HR enquiries: how widespread is it?
Two surveys with different populations (3 or more and 20 or more employees). The values are not directly comparable.
Share of companies with 3 or more employees (Bitkom, 2026).
Sources: Bitkom Research, press release of 18 August 2026 (602 companies with 3 or more employees, calendar weeks 18 to 25 of 2026); Bitkom Research, press release of 14 September 2026 (603 companies with 20 or more employees, calendar weeks 28 to 33 of 2026).
Does the chatbot have to say that it is a chatbot?
Yes, unless it is obvious. Art. 50(1) of the AI Act requires AI systems that interact directly with people to be designed so that the people are informed they are talking to an AI. Excluded is what is obvious to a reasonably well-informed, observant and circumspect person. The provision has applied since 2 August 2026. According to an analysis by Cypheron, the Digital Omnibus (Regulation (EU) 2026/1744, published in the Official Journal on 24 July 2026) did not change paragraph 1. A transitional period exists only for the marking of generated content under paragraph 2, until 2 December 2026 and only for legacy systems.
The wording addresses the provider. If you use HR software with a built-in assistant, you are a deployer within the meaning of Art. 3 no. 4, that is, someone who uses a system under their own authority. Still, do not rely on the provider to supply the notice. Whether it is visible is decided in your configuration. Whether a notice inside the application is “obvious” for your workforce is a question of interpretation, and in case of doubt I would not bet on the exemption. One sentence in the chat window costs nothing. Whether your own customisation makes you a provider is something your legal adviser should check before launch; I did not check that in this run.
The fines are high enough not to risk it. Infringements of the transparency duties under Art. 50 fall under Art. 99(4)(g) and can be punished with up to EUR 15 million or, for companies, up to 3 percent of worldwide annual turnover, whichever is higher. For an HR department, however, another point matters more: employees who do not know they are talking to a machine trust the answer like information from the HR department. That is exactly what makes wrong answers expensive.
Click through
The dates for an HR chatbot under the AI Act
Jul 24, 2026
Digital Omnibus published
Regulation (EU) 2026/1744 appears in the Official Journal and amends the AI Act, among other acts.
Sources: Regulation (EU) 2024/1689, Art. 50; Regulation (EU) 2026/1744 according to Cypheron and Cuatrecasas. As of 5 November 2026.
Is an HR chatbot a high-risk system?
A chatbot that explains policies is not, by the wording of the law. Under Annex III no. 4(b), high-risk systems are those that allocate tasks based on individual behaviour or personal traits, or that monitor and evaluate the performance and behaviour of employees. Someone who wants to know how many days of parental leave are possible does not trigger that. This is my reading of the wording, not a statement by an authority: I did not check the Commission guidelines on classification.
The line runs where the assistant passes judgement on people. A chatbot that reports “team risks” to managers, estimates the likelihood of resignations or distributes tasks according to how reliable someone has been is high-risk, as is profiling of natural persons. The exemption in Art. 6(3) for narrow procedural tasks without significant risk does not help then. For such systems the deployer duties under Art. 26 apply, including informing employee representatives and the persons affected before use at the workplace, after the Omnibus from 2 December 2027. The overall framework is described on the page on the EU AI Act in HR.
In practice this means: write down the purpose of the assistant before you switch on functions. Many HR suites ship an assistant that, besides giving information, also answers questions from managers, such as “who in the team has the most sick days”. The second function shifts the classification, the first does not. Choose which one you want and switch the other off. The self-check below leads to a first classification in three questions.
Self-check · interactive
Which tier does your HR assistant fall into?
Three questions for a first classification under Annex III of the AI Act and Art. 22 GDPR. Not legal advice; the assessment in an individual case belongs in the hands of your data protection and legal advisers.
Does the assistant evaluate or rank individual employees, or allocate tasks based on their performance or behaviour?
A “risk traffic light”, “resignation risk” or an absence ranking for managers also counts.
Which data may the assistant see, and what happens with sensitive questions?
Only what the asking person may see themselves, and only for the purpose of giving information. That is the basic rule from Art. 5(1)(b) and (c) GDPR: data is collected for specified purposes, not processed further in a way incompatible with them, and limited to what is necessary. For employees the legal basis is usually section 26(1) of the Federal Data Protection Act (BDSG), that is, necessity for carrying out the employment relationship. Consent is of little use, because its voluntariness is doubtful given the dependence of employees under section 26(2) BDSG.
The permission concept is the most common weak point. In its guidance on retrieval-augmented generation (October 2025), the German Data Protection Conference noted that a rights and roles concept can be applied to the vector database and the reference documents, but that who has access to what cannot be controlled inside the language model itself. In plain terms: if you put the management salary policy into the knowledge base, everyone who uses the chatbot can query it. Separate knowledge bases by audience, and let queries on personal data run only through the rights of the logged-in person. A large vendor documents for its product that the assistant shows only data the person has at least read rights to. That is a vendor statement about product behaviour, not a verification, and you should test the permissions in your own house regardless.
Sensitive questions will come. Employees ask the chatbot because it does not judge, and they type in illnesses, pregnancies or a severe disability. Health data belongs to the special categories whose processing Art. 9(1) GDPR prohibits in principle. In employment it is permitted under section 26(3) BDSG if it is necessary to exercise rights or duties under labour and social law. Consent must refer expressly to this data. The DSK considers Art. 9 data in RAG systems possible under conditions, because it does not remain permanently in the model, provided no targeted retraining with it takes place.
Rules for the design follow. On health questions the chatbot should point to the right contact instead of processing them. Inputs should not be used for training, and the input history should not be stored beyond the session, as the DSK recommends for default settings of employee accounts (guidance on AI of 6 May 2024). Whether your vendor can do that is a question for the contract. How to regulate handling in the house when employees already use their own tools is described in the article on AI use without approval in the team.
The chatbot should not answer these questions itself but hand them to a human:
- Illness, rehabilitation stay, return to workHealth data under Art. 9 GDPR. Refer to the HR department or the company doctor.
- Pregnancy and maternity protectionInformation on rights is possible; the processing of the person's own details belongs in the protected channel.
- Severe disability and equal statusSpecial category. The contact is the representative body for severely disabled employees, where one exists.
- Complaints about managersBelongs in the complaints procedure, not in a chat log.
- Another person's salaryMust never come from the knowledge base. Check permission, and when in doubt refuse.
Who is liable if the chatbot gives wrong information?
You are, as the employer. There is no express statutory rule for chatbot information in the employment relationship. However, in its decision of 18 February 2020 (3 AZR 206/18), the Federal Labour Court (BAG) reaffirmed the principle that an employer who gives information without being obliged to do so must give it correctly, clearly and completely, and is otherwise liable for the damage. In the case decided (company pension), the court found no breach of duty. The statement is a principle, not a ruling on chatbots. That it is transferred to the answers of an AI assistant is my classification and not a decided question of law.
From abroad there is an illustrative example. In Moffatt v. Air Canada (2024 BCCRT 149), the Civil Resolution Tribunal of British Columbia awarded a customer CAD 650.88 difference plus interest and fees after wrong chatbot information on the refund of a bereavement fare, and rejected the argument that the chatbot was a separate legal person. The ruling binds no one in Germany. It shows how a court argues when employees or customers trust an answer that stands on the company's website.
How often such errors occur cannot be proven for HR questions, because I know of no study specifically on HR matters. Two studies from other fields show the order of magnitude. The EBU and BBC examined more than 3,000 answers of AI assistants to news questions and found at least one significant problem in 45 percent. A Stanford study on AI legal research tools found hallucination rates of 17 to 33 percent. Neither field is HR. They do argue against putting an assistant in front of employees without countermeasures.
The countermeasures are unspectacular. The assistant answers only from a reviewed knowledge base, names the source in every answer, says “I don't know” instead of guessing, and hands over to a person on money, deadlines and dismissal questions. A notice in the chat window that answers are non-binding does not replace that. It cannot cancel the workforce's expectation, but it can dampen it.
Take a guess
How often do AI assistants deliver answers with a significant problem?
In what percentage of the AI answers examined did the study find at least one significant problem?
Source: EBU, study on the news integrity of AI assistants, published 22 October 2025 (more than 3,000 answers, 22 broadcasters, 18 countries, 14 languages). News questions, not HR questions.
How do you secure the vendor, the data and the technology?
With a data processing agreement, a check of where the processing takes place and a security concept against data leakage. If a vendor processes personal data of employees for you, Art. 28 GDPR requires sufficient guarantees and a contract that, among other things, provides that the vendor processes only on documented instructions, including for transfers to third countries. The basics on legal basis and data processing are described on the page on GDPR and AI in HR.
If the service or a subprocessor sits outside the European Economic Area, Chapter V GDPR applies in addition. Without an adequacy decision, appropriate safeguards plus enforceable rights and effective legal remedies are needed (Art. 46). For the USA, the adequacy decision on the EU-US Data Privacy Framework of 10 July 2023 currently applies. The General Court of the European Union dismissed the action against it on 3 September 2025 in case T-553/23. According to a specialist publication (EDPL), an appeal to the Court of Justice was lodged on 31 October 2025. I could not check where it stands. Plan the data transfer so that it survives a change in the situation.
On the technical side, the Federal Office for Information Security (BSI), in its study on generative AI models (version 2.0, 17 January 2025), names two risks that matter especially for an HR assistant: the unwanted leakage of inputs and outputs with cloud services, including possible use for training, and indirect prompt injection, where instructions are hidden in documents the model evaluates. The OWASP list of the ten biggest risks for LLM applications 2025 also names both (LLM01 Prompt Injection, LLM02 Sensitive Information Disclosure). For HR that means: anyone who loads documents such as applications, sick notes or emails into the knowledge base opens a door for hidden instructions. Load only reviewed documents of your own.
Art. 25 and 32 GDPR require data minimisation by design and by default and a level of protection appropriate to the risk. A data protection impact assessment is needed under Art. 35 where a high risk is likely. The joint mandatory list of the data protection authorities (as of 17 October 2018) names, among other things, the use of AI to steer interaction with data subjects and the extensive processing of employee data to evaluate work activity. The examples there do not fit a pure information chatbot directly, and the DSK guidance on AI says that an impact assessment will often be necessary for AI applications. So expect to carry one out and document the result. An infringement of the principles of processing can be fined under Art. 83(5) GDPR with up to EUR 20 million or 4 percent of worldwide annual turnover.
What does the works council co-determine with an HR chatbot?
The introduction, insofar as the assistant can make behaviour or performance visible, and that is the case with every chatbot that keeps logs. Section 87(1) no. 6 BetrVG gives the works council a co-determination right on the introduction and use of technical devices designed to monitor the behaviour or performance of employees. According to the Federal Labour Court (13 December 2016, 1 ABR 7/15), the objective suitability to collect and record behaviour or performance information is enough. Stored questions together with a user ID meet that.
Whether the works council also co-determines without logs is less clear. In interim proceedings, the Labour Court of Hamburg on 16 January 2024 (24 BVGa 1/24) denied a co-determination right where employees used ChatGPT voluntarily through the browser and private accounts (according to Gleiss Lutz). It saw no effect on work behaviour and missed monitoring pressure from the employer. For an assistant provided by the employer with login and logs, that does not carry, and the law firm itself warns against generalising. Anyone who relies on it risks an injunction at launch.
On top come the participation rights on AI itself. Where the works council has to assess the introduction or use of AI, calling in an expert is deemed necessary (section 80(3) sentence 2 BetrVG), and it must be informed about the planning of work procedures including the use of AI (section 90(1) no. 3). The works agreement applies directly and mandatorily (section 77(4)), and if no agreement is reached, the conciliation board (Einigungsstelle) decides at the request of either side (section 76(5)). Plan weeks, not days. Templates and structure are described on the page on the works council AI agreement.
The works agreement is not a free pass. On 19 December 2024 (C-65/23), the CJEU ruled that collective agreements on employee data protection must meet the requirements of Art. 5, 6(1) and 9 in addition to Art. 88(2) GDPR. On 8 May 2025 (8 AZR 209/21), the Federal Labour Court awarded an employee EUR 200 under Art. 82 GDPR because in the test operation of HR software more real data went to the parent group than the works agreement allowed. The amount is small. The lesson is not: even a test needs a framework.
Tap to open
Five participation points that come together with an HR chatbot
Sources: BetrVG sections 76, 77, 80, 87, 90; BAG 1 ABR 7/15; Gesetze im Internet. As of 5 November 2026.
What do vendors say about data protection and the AI Act?
Little, as far as it can be checked publicly. I went through 293 HR tools from my survey (core HR systems, payroll, engagement, learning and recruiting, archived entries not counted) to see whether, according to their vendor pages, they offer an AI assistant for employee questions. After manual cleaning, that is 23. For them I looked at what is stated publicly on data processing and the AI Act.
Six of the 23 state something on the data processing agreement: four publish it, one hands it out on request, one requires consent by click. For 17, no information is recorded. On the AI Act, two comment, one with a statement and one saying that, to the best of its knowledge, its AI functions are not high-risk. For 13 I found no statement, and for 8 it had not yet been checked. A page on the topic of the works council exists for one tool. On hosting, the location is recorded for 12: 2 in the EU or Germany, 3 in the USA, 7 in other places or several.
That does not mean the documents are missing. It means you have to request them before you feed in employee data. Ask each vendor in writing: where is the data processed and by whom? Are inputs used for training? How long do logs remain? Can the assistant be limited to information from released documents? Who sees questions with names? The answers belong in the data protection impact assessment and in the talks with the works council. No vendor is rated here, and the survey rests solely on public information.
Own survey
Of 293 HR tools, 23 offer an assistant for employee questions, 2 comment on the AI Act
Source: own survey by Nick Schaefer, as of 5 November 2026 (data per tool last checked between 25 September and 28 October 2026), 293 HR tools (core HR systems, payroll, engagement, learning, recruiting) excluding archived ones, assistants cleaned manually according to vendor pages. Recorded is what stands on the vendor pages. No information does not mean that the function or the document is missing.
Evidence9
What do you settle before the chatbot goes live?
Eight things, in this order: the purpose, the tier, the notice, the knowledge base, the permissions, the vendor, the works council and the exit. The purpose comes first because it determines the tier. “Explain policies” and “inform managers about teams” are legally two projects, even if the same tool can do both. Then the notice under Art. 50 and the knowledge base that a named person maintains. An assistant is only as current as its sources, and liability for wrong information follows from outdated policies just as from invented answers.
Permissions, vendor and works council follow the sections above. The exit means: you can switch the system off without losing documents or information, and logs are deleted after the period expires. Also train the people who maintain the knowledge base and spot-check the answers. The training duty under the AI Act is described in the article on AI literacy training for employees. Tick off in the list what is settled in your case. The status stays saved in your browser, and with the button below you copy it as text.
To take away
Checklist: introducing an HR chatbot for employee questions
Your progress stays saved in this browser.
When do you not need me for this?
When the assistant remains a pure information service and the rules are in place in the house. If it only answers questions from released policies, accesses no personal data and the works council is involved early, a clear AI policy, your data protection officer and a talk with the works council are usually enough. You need no external advice for that.
It is different if the assistant is to access HR data, trigger transactions or deliver evaluations to managers, and nobody in the house can look at Annex III, permissions, works council and data protection impact assessment together. Then I take over, as interim or fractional HR manager, the classification, hold the talk with the works council and build the processes so that in the end a person decides. Legal advice in an individual case stays with a lawyer.
My most important advice applies regardless of whom you bring in: start with the question of which questions your employees actually ask, and then with the smallest assistant that answers them. Everything it can do in addition is additional law that you have to settle.
Evidence7
Sources
- Bitkom Research, press release of 18 August 2026: AI in HR (in German)
- Bitkom Research, press release of 14 September 2026: AI use in companies (in German)
- EU AI Act (Regulation (EU) 2024/1689), Art. 50: Transparency obligations
- EU AI Act (Regulation (EU) 2024/1689), Art. 99: Penalties
- Cypheron, Digital Omnibus 2026: changes to the AI Act
- EU AI Act (Regulation (EU) 2024/1689), Art. 3: Definitions (no. 4 deployer)
- EU AI Act (Regulation (EU) 2024/1689), Annex III no. 4: Employment and workers management
- EU AI Act (Regulation (EU) 2024/1689), Art. 6: Classification as a high-risk system
- EU AI Act (Regulation (EU) 2024/1689), Art. 26: Obligations of deployers of high-risk AI systems
- Cuatrecasas, Digital Omnibus AI: Regulation (EU) 2026/1744 published
- GDPR Art. 22: Automated individual decision-making, including profiling
- GDPR Art. 5: Principles relating to processing of personal data
- GDPR Art. 9: Processing of special categories of personal data
- Federal Data Protection Act (BDSG) section 26: Data processing for employment-related purposes (in German)
- German Data Protection Conference, guidance on retrieval-augmented generation (October 2025, in German)
- German Data Protection Conference, guidance on artificial intelligence and data protection (version 1.0, 6 May 2024, in German)
- Microsoft Learn, data protection in Microsoft 365 Copilot (vendor documentation on product behaviour)
- Federal Labour Court, press release 8/20 on 3 AZR 206/18 of 18 February 2020: employer's duties to advise and inform (in German)
- The Guardian, 16 February 2024: Air Canada must pay after its chatbot gave wrong information (Moffatt v. Air Canada, 2024 BCCRT 149)
- EBU, 22 October 2025: AI assistants' systemic distortion of news, international study by public service broadcasters
- Magesh et al. (2024): Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools
- GDPR Art. 28: Processor
- GDPR Art. 46: Transfers subject to appropriate safeguards
- General Court of the European Union, press release 106/25 of 3 September 2025 on T-553/23 (Latombe v Commission)
- BSI, Generative AI models: opportunities and risks for industry and authorities (version 2.0, 17 January 2025, in German)
- OWASP GenAI Security Project, Top 10 for LLM applications 2025
- GDPR Art. 25: Data protection by design and by default
- GDPR Art. 32: Security of processing
- GDPR Art. 35: Data protection impact assessment
- German Data Protection Conference, list of processing operations under Art. 35(4) GDPR for the non-public sector (as of 17 October 2018, in German)
- GDPR Art. 83: General conditions for imposing administrative fines
- BetrVG section 87: Co-determination rights, para. 1 no. 6 (in German)
- BetrVG section 80: General duties, para. 3 experts and AI (in German)
- BetrVG section 90: Information and consultation rights (in German)
- BetrVG section 77: Implementation of joint decisions, works agreements (in German)
- BetrVG section 76: Conciliation board (in German)
- Federal Labour Court, decision of 13 December 2016, 1 ABR 7/15
- Gleiss Lutz, Labour Court of Hamburg on ChatGPT: no co-determination right of the works council (in German)
- Court of Justice of the European Union, judgment of 19 December 2024, C-65/23 (in German)
- Federal Labour Court, press release 20/25 on 8 AZR 209/21 of 8 May 2025: damages under the GDPR, works agreement on Workday (in German)
Frequently asked questions
Does an HR chatbot have to disclose that it is an AI?
Yes, unless it is obvious. Art. 50(1) of the AI Act requires that people are told they are interacting with an AI. The duty has applied since 2 August 2026 and by its wording falls on the provider. As a deployer, you should still set the notice in the chat window yourself.
Is a chatbot for employee questions a high-risk system?
A chatbot that only explains policies is not, by the wording of Annex III no. 4. If it evaluates the performance or behaviour of individual employees or allocates tasks accordingly, it is high-risk, with deployer duties from 2 December 2027.
Does an HR chatbot need the works council's consent?
As a rule yes, as soon as questions and answers can be logged. Under section 87(1) no. 6 BetrVG, the introduction of technical devices that can monitor behaviour or performance is co-determined. According to the BAG, objective suitability is enough (1 ABR 7/15).
Who is liable if the chatbot gives wrong information?
The employer. The Federal Labour Court has reaffirmed that information given voluntarily must be correct, clear and complete, otherwise the employer is liable for the damage (3 AZR 206/18). Whether that is transferred to chatbots has not been decided. Plan for a source in every answer, spot checks and handover to humans.
May an HR chatbot process health data?
Only under narrow conditions. Health data is a special category under Art. 9 GDPR. In employment, section 26(3) BDSG permits processing if it is necessary for rights and duties under labour and social law. It makes sense to hand such questions to humans.
Classify your HR chatbot
A free first call in which we sort your plan by disclosure duty, classification, data protection and works council. If an information chatbot with a clear policy is enough, I will tell you so there.
Classify your HR chatbot