Answers for employees, while decisions about people stay with people

An HR assistant takes routine questions off HR's desk. Employees ask how to apply for parental leave, what the travel expense rules say, or which works agreement covers mobile work. Each answer costs HR time, and the same question comes back again and again. An assistant answers from the approved documents whenever the question is asked, in the same words for every site, and hands individual cases to HR.

There is room to grow. In Bitkom's survey of 852 companies, published in January 2025, 9% already used a chatbot for internal HR questions and 49% could imagine doing so. AI for onboarding was used by 11%. In the Randstad-ifo survey for the third quarter of 2025, about 64% of HR managers rated AI's current benefit in HR as low or rather low, and 44% saw the most benefit in contract management and document creation. Bitkom's chief executive Bernhard Rohleder summed it up in 2025: “KI-Tools können Personalabteilungen von vielen Aufgaben entlasten”, AI tools can relieve HR departments of many tasks.

• Fewer repeated questions by mail and phone to HR, counted in your HR mailbox before and after.
• The same answer to the same question at every site, with the paragraph it comes from.
• New colleagues find onboarding information without asking around.
• HR documents drafted from your templates, so HR checks and signs instead of writing from scratch.

The assistant answers questions, guides new colleagues and drafts documents. It never judges a person. The version for employees has no access to personnel files: a question such as how many days of holiday I have left belongs in your HR system's self-service, and the assistant links there. We build it for the following tasks.

• HR policy questions on holidays, sick notes, travel expenses, company cars, mobile work and benefits, answered from policies, works agreements and collective agreements, with the paragraph cited and the version for the employee's site.
• An onboarding assistant: who to contact, which trainings are required, how to request IT access and where the forms are, built from your onboarding guides.
• HR administrative documents: confirmations of employment, standard letters and contract amendments drafted from your templates and the data in the HR system, for HR to check and sign.
• Employment references worded from the manager's own assessment. The manager does the assessing. The system only phrases what it is given and never rates anyone.
• Search for HR staff across policies, works agreements, collective agreements and process descriptions, so HR business partners answer consistently.

Every decision about a person stays with a person: hiring, pay, promotion, warnings, transfers and termination. The assistant does not prepare, rank or recommend any of them.

• Individual cases go to HR: as soon as a question concerns a specific person, contract or salary, the assistant hands it over.
• HR owns the content: only documents HR has approved are used, and a changed policy replaces the old version.
• Documents are drafts until HR checks and signs them. Nothing is sent to an employee automatically.
• Changes to the assistant's scope, sources or logging are agreed with the works council before they go live.

We do not build systems that screen or filter CVs, rank or preselect applicants, evaluate performance, or prepare decisions on promotion or termination. Under Annex III point 4 of the AI Act these are high-risk uses, and after the Digital Omnibus the obligations for them apply from 2 December 2027, as the law firm Lewis Silkin summarised in July 2026. Emotion recognition at work is prohibited (Grant Thornton, 2026).

For the company using such a system, the deployer in the AI Act's terms, that means the high-risk duties, including human oversight, from that date. Other laws apply as well. Under Art. 22 GDPR, rejections may not be fully automated without real human review. Selection guidelines need the works council's consent under § 95 BetrVG, and § 95 (2a) makes clear that this also applies when AI is used to draw them up. § 87 (1) no. 6 BetrVG applies to any system capable of monitoring performance or behaviour. And under the AGG, discrimination by a system counts as discrimination by the employer.

Companies do use AI there. In the Randstad-ifo HR survey for the third quarter of 2025, 33% of companies using AI in recruiting used it for CV analysis and 31% for preselection. We do not build these systems, and we say so before anyone asks. If you are assessing such a product, your data protection officer, works council and lawyers are the right first stop.

The assistant runs on your infrastructure and reads only the HR document set you approve.

• Sources: the approved HR document set on SharePoint or a file share, versioned, with each document's validity and the sites or employee groups it applies to.
• Answers with citations through RAG, which means answers drawn from your own documents with the source shown. When the documents do not answer the question, the assistant says so and offers a handover to HR.
• Access through Teams or the intranet with single sign-on, so each employee sees the rules for their own site and group, and no one else's.
• HR systems such as SAP SuccessFactors, Personio or Workday are connected only for document drafting, read-only and only for HR staff.
• Hosting on your own servers or in your own cloud account. Employee questions and HR documents do not leave your environment.
• Build or buy: HR suites and Microsoft 365 Copilot bring their own AI features. If they answer only from your approved documents and their logging fits your works agreement, use them. A custom build is worth it when answers must come strictly from an approved set with citations, when the logging has to follow a works agreement in detail, or when the data must stay on your own servers.

HR collects real questions from its mailbox, anonymised, and writes down the correct answer and source for each. That set is the acceptance test, and HR signs off the result.

After go-live the test runs again whenever a policy or works agreement changes, and after every change of model or prompt. We measure the effect on HR in totals, such as the questions reaching the HR mailbox before and after, never per employee.

• Correct answer with the correct citation, from the current version of the document.
• Correct handover when a question concerns an individual case or has no answer in the documents.
• No answers from another site's rules, and no content from documents the employee may not see.
• Questions that try to get the assistant to judge a colleague or an applicant are refused.

An HR assistant that only answers policy questions does not match the high-risk examples in the AI Act sources we rely on. It still involves the works council if its use can be tied to individuals: according to Bitkom's 2026 guide on AI and co-determination, § 87 (1) no. 6 BetrVG applies when a system is objectively capable of monitoring performance or behaviour, whatever the intent. In a 2024 case about ChatGPT used in the browser, the Hamburg labour court found (24 BVGa 1/24) that this right did not apply where the employer had no access to employee data through the AI. This is our reading of the sources and not legal advice. You decide with your works council, data protection officer and lawyers.

Employee data is protected under the GDPR and German data protection law, and employees must be told that they are talking to an AI system. That is the transparency duty under Article 50 of the AI Act, which has applied since 2 August 2026. We design the logging around the works agreement, which according to Bitkom's guide should say who may access which data and “wie Zugriffe protokolliert werden”, how access is logged.

• No names in the usage logs: pseudonymised IDs, or none where the agreement says so.
• No reports per person, and no way to look up who asked what.
• Retention set in the works agreement, with deletion that runs automatically.
• Access to logs only for named administrators, for troubleshooting, and every access is itself logged.
• A short system profile for the works council: purpose, sources, what is logged, who sees it, and the test results.

Start with the questions HR answers most often, and bring the works council in at the beginning, well before rollout.

• List the recurring questions from the HR mailbox and the documents that answer them.
• Check that those documents are current and approved. Outdated policies produce outdated answers.
• Agree scope, logging and retention with the works council and your data protection officer.
• Run the acceptance test with HR, then open the assistant for one site or one group first.