The question load nobody budgets for
In the HR teams I have led, recurring standard questions took, by my count, around three hours a week per HR headcount. That figure never appears in a workload plan because each question takes two minutes and arrives by chat, mail, or at the desk. Forty questions of two minutes each is a working afternoon, every week, spent restating what the handbook already says. The colleague answering them is usually the same person who is behind on reference letters and has a works council hearing to prepare.
The questions themselves are almost entirely predictable. Holiday entitlement and carry-over, how to report sick and when a doctor's certificate is needed, payroll dates, how the company pension works, how to book a training, where the travel policy is, what the notice period is, how parental leave is applied for. Roughly forty topics cover, in my experience, well over 80 percent of the inbound volume. That is what makes it a routine and not a support function.
Four layers, and the bot is the least important
- Curated knowledge baseA single library of approved, current documents: handbook, policies, process descriptions, the collective agreement summary if one applies. Nothing else. One owner, a change log, and a rule that a document is removed the day it is superseded.
- The botA language model restricted to that library, instructed to answer with the source document and section, and to say plainly when it has no source. No open internet, no improvisation, no memory of previous users.
- Escalation to a personA list of topics the bot recognises and does not answer. It states that it cannot help with this and gives the name and contact of the responsible HR colleague. The handover is logged so that HR can follow up within a working day.
- Monthly reviewHR reads a sample of the month's answers, checks them against the current documents, fixes the sources that produced wrong or vague answers, and reports the numbers to the works council. Thirty to sixty minutes a month.
Companies that buy a bot and skip the first layer get a fluent assistant that quotes the 2023 holiday policy with complete confidence. Companies that build the knowledge base first and put a plain search on top of it already solve half the problem. The bot is worth having because it understands questions asked badly, in German and English, at eleven at night. It is only worth having on top of a clean library.
What the bot may answer, and what it may not
Escalation: the part that protects people
An employee who types that they are being bullied by their manager is not asking a policy question, and a bot that replies with the section number of the code of conduct has failed them. The escalation list is therefore written first, before the knowledge base, and it is deliberately broad. In doubt, hand over. The cost of a false escalation is one extra message to HR. The cost of a missed one is an employee who was told by a machine that their problem was not recognised.
The handover itself has three parts. The bot says clearly that this topic is handled by a person. It gives the name, the channel, and the expected response time. And it logs the handover with a timestamp, not the content, so that HR can check that every escalation was followed up. In companies with a works council, the agreement usually gives the works council the right to see the escalation count, which is a useful signal in itself. A spike in questions about dismissal after a town hall tells you something the survey will not.
Introducing the bot in five steps
Two weeks of logging every inbound question by topic. The result is your version of the standard 40, and the escalation list.
Every document that answers a standard question is checked, updated, and moved into the single library. Superseded versions are deleted. An owner is named.
Use case, library, logging and retention, escalation, review rights. Shorter than for a screening tool because no decisions are made, but still negotiated.
The bot runs in Copilot or a comparable tool for one team. Every answer is read by HR. Sources are fixed. The escalation list is extended.
Sample of answers checked, library updated, numbers to the works council: questions answered, escalations, corrections.
Works council and GDPR
The bot logs which questions were asked and, unless configured otherwise, by whom. That log is a technical device capable of monitoring behaviour under Section 87 (1) no. 6 of the Works Constitution Act (BetrVG), and the works council (Betriebsrat), the elected employee body in German companies, has a consent right before it goes live. The agreement usually fixes that the log is anonymised or pseudonymised, how long it is kept, that no individual usage report goes to managers, and that the works council sees the monthly review. The four-step path to that agreement is under The Works Council AI Agreement, and the three criteria that decide whether co-determination applies under Co-Determination and AI Under German Law.
Under GDPR, the first version of the bot processes very little personal data: the question text and, if logged, a user identifier. That is manageable with an enterprise data processing agreement and a short entry in the record of processing activities. The picture changes the moment the bot is connected to the HR system to look up an individual's holiday balance or salary. That is a second project with access control per user, an impact assessment, and a longer conversation with the data protection officer. I recommend delivering the policy bot first, measuring for three months, and deciding on the integration with numbers. Under the EU AI Act, a bot that interacts with people carries a transparency duty: employees must know they are talking to a system, which is a one-line notice at the start of every conversation. Details under GDPR and AI in HR and The EU AI Act for HR.
HR knowledge bases and self-service processes built in HR Operations roles, including HR admin cut from 30 to 5 hours a week. Practical knowledge, not legal advice. More about me
40 standard questions an HR bot may answer
The question list from my inventories, sorted by topic, each with the document type that should answer it and a note on the German legal background where one exists. Plus the escalation list of topics the bot must hand over, and the one-line transparency notice. The starting point for your own inventory.
Frequently asked questions
Does an HR bot need a works council agreement?
In most cases yes. The bot is used by the whole workforce, it logs who asked what, and that log is a technical device capable of monitoring behaviour under Section 87 (1) no. 6 of the Works Constitution Act. The agreement typically covers the knowledge base, logging, retention, escalation, and the works council's right to see the monthly review.
How do we stop the bot from inventing answers?
By restricting it to the curated knowledge base and instructing it to say when it has no source. A bot that may search the open internet or improvise will eventually tell an employee they have 30 days of holiday when the contract says 28. The monthly review reads the answers it gave and removes or corrects the sources that produced errors.
Which tool do we use?
Usually the one your company already has. Microsoft Copilot with a SharePoint library as the knowledge base is the most common setup in the companies I work with. ChatGPT Enterprise and Claude for Work offer the same pattern with a document project. A dedicated HR bot product is justified when the volume is high enough to need a ticketing integration.
Can the bot see individual employee data, such as my remaining holiday?
It should not in the first version, and often not at all. Individual data means an integration with the HR system, access control per user, and a data protection impact assessment. Start with the bot that explains the holiday policy. The bot that looks up your balance is a second project.
What happens with sensitive questions?
The bot recognises the topic, gives no advice, and hands over to a named person with contact details. Conflicts, health, harassment, dismissal, pay disputes, and anything about a specific colleague are in the escalation list from day one. The handover is logged so that HR can follow up.
How much HR time does it actually save?
In my experience the recurring standard questions take around three hours a week per HR headcount in companies of this size. A bot that answers the standard 40 well takes most of that off the team. What it does not reduce is the sensitive cases, and it should not try.