The CEO comes back from a conference and says: "Let's build a chatbot for employees. Let HR and IT stop answering the same questions a hundred times a month." Sounds reasonable. In practice, before it works, there are a few awkward questions to answer.
An internal AI chatbot — answering employee questions about time off, hardware, procedures, benefits, VPN — has been one of the most commonly requested "AI projects" in Polish SMEs over the last year. It has good PR (visible, measurable, sounds modern) and seems simple: there's a knowledge base somewhere, the model reads it, problem solved.
Reality is messier. Not because the model doesn't work — because people don't ask questions the way you think they will, procedures change faster than the knowledge base, and one wrong sentence about compassionate leave can cost the company more than the whole project.
Where a chatbot actually helps
An internal chatbot makes sense where three conditions hold at the same time: there are many questions, the questions repeat, and the answers are unambiguous and rarely change. In Polish SMEs that area usually covers HR basics (how to file leave, when payday is, where to get a tax form), IT self-service (password reset, VPN setup, printer not printing) and health-and-safety/admin (parking, access badges, how to book a ticket).
Anything beyond that is a risk zone. Interpreting the work rules, bonus amounts, an individual employee's situation, anything that depends on tenure, contract or project — that is not a place for a chatbot. One "you are not entitled to extra leave" that turns out to be wrong will generate more complaints than a year without a chatbot.
- Do a month of stats: which questions HR and IT actually receive, and in what proportions.
- Extract the "top 20 questions" — they usually cover 70-80% of the traffic.
- Check how many of them have an unambiguous answer (not "it depends").
- If fewer than 10 of the top 20 pass the unambiguity test, it's not the moment for a chatbot yet.
A concrete example from the field
At one of our clients — a 35-person construction company in southern Poland — HR (2 people) was drowning in repetitive questions from technicians on site. Questions came in on WhatsApp, by phone and by email, mostly between 6:30-8:00 and 15:00-17:00, outside office rhythm. Scale: 60-80 requests a week, of which about 45 concerned 12 topics: leave, business travel, insurance, fuel expenses, accident reporting, clothing allowance and similar.
We deployed a simple chatbot: a language model connected to a 30-page Q&A document written together with HR, available in Microsoft Teams as a bot named "Zosia". Hard rule: the bot answers only those 12 topics; on anything else it says "I'll forward this to HR" and creates a Jira ticket with the person's name and the question.
After three months: 68% of questions closed by the bot without HR involvement, average response time 20 seconds (versus an average of 4 hours before), HR reclaimed about 5 hours per week. Project cost: 38,000 PLN one-off plus about 400 PLN/month for model tokens and hosting. It paid back in about 5 months counting only the two HR people's time.
Important, what the bot did not do: it did not interpret individual contracts, it did not say "you are entitled to", it did not calculate amounts. If someone asked about their specific case, the bot said "that depends on your contract, forwarding to HR" and stopped.
Why most projects fail
Three reasons, in this order.
First: no knowledge base. The company orders a chatbot, meets the vendor, the vendor asks "where is your knowledge base?" and it turns out to be nowhere. Rules are in PDFs on a shared drive, procedures in emails, "how to do X" in the heads of two HR people. The bot has nothing to read. If you write the knowledge base in parallel with the bot rollout, you triple the project length and get the worst version of both.
Second: no content owner. Even if the base exists, someone has to keep it current. Work rules change — who tells the bot? A new benefit appears, an old parking spot goes away, the insurance provider changes. Without one person (even part-time) responsible for knowledge quality, after three months the bot starts lying with full confidence. People will notice and stop using it.
Third: unrealistic scope. The ambition: "the bot will answer everything". The result: the bot answers 40% of questions correctly, 30% with a small error, 30% with full confidence and completely wrong. People lose trust within two weeks. A bot that openly says "I don't know, forwarding to HR" is far better than a bot that guesses.
What to do before you talk to a vendor
Before anyone writes you a chatbot quote, it's worth having four things in hand. They don't need weeks — a week of focused work is enough.
- A "top 20 questions" list with actual stats from the last month or quarter.
- A decision on which of those questions the bot handles and which must be handled by a human.
- A single document (a Word file will do) with the answers to the chosen questions — written in plain language, one style, signed off by HR and a manager.
- One person in the company who takes ownership of updating the content after rollout.
Without those four things, any chatbot rollout will be expensive and frustrating, regardless of vendor or model. With them — the vendor can start immediately and the rollout can close in 4-6 weeks.
The alternative nobody mentions
Sometimes the answer to "too many HR questions" is not a chatbot but a proper internal site with search. SharePoint, Confluence, Notion, whatever. A good full-text search over a 30-page Q&A document solves 60% of the problem at 5% of the chatbot cost. A chatbot makes sense when employees do not want or do not know how to search on their own — and that has to be assessed honestly, not assumed.
The real saving in this area does not come from an AI model. It comes from organizing knowledge so it stops living in two people's heads. The chatbot is only an interface. If a well-described base sits behind it, the model is the cherry on top. If nothing sits behind it, the model will simply lie loudly.
A chatbot that says "I don't know, forwarding to HR" is worth much more than a chatbot that guesses with a smile. People trust the first one. They stop using the second one within a week.