Building a Chatbot That Answers Employee Questions from Your Company Handbook
A step-by-step plan for an internal company chatbot: choosing scope, preparing documents, setting access rights, and handing off to humans when the bot doesn't know.
Every Monday morning, the HR inbox fills up with the same questions: "When do I need a medical certificate for sick leave?", "Does the group insurance cover dental?", "Where is the reimbursement form?" Every answer is already in the employee handbook, but nobody wants to open a hundred-page PDF. That is why a chatbot that answers from the company handbook is the first AI project many organisations choose. This article is a practical plan for building one.
What is a knowledge chatbot?
An internal knowledge chatbot is a chat assistant that answers employee questions based on company documents, such as the employee handbook, IT policies, purchasing procedures or SOPs. Most use the RAG technique behind the scenes: first search for relevant documents, then have the AI compose an answer from them and cite the source.
Step 1: Keep the scope narrow
A common mistake is trying to make the bot answer everything from day one. Start with a single category that has many repeated questions, clear documents and low risk, such as general benefits or basic IT questions like password resets and VPN connections.
Step 2: Get the documents ready
Find document owners
Every document needs someone responsible for keeping it updated. If a policy changes but the file in the library does not, the bot will confidently give outdated information.
Remove duplicates and old versions
Many companies have the 2020 and 2024 editions of the handbook in the same drive. If both go in, the bot may pick the old one.
Write so AI can read it easily
Use clear headings, one topic per heading, and convert tables or scanned files into readable text.
Step 3: Set access rights
This is the part you must not skip. A regular employee should not be able to ask the bot and get someone else's salary or confidential executive documents. The simple rule: the bot must not see more documents than the person asking is allowed to. Where personal data is involved, review it against Thailand's Personal Data Protection Act B.E. 2562 (PDPA) with your legal team or DPO.
Step 4: Design answers and hand-offs
- Every answer should include the document name and section it relied on.
- If nothing is found, say so plainly and give the contact for the person responsible.
- For matters with financial or legal consequences, have the bot point to the document and recommend confirming with staff.
- Put the bot in a channel employees already use, such as the company chat app, so they don't need to open a new system.
A lesson from abroad
The case Moffatt v. Air Canada (2024) at British Columbia's Civil Resolution Tribunal in Canada is often cited. The airline's chatbot gave information about bereavement fare discounts that did not match the actual policy. The airline argued the chatbot was like a separate entity it was not responsible for, but the tribunal rejected that argument and ordered the airline to compensate the customer. The lesson: an organisation is still responsible for what its own bot says. An internal bot does not talk to customers, but if employees act on a wrong answer, it can still cause damage.
Watch out for
- Don't open it to the whole company right away. Pilot with a small group and measure with real questions.
- Be careful sending internal documents to external AI services. Check the data-use terms and privacy settings clearly.
- The bot does not replace HR or IT. It reduces repeated questions so the team has time for matters that genuinely need people.
- Keep logs of questions and answers for review, and tell employees about this transparently.
A good bot isn't one that answers everything. It's one that knows when to hand off to a human.
How to get started
- Pull the most frequently asked questions from HR or IT emails and chats over the past three months.
- Choose one category, and gather the latest documents along with their owners.
- Define access rights and data policy together with the legal team.
- Build a prototype, test it with real questions, and open it to a pilot group.
- Review the questions the bot could not answer every week, then improve the documents or add a new category.
If you handle employee data, also read PDPA and employee data. And learn how to set privacy options when using AI in the course Set AI privacy so work data doesn't leak on Plants.work.