A business chatbot needs a reliable source of answers before it needs a confident tone. Feeding it every document in a shared drive can introduce old prices, conflicting instructions and material that was never intended for customers.
Build a smaller knowledge base around the questions the chatbot is meant to handle.
Collect questions from real work
Review common enquiries with the team. Group them into topics such as service coverage, booking steps, product information and support.
Write an approved answer for each question and identify the source behind it. Include related questions that need different answers rather than forcing them into one broad response.
Our chatbot lead-generation guide describes the wider enquiry journey. The knowledge base supports that journey; it does not replace the person responsible for the answer.
Give each source an owner
Record the document owner, review date and intended audience. Remove superseded versions from the chatbot’s approved collection.
Where information changes frequently, decide whether the chatbot should access an authoritative system or direct the customer to a person. A copied price list is only as current as its maintenance process.
For a hypothetical service provider, an approved coverage list might answer where the team works. A complex request outside that list should trigger a handover, not an improvised exception.
Define what happens without an answer
Write a fallback that explains the limitation and offers a useful next step. Ask only for the information needed to route the question.
Keep sensitive, contractual and unusual requests within the business’s agreed human-review process. Do not let the bot infer a refund decision or promise a service simply because the wording sounds helpful.
Specify which records may be created and who receives the handover. Test that staff can see the question and context.
Test against the source
Use a set of realistic questions, including ambiguous wording and missing details. Compare each response with the approved answer and supporting material.
Record unsupported additions, omitted qualifications and wrong destinations. Update either the content or the workflow, then repeat the affected tests.
Schedule reviews when services change, not only when somebody notices a bad answer. Keep the approved collection small enough that an owner can actually maintain it.
DonutLabs’ chatbot development service can help connect that knowledge with an enquiry process. Bring the recurring questions, current answers and the staff who own exceptions.
For the wider planning context, read How to choose your first business AI workflow.
