The question of how to teach a chatbot to answer customer questions almost always comes from the human side, not the technical one. An owner notices that a new hire is answering customers almost like a seasoned manager within a week — because there are people around, there are sample dialogues, there are unwritten rules. But a bot that was “just plugged in” answers politely and emptily: it confuses delivery terms, promises things that don’t exist, and meets any hard question with a generic “please contact a manager.” The difference isn’t the bot itself. The difference is that the manager was given context, and the bot wasn’t.
A good assistant isn’t model magic. It’s the quality of what you put into it: your customers’ typical questions and your business’s clear rules. This article is about that process — what to gather, and in what order, so the bot speaks in your voice and doesn’t improvise on the fly.
Why a Bot “Makes Things Up” With No Knowledge Base
A language model is built to always produce an answer — even when it doesn’t know it. That’s not a bug, it’s its nature: it constructs plausible text. If you never told it that the minimum wholesale order starts at a certain amount, the bot won’t stay silent. It will output something that sounds logical, and the customer will take it as fact.
The consequence is predictable. The customer gets a confident but wrong answer, brings it to you — and now you’re not answering a question, you’re managing a conflict and rescuing your reputation. A single case like that costs more than all the work of building a knowledge base.
So the first principle is simple: a bot needs facts and boundaries, not a “personality.” Personality is the nice wrapper. Accuracy depends on how fully you’ve described your business in its own words.
Step One: Collect Real Questions, Not Imagined Ones
The biggest mistake is sitting down and inventing questions “off the top of your head.” In an owner’s mind lives an ideal customer who asks to the point. A real customer asks differently, in different words, and often about what seems obvious to you.
You already have the source of the right questions. It’s your managers’ chats, messenger threads, comments under posts, calls that someone takes notes on. Pull up the dialogues from the last few months and copy out the live wording — verbatim, with the typos and slang customers use. That’s exactly how people will write to the bot.
Then group them. You’ll quickly see that 80% of questions are 15–20 recurring topics: price and what it depends on, availability, delivery times and cost, payment, warranty and returns, whether the product fits a specific need. That short list of recurring topics is the backbone of your base. Learning how to teach a chatbot to answer questions starts with this work, not with settings.
Step Two: Describe Business Rules Unambiguously
Questions are half of it. The other half is your rules: what the bot is allowed to say and what it isn’t, and by what logic.
One rule applies here: if a condition can be read two ways, the bot will read it both ways and pick the wrong moment for each. “Free delivery over a certain amount” — over what amount, under what conditions, does it include installation, does it apply to every region? What seems self-evident to you doesn’t exist for the bot until it’s written down.
Describe the boundaries explicitly. Not just what’s allowed, but what to do when a question falls outside them: a custom price, a non-standard order, a complaint, a legal matter. The right behavior here isn’t to invent an answer, but to honestly say “a manager will calculate that” and hand the conversation to a human. A bot that knows the limits of its knowledge earns more trust than one that answers everything confidently.
Separately, record what must never be said: promises about timelines you don’t control, discounts that don’t exist, opinions about competitors. These prohibitions matter just as much as the answers themselves.
Step Three: Set the Voice and Tone
Once facts and boundaries are in place, intonation remains. Your manager speaks a certain way — calm or lively, formally or plainly, in short phrases or at length. The bot has to sound the same, or the customer will sense the substitution even when the answers are correct.
The simplest way to set tone isn’t to describe it with adjectives (“be friendly and professional” means almost nothing) but to provide 5–10 sample dialogues: a customer’s question and the answer your best manager would give. A bot learns from examples far more precisely than from instructions. At the same time you’re fixing the answer length, the level of formality, and how gently to move toward a sale.
Step Four: Test Like a Demanding Customer
A knowledge base assembled at a desk always has gaps. You can only find them in the field — by hitting the bot with the questions real people ask, especially the awkward ones.
Go through it yourself: try to confuse it, ask ambiguous questions, ask about something you deliberately left out. Every answer that disappoints you isn’t a reason to swap the model — it’s a signal that a fact or rule is missing from the base. You add it, and the bot gets sharper. This isn’t a one-time act: the business changes, new products and terms appear, and the base needs topping up the same way you’d keep training a live employee.
This is where experience and your own support infrastructure matter. We at lpf.com.ua have worked since 2018, and we’ve seen that the difference between an empty bot and a helpful one comes not from a flashy technology, but from disciplined work on the knowledge base and honestly drawn boundaries.
Where to Start
Don’t try to describe the whole business at once. Take the 20 most frequent questions and clear rules for them — that’s enough for the bot to handle most inquiries from day one and honestly pass the rest to a manager.
If you’d like, we can start with a short conversation: we’ll look at which questions come to you most often, and how to build a base so the bot truly speaks in your business’s voice instead of improvising on the fly.