How to Train an AI Chatbot on Your Own Business Data (FAQs, Docs, and Policies)

You usually don't "train" a business chatbot at all. In most cases you give a language model a curated set of your own information (FAQs, price lists, policies, product details) and instruct it to answer only from that material. The technique is called retrieval: the bot looks up the relevant passages for each question and writes its answer from them. So accuracy depends far more on how clean and current your content is, and on the rules you set, than on any technical training step.

Do you need to train the model, or just give it your documents?

For almost every small or mid-size business, the answer is to give it your documents. Fine-tuning a model means retraining it on examples, which is slow, costly to repeat and a poor way to store facts that change, like a price or an opening hour. Retrieval works differently: your content sits in a searchable knowledge base, and when a customer asks something, the system pulls the few most relevant passages and hands them to the model along with the question. Change a document and the answers change the same day.

Fine-tuning can still make sense for tone or a very specific output format, but it is not the right tool for facts. If a vendor tells you that you must train a custom model before the bot can know your return policy, ask why retrieval would not do. The OpenAI and Anthropic documentation both describe retrieval-style approaches, giving the model reference material in the prompt, as the standard way to ground answers in your own data. If you want the next level up, where the bot decides what to look up and in what order, we cover it in Agentic RAG: The Evolution from Passive Search to Active Reasoning.

What business information should you collect first?

Start with what customers actually ask, not with everything you have. Open your inbox, WhatsApp history, Instagram DMs and any call notes from the last two or three months, and list the questions that come up repeatedly. For most businesses, a surprisingly small set covers the majority of conversations: hours, location, prices, delivery or booking steps, returns and cancellations, and how to reach a person. Build the first version of the knowledge base around those, and add more later based on what the bot fails to answer.

Gather the sources into one place and note who owns each one. Typical inputs look like this:

• FAQs: your website FAQ page plus the answers you type again and again in chats.
• Prices and services: the current price list, packages, what is included and what costs extra.
• Policies: returns, refunds, cancellations, warranties, delivery zones, payment methods.
• Process documents: how to book, how an order is handled, what the customer must bring or send.
• Product or service descriptions: specs, sizes, ingredients, requirements, limits.
• Escalation info: who handles complaints, urgent cases or custom quotes, and when they are available.

Leave out anything outdated, anything contradictory you have not resolved, and anything sensitive such as internal margins, staff data or private customer records. A bot can only repeat what it has been given, so the sensitive material should never go in.

How do you clean and structure the content so the bot answers correctly?

Rewrite your content as short, self-contained answers, one topic per entry. A 12-page PDF that mixes shipping, returns and warranty forces the retrieval step to guess which paragraph matters, and that is where wrong answers come from. A better format is a question-style heading followed by two to five sentences that fully answer it, including the conditions. For example, "Can I return a sale item?" followed by the exact rule, the time window and how to start the return.

Resolve contradictions before you upload anything. If your website says delivery takes three days and an old PDF says five, the bot will sometimes pick either one, and the customer will screenshot the wrong one. Put dates on documents that change, such as price lists or seasonal hours, and delete the old versions instead of keeping them "just in case". Spell out exceptions in the same entry as the rule, because a bot that finds the rule without the exception will state it as absolute.

Tables are fine for prices, but keep them simple and label every column clearly. Merged cells, images of price lists and scanned PDFs often extract badly, so convert them to plain text first.

What rules should you give the bot for things it doesn't know?

Write explicit instructions for the moments when the knowledge base has no answer, because that is where chatbots invent things. The instruction should say, in plain terms, that the bot answers only from the provided information, says it doesn't have that detail when it can't find it, and offers a handoff instead of guessing. Without this, a model will often produce a plausible answer from general knowledge, which for a business is the worst outcome: a confident wrong price or policy.

Beyond the "I don't know" rule, decide in advance which topics the bot must never answer on its own. Common examples are custom quotes, legal or medical advice, refunds above a certain amount, complaints and anything involving a specific customer's account. For those, the bot should collect the details and pass the conversation to a person. How to design that handoff is covered in How to Automate Customer Support with an AI Chatbot (and When to Hand Off to a Human), and the broader question of how much freedom to give the bot is in Autonomous vs. Semi-Autonomous Agents: Finding the Right Balance.

How do you test accuracy before customers see it?

Build a test set of real questions and check the answers against your own documents. Take 30 to 50 questions from your actual chat history, including the awkward ones: misspellings, two questions in one message, vague wording like "how much is it?", and questions the bot should not be able to answer. For each, write down the correct answer or the correct behaviour (for instance, "hand off to a person"), then run them through the bot and mark every miss.

1. Run the full list and record which answers are right, partly right, wrong or invented.
2. For each failure, find the cause: missing content, contradictory content, a vague entry, or a missing rule.
3. Fix the source content or the instructions, not just the single answer.
4. Run the whole list again, since a fix in one place can change answers elsewhere.
5. Have someone who did not write the content try to break it with unusual questions.

Pay most attention to anything involving money, deadlines and promises, such as prices, refund windows or delivery dates. A small error in tone is cosmetic; a wrong number can cost you a sale or create a dispute. If you want a more systematic way to score answers over time, see Evaluating AI Agents: Metrics and Frameworks Beyond Simple Accuracy.

How do you keep the chatbot's knowledge up to date?

Treat the knowledge base as a living document with an owner, not a one-time upload. Name one person responsible for it, and tie updates to the events that change your information: a price change, a new product, a holiday schedule, a policy revision. The simplest habit is that nobody changes a price on the website or in a quote template without also changing it in the bot's source content on the same day.

Then use the conversations themselves as feedback. Review a sample of chats every week or two in the beginning, and list the questions where the bot handed off or answered badly. Those gaps are your to-do list for new entries. Over time the review can become monthly. It also helps to know how the bot keeps context inside one conversation versus across visits, which is explained in The Role of Memory in Persistent AI Agent Workflows.

If the bot lives on WhatsApp or Instagram, the same knowledge base can serve every channel, so you only maintain one set of answers. For channel-specific setup, see How to Set Up a WhatsApp Chatbot for Your Small Business.

What usually goes wrong?

The most common problems are not technical. Content is out of date, two documents disagree, the instructions allow the bot to improvise, or nobody tests with real questions before launch. A fourth is overloading the knowledge base with every document the company owns; more material is not better, because irrelevant passages compete with the right ones and dilute the answer.

Another is expecting perfection. Even a well-built bot will occasionally misread a question or fail to find an answer that exists. The goal is that errors are rare, low-stakes and caught by good fallback rules, not that they never happen. If you are unsure whether a bot is the right investment at all, AI Chatbot vs. Hiring Help helps frame that decision.

If you would rather not prepare and test this alone, Kuan AI works through it in stages: a diagnosis of what your customers ask, design of the knowledge and rules, implementation with the tools you already use, and then measuring and adjusting. You can reach us through the form on kuan-ai.com and we reply within 24 hours.