How to Train an AI Chatbot on Your Business Data

"Training" an AI chatbot on your business usually doesn't mean retraining the AI model. It means giving it a knowledge base (your documents, website pages and FAQs) that it searches before answering, so its replies come from your facts instead of general knowledge. This is called retrieval-augmented generation, or RAG. The quality of what you put in decides the quality of what comes out.

I've watched a lot of chatbots give vague or wrong answers, and nearly every time the cause was the knowledge base, not the model. Here's what I've learned about building one that works, using ReplySetter as the example.

How "training" actually works

1. You add content: documents, web pages, FAQs.
2. It's split into chunks and turned into embeddings (numbers that capture meaning), stored in a vector database.
3. When a message comes in, the system searches for the chunks most relevant to the question.
4. The AI writes its answer using those chunks plus its instructions.

No model is retrained. That's good: you can update a price in seconds, and the AI uses the new one immediately.

What to include

The things leads and customers ask about most:

• Services or products: what they are, who they're for, what's included.
• Pricing: prices or ranges, what affects them, payment options.
• Logistics: location, service area, hours, booking process, what to bring.
• Policies: cancellation, refunds, guarantees, warranties.
• Proof: results, case studies, testimonials (with permission).
• Objections: common concerns and your best answers.
• About you: team, experience, credentials.
• The top 20–30 questions people actually ask, with real answers.

The fastest way to find those questions: read your last 50 conversations and write down every question.

How to write documents the AI uses well

This made more difference than anything else in my testing:

• One topic per section, with a clear heading. "Teeth whitening: price" beats a long page that mentions whitening in paragraph 9.
• Write the question and the answer: "How long does a kitchen remodel take? Usually 4–8 weeks, depending on…"
• Be specific. "Most clients see results in 6–8 weeks" beats "results vary".
• Include the "it depends" factors explicitly.
• Avoid contradictions between documents. Old price lists cause wrong answers.
• Keep facts out of the prompt. Put behaviour (how to talk) in the instructions and facts (what's true) in the knowledge base.

Website crawling

Most platforms can crawl your website to build the knowledge base. Useful, with caveats:

• Marketing pages are often vague. Add a dedicated FAQ document as well.
• Exclude pages that confuse things: old blog posts, job listings, legal pages.
• Re-crawl when the site changes.

ReplySetter can crawl a site into a knowledge base, and you can add documents alongside it.

Telling the AI what to do when it doesn't know

Every knowledge base has gaps. Add a clear instruction:

If the answer isn't in the knowledge base, don't guess. Say you'll check with the team and hand the conversation over.

This one line prevents most of the embarrassing answers I've seen. More in how to stop an AI agent from hallucinating.

Testing the knowledge base

1. Ask the 30 most common questions and check each answer.
2. Ask questions in different words: "how much" vs "what's the cost" vs "pricing?"
3. Ask things that aren't covered and check it says so instead of guessing.
4. Look at what it retrieved. If an answer is wrong, check whether the right chunk was found. If not, rewrite the document; if it was, fix the instruction.

ReplySetter's knowledge bases have a search tester, and the reply trace shows which chunks each answer used.

Keeping it up to date

• Assign an owner for the knowledge base.
• Update prices and offers immediately when they change.
• Review weekly: read a sample of conversations and add missing answers.
• Remove outdated documents rather than adding contradicting new ones.

Where to store the knowledge

Most platforms include a built-in store. If you already run a vector database, some platforms connect to it. ReplySetter has a built-in knowledge base on every plan, and on paid plans can also use pgvector, Supabase, Pinecone or Cloudflare Vectorize.