How to train an AI agent on your business knowledge

Why the knowledge base matters more than the AI

The quality of your AI agent is almost entirely determined by the quality of its knowledge base. The underlying language model (GPT-4, Claude, etc.) is powerful — but it only knows what you tell it about your business. If your knowledge base is vague, outdated, or incomplete, your AI agent will give vague, outdated, or incomplete answers.

Think of it this way: the AI is a brilliant employee with perfect memory and instant recall, but they've never worked at your company before. On their first day, they can only answer questions based on what you've told them. The more specific and organized that information is, the better they perform.

Step 1: Gather your business information

Before you upload anything to a knowledge base, collect the information your customers actually ask about. Here's a practical checklist:

Core business information

• Business name and what you do (a clear, specific description)
• Services or products offered (with descriptions and pricing if applicable)
• Business hours and location(s)
• Contact information (phone, email, website)
• Appointment scheduling process (if applicable)

FAQs your customers actually ask

Go through your last 50 customer conversations — emails, phone call notes, chat messages — and list the questions that come up repeatedly. These are the most important entries in your knowledge base. Common ones include:

• "Do you accept [insurance/payment method]?"
• "What are your prices for [service]?"
• "How do I schedule an appointment?"
• "What's your cancellation policy?"
• "Do you offer [specific service]?"

Policies and procedures

• Cancellation and rescheduling policies
• Refund policies
• Privacy policies (in plain language, not just legal text)
• What happens in specific scenarios (e.g., "What if I'm late for my appointment?")

Step 2: Organize information by topic

Don't just dump everything into one big document. Organize your knowledge base into logical topics. This helps the AI find the right information quickly and reduces the chance of giving wrong answers.

A good structure for most businesses:

• About — Business description, mission, team (if relevant)
• Services/Products — Detailed descriptions, pricing, what's included
• Appointments — How to book, reschedule, cancel, what to expect
• Policies — Cancellation, refunds, privacy, terms
• FAQ — Common questions and answers
• Contact — How to reach a human, emergency contacts, escalation paths

Step 3: Write clear, specific entries

The way you write knowledge base entries matters. AI models work best with clear, specific, factual information. Here are examples of good and bad entries:

Bad entry

We offer various dental services at competitive prices.

Why it's bad: Vague, doesn't answer the actual question, includes marketing fluff the AI might repeat awkwardly.

Good entry

We offer the following dental services:
- General checkup: $85 (includes exam, cleaning, X-rays if needed)
- Teeth cleaning: $65
- Filling: $150-$300 depending on size and location
- Root canal: $800-$1,100 depending on tooth
- Extraction: $200-$400 depending on complexity
We accept most major insurance plans. Call (555) 123-4567 to verify your coverage before your appointment.

Why it's good: Specific, answers the actual question, includes prices and details the AI can quote accurately.

Step 4: Upload to your AI platform

If you're using Zepiva, the upload process is straightforward:

1. Log in to your Zepiva dashboard
2. Go to the Knowledge Base section
3. Upload documents (PDF, Word, text files) or paste text directly
4. Organize entries by topic using the built-in structure
5. Review the AI's suggested responses to test questions

The key is to upload the most specific version of your information. Don't upload a 20-page brochure when you could upload a 2-page FAQ that answers the actual questions.

Step 5: Test with real questions

Once your knowledge base is uploaded, test it. Don't just check if it works — check if it works well. Here's a testing process:

1. Ask common questions — Use the actual questions from your customer conversations. Does the AI answer them accurately?
2. Ask edge cases — What about "Do you do emergency appointments on Sundays?" If the answer isn't in the knowledge base, what does the AI say? It should be honest: "I don't have information about Sunday appointments. Let me connect you with someone who can help."
3. Ask ambiguous questions — "How much?" Without context, a good AI asks for clarification: "How much for which service?"
4. Test the tone — Does the AI sound like your business? If you're casual and friendly, the AI shouldn't sound like a corporate robot.

Step 6: Update regularly

A knowledge base isn't a "set it and forget it" thing. Update it when:

• Your prices change
• You add new services or products
• Policies change (cancellation, hours, etc.)
• You notice the AI giving wrong or incomplete answers
• Customers ask new questions that aren't covered

The best practice is to review your conversation logs weekly for the first month, then monthly after that. Look for questions the AI couldn't answer well and add those to the knowledge base.

Common mistakes to avoid

Uploading marketing material instead of factual information

Your website copy is written to sell. Your knowledge base should be written to inform. Strip out the marketing language and focus on facts, numbers, and specific answers.

Being too vague

"We offer competitive pricing" is useless. "$85 for a general checkup" is useful. AI can only work with specific information.

Ignoring what customers actually ask

You might think the most important thing is your company history. But customers ask "How much?" and "When are you open?" more than "When was the business founded?" Build the knowledge base around customer questions, not what you think is important.

Not including escalation paths

Every knowledge base should include: "If the customer asks about [X], transfer them to [person/department]." The AI should know when to stop trying and connect the customer to a human.

The bottom line

Training an AI agent on your business knowledge is not technically difficult — it's mostly about gathering and organizing the information your customers actually need. The time investment upfront pays off in accurate, helpful AI responses that save you time answering the same questions repeatedly.

Start with the 20 most common customer questions. Write clear, specific answers. Test with real scenarios. Update as things change. That's the practical foundation of a useful AI agent.