How to prepare your business information for an AI agent that does not make things up

An AI agent does not get better because it has more text. It gets better when it has clear information, defined limits and real examples of how it should respond to customers.

Preparation does not have to be technical. The important part is turning what your team already knows into an organized knowledge base: what you sell, what you can promise, what the agent should not say and what information it should collect before handing a conversation over.

The goal is not to document everything

The common mistake is starting with a huge document and expecting the agent to understand everything on day one. A smaller, reviewable version works better.

To start, Captia needs four blocks:

1. What your business offers: services, products, categories, areas, languages and availability.
2. Which rules it must respect: pricing, bookings, cancellations, guarantees, shipping, returns or limits.
3. Which questions repeat: the questions your team answers every week on WhatsApp, web, email or phone.
4. What should happen next: book, collect details, notify the team, create follow-up or hand off to a person.

If an answer depends on information that is not written down, the agent should not improvise. It should acknowledge the limit and prepare the conversation for the team.

What information to prepare

Think of the agent as a new person joining your team. If that person started tomorrow, what would they need to answer without interrupting everyone every five minutes?

Block	What to include	Useful example
Services or products	What you sell, variations, duration, requirements and availability	”Men’s haircut: 30 min. Color: requires a prior consultation.”
Pricing	Fixed price, starting price, range or when a quote is prepared	”First dental visit: $45. Quote after diagnosis.”
Hours	Real hours, holidays, areas, timelines and capacity	”We work Monday to Friday. Shipping takes 24-72 h depending on stock.”
Rules	What it can confirm and what a person must review	”Can suggest slots. Cannot confirm surgery without reception.”
Tone	How it should sound and which phrases to avoid	”Warm and direct. Avoid jargon unless the customer uses it.”

It does not have to be perfect. It does have to be explicit. “It depends” is not useful unless you explain what it depends on.

Define limits before going live

The most important part is not what the agent knows. It is what the agent is not allowed to invent.

Write down these rules:

• Which prices it cannot give if they are not published.
• Which availability it cannot confirm without a real calendar.
• Which cases must be passed to a person.
• Which legal, medical, technical or sensitive topics require care.
• Which personal details it may ask for and why.

At Captia, we prefer the agent to say “I do not have that information right now” instead of sounding confident with a false answer. That protects the sale and avoids conversations your team has to correct later.

Use real conversation examples

The best examples come from existing messages. You do not need to share the full history: copy typical questions and write the ideal answer.

A good format:

• Customer question.
• Information the agent should check.
• Desired answer.
• Detail it should ask for if something is missing.
• Next action for the team.

Example:

Customer: “Do you have any haircut slots this week?”

Agent: checks services, duration and availability. It replies with concrete slots, asks for the name if the customer chooses one and leaves the booking ready to confirm.

Final review before production

Before publishing, test the agent with real questions, incomplete questions and questions it should not answer.

A good review does not try to make the agent sound clever. It checks that the agent is clear, useful and safe:

• It answers with business information.
• It does not invent prices, timelines or availability.
• It asks for the right detail when context is missing.
• It leaves every conversation with a clear action.
• It knows when to notify the team.

When this is prepared, the agent stops feeling like “a chatbot” and starts behaving like an operational extension of the business.