Knowledge base for AI agents: how to build one that sells

A knowledge base for an AI agent is the organized set of information the agent looks up to answer customers: products, prices, delivery times, policies and sales rules. It determines the quality of service more than the model you use. Confusing, outdated or incomplete content becomes a wrong answer and a lost sale.

Most companies have this knowledge scattered around: some in PDFs, some in pinned messages in the team's group chat, some in the head of the longest-serving sales rep. Dumping all of it into a tool and expecting it to work is the fastest route to an agent that confidently answers what it shouldn't. This guide shows how to turn that material into a base that's useful for serving customers and, above all, for selling.

What makes a knowledge base for AI different from an internal manual?

A manual is written for a person who has context. They read "the standard delivery time applies" and know what that time is. An AI agent searches for relevant excerpts and builds the answer from them. If the retrieved excerpt doesn't make sense on its own, the answer comes out incomplete or wrong.

That's why every block in the base has to be self-contained. Compare:

The numbers in the table above are illustrative only. The principle is what matters: no hidden references, no dangling pronouns, no dependence on another document.

What needs to go into the knowledge base?

Think in five layers. The first two usually exist; the last three are the ones most often missing and the ones that connect the agent to the sale.

1. Business facts

Who the company is, address, hours, channels, regions served and payment methods. These are simple, frequent questions. Getting them wrong comes across as careless.

2. Products and services

For each item: what it is, who it's for, what it includes, what it doesn't include, authorized price, delivery time and requirements. The "what it doesn't include" field prevents promises that later turn into complaints.

3. Policies

Exchanges, cancellations, warranty, refunds and rescheduling. Write the rule and the exceptions, and state what the agent should do when the case falls outside them.

4. Sales knowledge

This layer is what separates an agent that informs from an agent that helps sell:

• the ten most common objections and the approved answer for each;
• qualification criteria, that is, which questions identify a good customer;
• a comparison between your plans and when to recommend each one;
• arguments by customer type;
• the desired next step: book a meeting, send a proposal or transfer.

Much of this has already been written if the company has a sales playbook. If it doesn't, read closing conversations and extract what the best sales reps do. It's also worth looking at the objection patterns that show up in conversations.

5. Limits and transfer

What the agent can't say and where the conversation goes in each situation. Include discounts, special delivery times, legal topics and serious complaints. The way out to a person deserves its own attention: see how to hand off from AI to a human agent.

How to build the knowledge base in seven steps

1. Start from the questions, not the documents

Export recent conversations and list the real questions, using the customer's own words. Customers say "can I pay in installments?", not "what are the payment terms?". Group them by subject and count the frequency. That list sets the order of the work.

2. Take inventory of your sources

Gather the website, catalogs, price lists, contracts and standard messages. For each source, note the date and the owner. A source with no owner and no date doesn't go in until someone confirms it.

3. Resolve the contradictions

This is the most laborious step and the most valuable. When the website says one thing and the price list says another, the agent may use either. Decide which is the truth and fix the source. It's common for an AI project to reveal that the team had been giving the same customer different answers.

4. Rewrite in short blocks

One subject per block. Use a title phrased as a question, the direct answer first and the details after. A simple template:

Question: Do you serve areas outside the capital? Answer: We serve the cities listed in region X. When to transfer: if the customer is outside the list and interested in an annual contract. Next step: ask for the city and offer to book an appointment.

The "next step" field is what makes the answer move the funnel forward instead of ending the conversation.

5. Add metadata

Tag each block with product, audience, expiration date and owner. A promotional price with no expiration date is the classic source of trouble: the campaign ends and the agent keeps offering it.

6. Test with real questions

Build a test set with the questions gathered in step 1, plus difficult cases. For each one, record the expected answer. Evaluate four points: is it correct, is it complete, does it use the brand's tone and does it lead to the next step?

7. Publish by subject

Release one part of the base, read the conversations and expand. That's worth more than trying to cover everything before launch.

Take Nexo's free assessment and find out whether your operation already has its knowledge organized well enough to put an AI agent live.
Take the free assessment

Practical example: from a price table to a block that sells

Imagine a language school with two plans, regular and intensive. The original material is just a table with names and prices. A customer asks: "Which one is better for me?". With the table, the agent repeats prices. With a well-written base, it has this block:

• Regular: for people with up to three hours a week and a long timeline.
• Intensive: for people with a deadline, such as a trip or a job interview.
• Qualification question: "Do you have a date in mind for when you'll need the language?"
• Next step: offer a trial class and book it on the calendar.

The information is the same a good consultant would use. The difference is that it's been written down. The example is hypothetical, but the method works for any business that has more than one offer.

How do you keep the knowledge base up to date?

The base starts aging on day one. Without a routine, the agent gradually starts making mistakes, and nobody notices until a complaint shows up. Define four things:

Also create a channel for the team to report errors. The people serving customers notice first. And include the base in the marketing calendar: every campaign needs a block with the offer, rules and expiration date before it goes live. If the ad promises something the agent doesn't know about, the money paid for the click is lost in the first answer.

Which signs show the knowledge base is weak?

• The customer asks the same thing again in different words.
• The agent answers "I don't have that information" to frequent questions.
• Sales reps correct the agent after the transfer.
• There are many requests for a human agent on the same subject.
• The answer is correct, but the conversation doesn't move to the next step.

These signs can be measured. See how to measure AI customer service.

What data should stay out?

The knowledge base is no place for customers' personal data, passwords, individual contracts or internal margins. Whatever the agent looks up can end up appearing in an answer. Customer-specific information should come through a controlled integration with the CRM or another system, with defined permissions. For these precautions, see LGPD (Brazil's data protection law) and AI on WhatsApp.

Common mistakes

• Uploading all of the company's PDFs and calling the job done.
• Copying marketing copy, full of adjectives and short on rules.
• Forgetting the expiration date on offers.
• Not writing down what the product doesn't do.
• Leaving the base without an owner after implementation.
• Evaluating the agent only on correct answers, without checking whether the conversation moved forward.

Conclusion

A good knowledge base is the company's sales knowledge written in a way a machine can use. The work of organizing it usually benefits the human team too: answers become consistent, training gets easier and marketing and sales start speaking the same language.

Want to put an AI agent on customer service without risking your sales? Start with Nexo's free assessment.
Take the free assessment

To see the full project, read how to implement AI in customer service and check out AI agents.