The knowledge base your AI sales agent needs: 12 facts to write down first

An AI sales agent can only be as accurate as the material it can search. Before you switch it on, write down twelve facts in plain text, put them in a knowledge base, attach it to the persona and test it. Most bad agent replies are not a model problem. They are a missing fact, a stale fact or a fact written in a way nobody could use.

This post lists the twelve facts, how to write each one, and what to leave out. It is written for Sellinger, where a knowledge base is a collection of sources attached to a persona, but the list works for any agent that answers from documents.

What is a knowledge base for an AI sales agent?

It is the set of documents the agent searches when it writes or answers: a price list, product pages, policies, answers to common questions. In Sellinger a collection can hold a file (PDF, TXT or DOCX up to 10 MB), a link, a website that is crawled on a refresh interval, or a text article. The collection is attached to a persona, not directly to a campaign. The agent searches it for what is relevant to the question rather than reading everything each time.

The persona says who the agent is and how it talks. The knowledge base says what is true. Keep the two apart, and keep rules, the things the agent must or must not do, in the rules.

Which twelve facts come first?

Write each as a short text article with a clear title, so a search finds it.

1. What you sell, in two sentences. Product, who uses it, the main outcome. If a person cannot say it in two sentences, the agent will ramble.

2. Who it is for, and who it is not for. The second half matters more. An agent that knows who to disqualify wastes less of everyone's time.

3. Prices you may publish. A table: plan, price, what the price counts. If a price depends on a quote, say "quote only" and let the rules hand it to a person. The agent should not guess a price.

4. What is included and what is not. Most disputes begin with an assumed inclusion. Write the exclusions as plainly as the inclusions.

5. How booking works. Which link, which calendar, how long the call is, who takes it, what happens if the person cannot make it.

6. Your terms. Contract length, how to cancel, refunds, what happens to data. Copy the wording from your terms, do not paraphrase it.

7. Proof you can show. Case studies with the number and the base it was measured on. Link them. If you have few, say so in the article, so the agent does not stretch one. See the template for replying to a request for case studies.

8. The five questions you hear most, with your best answers. Take them from your inbox, not from your imagination. Write the question as the customer words it and the answer in two or three sentences.

9. What to hand to a person. Pricing outside the table, complaints, legal questions, anything about a contract. Write it as a list. Then set the matching needs attention rules.

10. What never to say. Claims you cannot support, names you may not use, promises about dates, comparisons with named competitors. A short list of forbidden phrases is easier to follow than a long essay on tone.

11. How to talk. Formal or casual, first name or surname, language, the length of a message. This usually belongs in the persona, but an example of a good reply in the knowledge base helps.

12. Who you are. Company name, location, who runs it, how to reach a person. When someone asks "who are you?", the agent should answer from a source.

What should you leave out?

• Old decks. A 40 slide deck from last year contradicts this year's prices. Delete it from the collection, not just from the folder.
• Internal notes. Margins, discount ceilings and negotiation plans do not belong where an agent can quote them.
• Anything you would not say aloud to a customer.
• Duplicates. Two documents with two prices produce two answers.

How do you write a fact the agent can use?

Write it the way a customer would ask it, then answer in the first sentence.

• Weak: "Our flexible terms are designed to accommodate a variety of needs."
• Better: "Can I cancel any time? Yes. You can cancel, upgrade or change modules at any time. There are no lock in contracts and no setup fees on self serve plans."

The second version can be searched, quoted and checked. This example is Sellinger's own pricing answer, which is also kept in the fact register so that it cannot drift between pages.

How do you test it?

Do this before the first send and after every change.

1. Ask a question that only your documents answer, for example "what is your cancellation policy?", and read the reply.
2. Ask a question your documents do not answer. The right result is "I do not know, let me get someone", not a guess.
3. Ask for a price outside the table. It should hand over.
4. Ask something you listed under "never say". It should refuse or hand over.
5. Read the first real sends in approval mode, where every message waits for a person, before you let a campaign send on its own.

In Sellinger you can preview the answer in a live chat widget, which is the fastest way to run these checks.

What does this change in practice?

A knowledge base turns "the agent said something odd" into "which fact was missing". That is a better problem, because you can fix it once.

At Hotel Pirin, a related party, the reservations agent answered guests directly. Before a rule to quote as soon as it had dates and party, many replies promised that a colleague would send an offer, and offers took from 3 hours to 7 days. After the rule, the share of replies that carried a price rose from 50% to 92% (33 of 36). The prices came from a source the agent could search, and the rule said when to use them. The groups are small and the comparison is not causal, but the pattern is the one to copy: a fact, plus a rule that says when to use it. See should an AI agent give a price quote or stall?.

How often should you update it?

Quarterly at least, and whenever a price, a term or a product changes. A website source can be refreshed on an interval, so a change on your site reaches the agent without a manual step. Text articles and files need an edit. Put a date in the title of each article, such as "Prices, checked 25 October 2026", so you can see at a glance what is stale.

A one page checklist

• [ ] Offer in two sentences
• [ ] Who it is for and not for
• [ ] Published prices and "quote only" items
• [ ] Inclusions and exclusions
• [ ] Booking steps
• [ ] Terms, copied
• [ ] Proof with its base
• [ ] Top five questions and answers
• [ ] Hand over list
• [ ] Forbidden claims
• [ ] Tone example
• [ ] Who you are

If you can tick all twelve, the agent has what it needs to answer. If you cannot, you have found the gap before a customer did.