Articles  /  Setup

Preparing a Knowledge Base for AI Phone Calls

Setup 7 min read By the TelAI team

A caller asks whether they can return an opened item. The AI agent on the line can only answer correctly if the answer is somewhere in the material it was given, written clearly enough to say out loud. The knowledge base is that material: the product docs, policies and FAQs the agent works from on the call.

Most teams already have the documents. The work is choosing the right ones, fixing the parts that don't hold up when spoken, and keeping them current.

Start from what callers actually ask

The best starting point is not your document library but your phone calls. Ask the people who answer the phones today what they hear most. Look at the subjects of support emails. If you have call recordings or notes, skim a week of them.

You will usually find a short list of topics that covers most calls: order status, returns, hours and locations, pricing, how a product works, what to do when something breaks. Make sure each of those has a clear, current answer in the knowledge base before you worry about anything else.

What to include

  • Policies such as returns, exchanges, warranties, cancellations and shipping.
  • Product details that callers ask about, including sizes, compatibility, care and common problems.
  • Practical facts like hours, locations, holiday closures and how to reach a specific department.
  • Your existing FAQ, after checking that it is still accurate.
  • Names and terms that are specific to your business, such as product names, plan names and place names.

The last item helps more than people expect. Phone audio garbles words, and an agent that knows your product is called "Otto" is much better at telling it apart from "auto" when a caller says it. TelAI reads unclear words against the context of the call and the knowledge base, so having the right names in there gives it something to work with.

What to leave out

More material is not always better. Leave out anything you wouldn't want read to a customer, such as internal notes, margin figures, staff names and escalation gossip. Leave out old versions of policies too. If last year's return policy and this year's are both in the knowledge base, the agent has to guess which one applies, and so does anyone reviewing the call later.

Drafts, meeting notes and long marketing pages rarely help. They add words without adding answers.

Resolve conflicts before the agent finds them

Documents written by different teams at different times tend to disagree. The website says returns within 30 days, the printed receipt says 60 days for members, and the warranty guide says defective items can come back any time in the first year. A person on the phone knows how these fit together. The agent only knows what is written down.

Read through the policies that matter most and make sure they agree, or that the exceptions are spelled out. "Returns are accepted within 30 days. Members have 60 days. Defective items are covered by the one-year warranty instead." is something an agent can say correctly.

Write so it can be said out loud

Written policy language often falls apart when spoken. Long sentences, nested exceptions and references like "see section 4.2" don't work on a phone call. You don't need to rewrite everything, but for the topics callers ask about most, a short plain-language version helps a lot:

  • Put the answer first, then the conditions.
  • Use specific numbers and dates instead of "a reasonable period".
  • Say what the caller should do next, such as bringing the item to a store or emailing a photo.

The agent keeps its answers short and conversational either way, but it can only be as clear as the material behind it.

How much material it can hold

TelAI can load a knowledge base of tens of thousands of words for each call and stay fast from the first reply to the last. That is enough for a full policy manual and a product guide together. The limit that matters in practice is quality, not size. Twenty thousand words of clear, current material will serve callers better than eighty thousand words that contradict each other.

Use a different knowledge base for different calls

Because the knowledge base is loaded per call, you don't need one document that covers everything. A returns line can work from return and warranty policies, while an outbound campaign about delayed orders works from shipping information and the offer you are making. Smaller, focused material is easier to keep accurate and easier to review.

Keep it current

A common reason an agent gives a wrong answer is that the material behind it is out of date. Put someone in charge of the knowledge base, and add it to the checklist whenever a policy, price or product changes.

Your call history tells you where the gaps are. Search transcripts for phrases like "I'm not sure" or "someone will follow up" to find questions the agent couldn't answer, and look at calls with high frustration scores to see what callers were asking about. Each of those is a candidate for a new entry.

Before going live: pick the twenty questions callers ask most, call the agent and ask each one, and check the answers against the people who answer the phones today.

The knowledge base covers what the agent knows. How it behaves is set in its instructions, which we cover in writing a scenario and instructions for an AI phone agent.


TelAI loads your knowledge base for each call and stays fast with tens of thousands of words. See knowledge and instructions

Hear TelAI on your own kind of call

TelAI puts a natural-sounding AI voice agent on your phone line, with live controls for your team and an analysis of each call. You can also see how it works first.

Request a demo