Building a Knowledge Base for Support and AI: A Practical Guide
Build a knowledge base that agents and AI assistants can use on calls: what to include, how to write answers, structure, ownership and review.
Ask an experienced agent where they find answers and you will often hear "I just know" or "I ask Priya." That is a fragile way to run a contact center. A good knowledge base puts the answers in a place everyone can reach, and it lets new staff, and AI assistants, give the same answer as your most experienced colleague.

This guide covers what to collect, how to write it, how to keep it alive and how to use it with AI. For turning sources into answers inside an AI agent, see Using Knowledge Sources in Call and RAG for voice agents.
Start from questions, not documents
Do not begin by uploading every document you own. Begin with what people ask. Read a sample of recent call notes, tickets and emails, group them by topic and count. The top twenty topics usually cover most volume. Those become your first articles.
A path from questions to articles
- CollectGather real questions from calls, tickets and chats.
- GroupCluster them by topic and note how often each appears.
- ChoosePick the most frequent and the most costly when wrong.
- DraftWrite a short answer with the agents who know it.
- TestAsk someone new to use it on a real question.
The same exercise finds gaps. Calls where the agent hesitated, asked a colleague or gave an inconsistent answer point to missing or unclear articles. Finding knowledge gaps from calls explains the habit.
Writing articles that work under pressure
An agent has seconds, not minutes. An article should be readable at a glance, with the answer at the top and detail below.
- Title it as the question a caller would ask, not as an internal topic name.
- Put the answer in the first two sentences.
- Follow with steps, exceptions and who to ask if the answer does not fit.
- Use short sentences and plain words; avoid internal jargon.
- Include the date it was last checked.
Weak and strong articles
| Weak | Strong | |
|---|---|---|
| Title | Refund policy v3 final | Can a customer get a refund after thirty days? |
| Opening | A history of the policy | Yes, in these cases; otherwise no |
| Layout | One long paragraph | Short answer, steps, exceptions |
| Owner | Unknown | Named, with a review date |
| Language | Internal abbreviations | The words callers use |
Write for the ear as well as the eye if the content will feed a voice agent. Short, complete sentences work for both readers and listeners; tables and long lists do not survive being read aloud.
Structure and findability
Even good articles fail if nobody finds them. Use a small number of categories that match how agents think, such as billing, delivery, accounts and complaints, rather than how the company is organised. Add search terms people actually use, including common misspellings and old product names. Link related articles to each other, so someone reading about refunds can reach cancellations in one step.
Four habits that keep a knowledge base findable
- One topic per articleAvoid mixed articles that answer three questions badly.
- Plain titlesUse the caller's wording, not the internal label.
- SynonymsList other names for the same thing.
- Few categoriesFive clear groups beat twenty vague ones.
Ownership and review
A knowledge base without owners decays quietly. Give each article an owner, usually the person closest to the subject, and a review date. Short cycles for fast-changing topics such as promotions or opening hours, longer ones for stable topics. When a policy changes, update the article the same day and tell the team.
A light governance routine
- Every article has an owner and a last-checked date
- A weekly slot to add or fix articles from the week's calls
- A monthly look at the oldest articles
- A simple way for agents to flag an error in one click
- A rule for retiring outdated articles rather than hiding them
Using the same knowledge base with AI
An AI assistant answers from what you give it, so a clean knowledge base is a strong prompt you can write. Good human-facing articles usually make good AI sources. A few extra points help.
- Be explicit about what the assistant must not say, such as promises about outcomes.
- State what to do when the answer is not in the base: hand over to a person or take a message.
- Keep prices and dates out of long-lived articles, or label them as changing, so the assistant does not repeat stale figures.
- Avoid contradictions between articles; an AI cannot ask which version is current.
- Keep personal customer data out of the base.
Test with real wording. Ask the assistant ten questions in the words callers use, and read the answers critically. Voice AI hallucinations explains why an assistant may sound confident while wrong, and FAQ content for an AI receptionist shows how to write answers for it.
An example article
Here is how a weak article becomes a strong one, using a hypothetical dental clinic. The weak version is titled "Cancellation policy" and opens with three paragraphs of background. The strong version is titled "Can a patient cancel on the day of the appointment?" and begins: "Yes. Tell them the clinic asks for a day's notice, but never refuse a cancellation. Note it in the record and offer a new time." Below that come the exceptions, such as emergencies, and who to ask for anything unusual.
Notice what changed. The title matches the question. The answer comes first. The advice says what to do, not just what the policy is. And it tells the agent what not to do. That is the difference between a document and a tool.
Common mistakes
- Copying every policy document in unedited and calling it a knowledge base.
- Writing for auditors rather than for people in a hurry.
- Letting two articles disagree, then wondering why answers differ.
- Having no process for retiring old content, so the base fills with ghosts.
- Treating the launch as the finish line instead of the start of upkeep.
- Making edits so hard that nobody bothers to fix errors they spot.
The remedy is mostly habit, not technology. A small team that spends thirty minutes a week on the knowledge base will beat a large team that does a big clean-up once a year.
Measuring whether it helps
A knowledge base should change what happens on calls. Look for fewer escalations on topics you documented, shorter time to find answers, fewer inconsistent replies and fewer repeat calls. First call resolution and handle time are natural measures, but read them carefully: faster is not better if the answer is wrong. Pair them with spot checks of real calls.
Doing it in Call
In Call, an AI agent's knowledge base is built from sources: files, website pages, text snippets, Q&A and imported Notion pages. Q&A suits exact answers, while files and web pages carry background. A check-a-question tool shows what the knowledge base finds for a question without a call, and suggestions built from caller questions can point to gaps. Only managers can change sources. Keep one owner for the content, and review it on the same rhythm as your call reviews.
Frequently asked questions
What should go into a knowledge base first?
The questions that come up most often and the ones that are costly when answered wrongly. Start from real calls and tickets, not from a document dump.
How long should an article be?
As short as it can be while still answering the question and giving the exceptions. If an agent cannot find the answer in a few seconds, shorten or restructure it.
Who should write the articles?
The people closest to the subject, with an editor who keeps the style consistent. Agents are often strong writers because they know the real questions.
Can one knowledge base serve both agents and AI?
Yes, if it is accurate, plain and free of contradictions. Add explicit instructions for the AI about what not to say and what to do when the answer is missing.
How often should I review it?
Fast-changing topics monthly or sooner, stable ones less often. The most important rule is to update an article the day the underlying policy changes.
How do I find what is missing?
Look at calls where agents hesitated or escalated, and at questions an AI assistant could not answer. Those are your gaps.