Building a knowledge base for AI customer service in 1 hour
Published on November 20, 2026

Imagine you hire a new part-timer for support tomorrow. She is friendly, writes without typos and has worked in three other stores. What she does not know: how long your return window is, whether you ship to Austria, what happens when a customer wants to cancel an hour after ordering, and whether a broken glass needs a photo or just a replacement. Without those answers she either writes filler or asks you on every second email.
An AI agent is in exactly the same position. It can write, it can look up orders in Shopify, it can pull the shipment status. But it does not know your rules until you have written them down. Most stores that fail with AI in support do not fail because of the model. They fail because of a knowledge base that is either missing or consists of 40 pages of terms and conditions in which the one line that matters gets lost.
This article shows what belongs in the knowledge base, what does not, and the order in which to build it in roughly an hour. At the end you have a document that helps not only an AI agent but every temp you ever hire.
What a knowledge base is, and what it is not
A support knowledge base is not an FAQ page and not a copy of your terms. It is the collection of decisions you make in support every day, written down so that someone else would make them the same way.
The difference matters. Your terms say: “The withdrawal period is 14 days.” Your knowledge base says: “We accept returns up to 30 days after delivery. Up to day 45 as a goodwill gesture if the item is unused. After that, no, politely, with a note about the deadline.” One is law, the other is your practice. The agent needs the practice.
Three things do not belong in it:
- Product descriptions. They live in the store, and a good agent can read them there. You do not need to retype that the T-shirt is organic cotton.
- Order data. That comes from Shopify, live. An agent that is supposed to learn orders from a document is built wrong.
- Anything you “might need someday”. The knowledge base grows with the cases that actually arrive. Start small.
The 20 questions that come in every day
Before you write, look at your inbox. Take the last 50 emails and sort them roughly. In almost every store the same topics come out, and for each topic you need two to four facts. Here is the list that is enough in practice:
- Delivery time. How many working days from order, from dispatch, separated by country. What applies before public holidays.
- Shipping costs and free-shipping threshold. Per country. What happens on a partial refund that drops below the threshold.
- Carriers. Who delivers where. Whether customers can choose the carrier. Whether parcel lockers are supported.
- First scan. How long it typically takes for a shipment to show a status after the label is created. One sentence on this saves you many delivery questions, more on that in the post on WISMO requests.
- Return window. Legal, voluntary, goodwill margin.
- Return route. Portal, label by email, customer pays their own postage. Which address. Whether original packaging is required.
- Return shipping costs. Who pays, in which case. Exceptions for defects.
- Refunds. How fast after receipt, by which method, whether store credit is an option.
- Exchanges. Whether you exchange directly or do return plus new order.
- Cancellations. Until when a customer can cancel. What if the parcel is already packed.
- Address changes. Until when possible, who changes it, what happens after dispatch.
- Damaged goods. Photo yes or no. Above which order value you reship without asking for a return.
- Wrong or missing items. Same pattern.
- Discount codes. Whether they can be applied retroactively. Whether they combine. What about expired codes.
- Payment methods. Which ones exist. What happens with invoice payment and reminders.
- Invoices. Where the customer finds them, whether a company address can be added afterwards.
- Availability. Whether items are restocked, whether pre-orders are possible, how you answer “sold out”.
- Sizing and fit. Only if you sell fashion or shoes: does the brand run large or small.
- Contact. Opening hours, which channels exist, when a human answers.
- Tone. Formal or informal address. What you call yourselves. Two sample emails that are typical for you.
If you write two or three sentences for each of these lines, you have about two pages. That is the knowledge base. Nothing more.
The hour, broken down
It helps to set a time limit. Otherwise “quickly write down the rules” turns into a project that never finishes.
Minute 0 to 10: skim the inbox
Open the last 50 emails. Do not count precisely, estimate. Which five topics come up most often. Write those first. Everything else can wait, because those are the cases that rarely occur and should land with you anyway.
Minute 10 to 40: write down the facts
Go through the list above, starting with the five most common topics. Write short sentences in the voice you use with customers. An example for returns:
We accept returns 30 days after delivery, unused and in the original packaging. The customer registers the return through the portal and receives a DHL label. The customer pays return postage, except for defects or wrong deliveries. Refund within five working days of receipt to the original payment method. Up to day 45 we take unused goods back as a goodwill gesture, after that we do not.
Five sentences, and an agent can answer nine out of ten return questions with them. The tenth case is the one that should come to you.
Minute 40 to 50: mark the limits
This is the part most people forget. Write down what the agent should not decide. For example:
- Refunds above a certain amount: only with approval.
- Complaints containing words like “lawyer”, “consumer protection”, “review”: to a human.
- Anything that needs the word “goodwill” and is not covered by a rule: to a human.
- Business customers, bulk orders, press enquiries: to a human.
This list is as important as the facts. A knowledge base without limits produces an agent that answers everything, including what it cannot answer. How to decide which cases may go out automatically is covered in the post on auto-send.
Minute 50 to 60: set the tone
Two or three emails from your inbox that sound the way you want to sound. Not the perfect ones, the typical ones. Add three sentences: formal or informal, what you call yourselves (“we” or “the team at …”), which phrases you never use. More on this in the post on brand tone of voice.
Common mistakes when building it
From the knowledge bases we have seen, a few patterns repeat:
- Too much text. Forty pages listing every exception. An agent does not find the right rule faster when it sits between thirty irrelevant ones. Cut to what is actually asked.
- Contradictions. 14 days on the FAQ page, 30 in the knowledge base, 45 in the last email to a customer. An agent cannot know which one is true. Clean that up before you automate. It is, by the way, the most common cause of wrong answers, more in the post on hallucinations in customer service.
- Rules nobody follows. If in practice you always take returns up to 45 days, write 45, not 30. The agent should reflect your practice, not an ideal you do not live by yourself.
- No date. Shipping costs change, deadlines change. Add to the end of every rule when it was last checked.
- Missing country rules. If you ship to Austria and Switzerland, you have different delivery times, different shipping costs and customs for Switzerland. One sentence per country is enough, but it has to be there.
How the knowledge base keeps growing
After the first hour the knowledge base is not finished, it is usable. The rest emerges in daily operation. Every time a case arrives that the base does not cover, you have two options: answer it manually once and forget it, or answer it and add the rule. The second option costs thirty seconds more and saves you the same question next time.
This is how it works at Humane: during setup you connect the store, answer a few questions on exactly these topics, and training starts. The first week is a training week in which the agent writes drafts and you approve or correct them. Every correction becomes a rule. If you change “14 days” to “30 days” in a draft, that does not need to happen a second time. The business rules engine then checks every reply against the stored rules, whether they came from onboarding or from a correction. How that looks in detail is in the post Correct once, right every time.
The principle applies regardless of the tool, though. A knowledge base that is written once goes stale. One that grows by a sentence with every exception will be better after three months than any document you could have written in a single afternoon.
What an agent still cannot know
To be honest: there are things that are not in any knowledge base and should not be. Whether you make one more concession to that one regular customer after the third complaint. Whether the batch that went out last week really had a problem. Whether the tone of an email is merely annoyed or already desperate. Those are the decisions you sit in support for.
A good knowledge base makes sure the agent answers the other cases reliably and puts exactly these cases in front of you, with all the data but without a finished reply. More on where that line runs in the pillar post AI in customer service: what actually works in 2026.
What you can do now
- Take the last 50 emails and write down the five most common topics.
- For those five topics, answer the questions from the list above, in short sentences, in your tone.
- Add what the agent is not allowed to decide.
- Clean up contradictions between store, FAQ and practice.
- With every new case the base does not cover: add one sentence.
If you want to see how two pages of rules turn into finished replies in your inbox: try 100 customer requests for free, no credit card, setup usually takes under 15 minutes.