No. An AI agent is only as good as its knowledge base, and on that point the cautious people are right: it needs information that is accurate, complete and organized. But you don't need to spend three or six months rewriting everything before you start. Your help center, your team's replies and your product manuals are enough to begin with, as long as AI helps organize them and your team approves what goes live.
Two things you hear about AI agents
When a brand starts looking into AI agents, it usually hears two stories.
The first comes from the dream sellers. AI fixes everything: plug it in, it understands your customers, and your customer service runs itself overnight. To hear them tell it, it's close to magic.
The second comes from people who have actually deployed one. The more experienced the person, the more you hear the same warning: a good agent needs well-organized, complete, reliable data, and that's harder than it sounds.
Honestly, the second story is right. It's the part that matters most in any AI agent project. Where I disagree is the conclusion people often draw from it: six months of knowledge base cleanup first, then the agent.
Why does the knowledge base matter so much?
Because an AI agent answers with what it reads. It doesn't know your return policy, your delivery times or your warranty terms: it finds them in your documents, or it doesn't.
In the brands I work with, as in the customer service team I used to run, I find the same three flaws almost every time:
- articles that are out of date: a delivery time, a carrier, a condition that changed after they were written,
- documents that contradict each other, because different people wrote them at different times,
- topics nobody ever wrote down, because the team knows them by heart.
A support agent makes up for these flaws without thinking about it. They know which version is current and who to ask when they don't know. An AI agent only has the text: it repeats what is out of date and may give either version of a rule. When information is missing, a well-built agent hands over to the team, and each of those handovers shows exactly what needs writing.
Do you have to reorganize everything before you start?
No. A brand that has run customer service for a few years already has almost everything it needs, scattered across four places:
- the help center and the team's saved replies,
- past tickets, meaning what your agents actually answer, day after day,
- product documents: manuals, spec sheets, warranty terms,
- a few people's heads, for the processes nobody ever wrote down.
Of the four, past tickets are often the most reliable. At one of the brands we work with, before the agent handled a single warranty claim, we went back over a few hundred past claims. The rules as they had been written down at the start rarely matched what the team actually did. The rules drawn from the claims themselves matched the team's practice in the vast majority of cases.
So the material is there, but in a shape that's hard to use: hand-drawn flowcharts, Google Sheets you have to pick apart line by line, articles that don't agree with each other. Rewriting it all by hand takes months, and meanwhile your customers keep asking the same questions.
I don't ask teams to spend three or six months reorganizing a messy knowledge base before they launch anything.
Start from what exists, with AI
In practice, I use AI for the heavy sorting and keep the team for the calls only it can make.
AI proposes the structure. From the help center, the saved replies and a sample of tickets, it groups topics and proposes an outline: delivery, returns, warranty, care, payment. The team corrects the outline instead of writing it.
Team replies become articles. What agents type by hand every day, and that no article covers, becomes a draft. It's approved when several agents give the same answer. When the rule depends on a decision by the brand, it waits for that decision.
Contradictions go into a table. Each disagreement between two sources becomes a row: what one says, what the other says, and a box for the team to pick the right version. The article is updated afterwards.
Product manuals become drafts. Dropped into a shared folder, they're read by AI, which turns each one into a draft article in every language and flags when the catalog listing promises something the manual doesn't describe. Nothing reaches the agent until someone approves it.
A few targeted meetings. There are always processes nobody wrote down: who decides on a goodwill gesture, what happens with a damaged parcel, when a refund kicks in. For those, a few short meetings with the right people are enough. No need for fifty of them.
I've learned not to overdo it. At an online retailer, we first generated a map, from the agent's conversations, of everything it didn't know. It was unreadable. We replaced it with a simple table, one row per customer situation: the information needed, the decision rule, who decides. Then I realized the brand had already set out most of those situations in its guarantee policy, written a few weeks earlier. The table's rows now come from that document.
What AI won't do for you
AI won't decide your rules. When two sources contradict each other, it can spot it, but it can't tell which one is right. Your team makes that call.
It won't make up missing information. If a product detail isn't written anywhere, someone has to ask the product team, then store the answer with the product data rather than in a general article.
It won't approve what commits the brand. Refunds, warranties, claims about what a product does: a draft based on a manual can repeat a catalog promise the product doesn't keep. A person reviews it before it goes live.
And approval takes some of your team's time. In theory that time shrinks week after week, because the agent only flags the real gaps.
Where should you start?
If you want to launch an AI agent in the coming weeks, I'd start with four things:
- gather what exists: the help center export, saved replies, a few weeks of tickets, product manuals,
- have AI propose a structure, then correct it as a team,
- list the contradictions and settle them before adding content,
- note the processes nobody wrote down, and plan a short meeting for each.
Then launch the agent on the topics that are best covered. The questions it can't answer will tell you what to write the following week. Our AI agent page explains how we build an agent from your own data.
FAQ
How many articles do you need to launch an AI agent?
There's no minimum. What matters is that your customers' most frequent questions are covered: delivery, order tracking, returns, payment, warranty. The rest comes over the following weeks, from the questions the agent couldn't answer.
Can you use past tickets to build the knowledge base?
Yes, they're the best source for what your team actually tells customers. An article drawn from tickets keeps only the rule or the information, never a customer's details, and a person approves it before the agent uses it. Our security page describes how we handle data.
Who should approve articles written by AI?
The head of customer service, or an experienced agent they choose. Review matters most for anything that commits the brand: refunds, warranties, goodwill gestures, product claims.
What should you do when your help center contradicts itself?
List the contradictions, have the team settle each one, then keep a single version in one place. The agent will read that version, and so will your customers.