Your veteran customer-service agent is leaving.
Only now does the owner realize: how to judge a return, how to calculate shipping, which customer can get an exception — all of it lives in one person's head.
Day one for the new hire: a customer asks "can I return it even after opening the package?" and the new hire answers "no". The customer fires back: "Your Xiaowang returned it for me last time!"
That is exactly the problem a customer-service knowledge base exists to solve.
First, a Counterintuitive Point
When organizing a knowledge base, do not start with process.
Start with exceptions.
Do not start with policy.
What actually makes a veteran agent valuable is not that they know the process — it is that they know when the process can be broken. She returned an opened package because she recognized this customer as a long-time buyer, the item was a standard product, and the opening did not involve hygiene concerns.
Ask her "what is our after-sales process?" and she will recite the policy at you. Ask her "what kinds of cases do you make exceptions for?" and only then will she hand you the real thing.
So ask the right questions in the interview:
- In what situations do you refuse a customer?
- In what situations do you compensate them on your own initiative?
- Which type of customer do you know, the moment you hear their voice, to handle with care?
Once these three questions are answered, the knowledge base has a skeleton.
Only then do you fill in the flesh.
Step 2: Write Answers as Cards, One Question per Card
One question, one answer.
Do not write a book. A book nobody ever opens — and the AI will not retrieve it accurately either.
Cards must be short, and they must carry numbers.
"Opened packages can be returned, provided: standard product, packaging intact, within 30 days, not custom-made. Custom-made items are never returnable."
There is only one test: can a new hire, following this card alone, answer the customer directly? If yes, it passes.
Step 3: Mark the Exceptions — Exceptions Are the Value
Attach an "exception" note under every rule.
"7-day no-reason return" is the rule. The exceptions are: custom-made items cannot be returned, fresh produce cannot be returned, activated software cannot be returned.
Why are exceptions so important? Because almost every customer complaint happens on an exception.
You can recite the rules as fluently as you like — hit an exception and answer it wrong, and you get scolded just the same.
The exception is the dividing line.
Step 4: Separate Three Layers, Do Not Mix Them
Most teams skip this step, and it is quick to do:
- The wording layer: the unified external message. Price, delivery times, returns and exchanges.
- The rules layer: the red lines. What you cannot promise, what you cannot discuss.
- The evidence layer: the source text. Platform rule documents, contract clauses, national standards.
The evidence layer sounds redundant. But the moment a customer gets tough, or a platform comes inspecting, the side that can produce the original text does not lose the argument.
Take AI customer service, for example: the national standard behind it is GB/T 47746—2026 (effective 1 September 2026) — 61 requirements in total, 48 shall, 4 should, 9 may, including 5 deal-breakers, among them the scenarios that must be handed over to a human automatically.
Copy those clauses into the evidence layer, and your AI's answers will have a backbone.
Step 5: Review Weekly — Do Not Build It and Abandon It
A knowledge base is not something you build once and forget.
Shipping prices change, promotions rotate, new products launch — if the library is not updated, the AI will keep answering from the old version.
Assign one person thirty minutes a week. Thirty minutes a week can save you a complaint dispute.
Two Traps
Trap 1: treating the SOP as the knowledge base.
An SOP is written for the people doing the work. A knowledge base is for the people answering customers (or the AI). Your "follow company policy, chapter 3" is something customers cannot parse and the AI cannot retrieve.
Trap 2: writing rules without scenarios.
"No-reason return does not apply" is a useless sentence. You have to write down which product, in which state, at what time the exemption applies. Miss one condition and the AI will answer it wrong.
One Last Line
What a veteran employee takes with them when they leave is not their length of service.
It is the hundreds of "in what situation can I break the rules" cases in their head.
While they are still here, ask for them.
Does your company have that one person who knows things no one else does? Tell us in the comments what they guard, and we may write about it in the next piece.
If this was useful, give it a like so more business owners see it.
Who we are: we help small and micro businesses build enterprise knowledge bases and AI customer service that actually ships — turning the wording, rules and scripts scattered in employees' heads into a traceable base the AI can look up, plus red-line interception and human-handover judgment.