Last week the owner of a chain of stores asked me: a company had quoted him an "AI customer service" for only 99 yuan a year, cheaper than hiring a part-time student, and he wanted to know whether he should replace his current setup.
I asked him one question: with your current customer service, is it that customers' questions go unanswered, or that they are answered and you simply think it is expensive?
He thought for a few seconds and said: the questions go unanswered.
Then it is not a price problem. A 99 yuan customer service and a 9,999 yuan customer service use the same brain — and what that brain is filled with is your own company's material.
First, a fact: AI's "utility bill" really has collapsed
The liveliest thing in the industry over the past fortnight is large model API prices falling off a cliff. (Large model = the "engine" behind the kind of AI that talks, and DeepSeek, Tongyi Qianwen and Doubao all count; API = the channel programs use to call each other.)
- Start with 8 September, when DeepSeek adjusted pricing for the Flash series: the input price for a "cache hit" fell to 0.02 yuan, against 1 yuan for a miss — asking the same question a second time costs 1/50 of the first time. (In plain language, a "cache hit" means you asked this question before, the machine still holds the answer, so it does not have to think it through again.)
- On 19 September, Alibaba's Qianwen released an omni-modal model and cut audio and video API prices by 98%.
- Around the same time, Xiaomi's large model announced a permanent 99% price cut; even People's Daily published a piece saying this wave of API price cuts (API being the channel programs use to call each other) sends three signals.
In language an owner can follow: AI's "call charge" is now so cheap it can be ignored. Have AI answer ten thousand customer questions today and the model bill may come to less than a meal.
So here is the question: if it is that cheap, why does your AI customer service still fail to answer?
I gave it a name: **the last-mile bill**
Think of AI customer service as ordering takeout.
The model call fee = the delivery rider fee. The rider fee has now fallen to a few cents, or even to zero.
But for the food to reach your door, you first have to write the address clearly — building, entrance, door number. Get it wrong and the rider just circles the estate.
That address is your company's knowledge. When a customer asks "can that plan of yours be refunded", the answer is in a contract you signed three years ago, in a veteran employee's head, in an Excel file nobody ever cleaned up — however cheap the rider fee is, the food still does not arrive.
So the real bill looks like this:
| Line item | Then | Now |
|---|---|---|
| Model call fee (rider fee) | The bulk of it, enough to put small businesses off | Approaching zero |
| Knowledge organisation and maintenance (writing the address) | Ignored by everyone | The only large item left |
| Acceptance testing (did it arrive) | Based on gut feel | Now you have to produce evidence |
What got cheaper is not "customer service". Only the "mouth" got cheaper. The brain is still yours to supply.
Where the money actually goes: three line items
When we build this kind of thing for shops and small businesses, the money really goes to three places:
The first: turning scattered material into something searchable. The product manual is a PDF, after-sales rules live in Excel, the official line on policy lives in a veteran's experience. To make it answerable it first has to be broken into individual retrievable items. A store-level business usually yields 200–400 Q&A items, maintained once a month afterwards. (The unit is "items", not "books" — because customers do not ask a whole manual at a time.)
The second: acceptance testing. Before delivery we run two kinds of test: 60 real colloquial questions (the kind customers actually ask, such as "can you take a look at this for me" or "can that one of yours be refunded"), to see how many it drops; and 22 automated acceptance checks (whether the deployment is healthy, whether the handover to a human actually works, whether any key has leaked into the browser, and so on).
Why go to the trouble? Because on 17 September someone who has spent two years cutting customer service costs with AI published a piece with a very honest title: "To be honest: I have never once been able to calculate that 70% figure." The "70% cost reduction" repeated everywhere in the industry cannot be reproduced even by someone working in it. Only what can be verified counts as a result; what cannot be calculated is just a story.
The third: maintenance. This is the one most easily ignored: leave a knowledge base alone for three months and it starts to rot. Prices changed but were not updated, a promotion ended but was not taken down, a new policy never got entered — a customer getting an expired answer hurts more than getting no answer. So we usually suggest setting aside one hour a month to go through the new questions customers have asked and add them in.
Three things you can do now (they cost nothing)
One: write down the 20 questions customers ask most, and answer them by hand. It takes an hour, and it becomes your "acceptance standard" — if you cannot produce those 20 items, you are not ready for any AI customer service, and whichever model you buy first is a waste.
Two: take those 20 questions and test any AI customer service with them. Do not listen to the sales pitch — read the questions out and count how many it gets right. Twelve or more is a pass; five or fewer and you should not buy it however cheap it is.
Three: ask for a reproducible acceptance checklist. Not a claim like "a 70% improvement", but "which 60 questions you will run, and whether I can see the results on the spot". That one sentence filters out nine out of ten suppliers.
In fairness
Cheaper models are genuinely useful. Their greatest value is that small businesses which never dared try before can now start for a few hundred yuan — unthinkable two years ago.
I am not saying the price cuts are useless. I am saying: do not mistake "cheap" for "usable". An AI customer service that cannot answer customers' questions is expensive even at 99 yuan a year; one that answers "can this be refunded" clearly is cheap even at 3,000 yuan a year.
The model is a rented brain; the knowledge is your own family capital.
The national AI customer service standard that took effect on 1 September requires that "handover to a human must have a clear entry point" — note that it governs whether the entry point is clear, not whether the AI is clever. The standard never expected AI to know everything; it only requires that you do not use AI as a shield.
SavantCat · focused on knowledge bases and AI customer service for small and micro businesses: product manuals, after-sales rules and veteran employees' experience turned into a knowledge base that answers correctly on the first try, with a real colloquial regression run and automated acceptance before delivery.
Has an AI customer service ever wound you up in your own shop? Was it that you could not reach a human, or that the answers were completely beside the point? Comment with the exact sentence and I will pick a few typical ones to show you how to fix them. If this was useful, hit like so more owners who have been driven up the wall by AI customer service see it.