Zhang runs a shop selling digital accessories. His store has exactly one customer service person, who also handles packing.
Last month he received a fine notice: the service rep had written "the lowest price on the whole internet" in a chat. Under the platform's new quality-check rules that counts as an absolute-term violation, and the fine was 1,200 yuan. He was still dazed when he showed me the screenshot — "I shipped the goods and I refunded the buyer. How is that a violation?"
One sentence, 1,200 yuan. That is the most concrete change in the service industry this year.
First, three things need to be clear
The "new Taobao customer service quality-check rules" circulating in the trade come down to three points: when a service rep uses a banned word, makes a false promise or shows a poor attitude, a fine is levied per ticket, 200 to 2,000 yuan each; the platform no longer samples but puts 100% of all conversations through quality checking; and a new emotional service assessment requires that a buyer's mood shift be detected and responded to within 8 seconds.
(These points come mainly from industry interpretations and vendor marketing, which makes them a market signal rather than the official text. But for an official line: the national standard GB/T 47746—2026 took effect on 1 September, and it explicitly names 5 categories of scenario that must hand over to a human automatically — platform rules will only get finer, and that direction is certain.)
The problem is not "was the attitude good" but "can you hold up"
We used to call customer service quality "service level". Now it has become a compliance cost. And the two are met in conflicting ways:
A good attitude relies on people, on training, on experience.
Holding up relies on systems, on full coverage, on traceability.
A shop can only afford one person. Are you going to have him read 8,000 chat records a day? He cannot finish. Are you going to have him judge within 8 seconds whether the other person is angry? He is not a machine. Once the rules are upgraded, the ones who lose out most are precisely the small merchants with the fewest hands.
Two actions you can take today
One: move "things you cannot say" out of training and into the system.
Training lives in people's heads, and people forget the moment they are under pressure; red lines written into the system mean every reply gets checked. The difference between the two is the difference between "explaining afterwards" and "blocking on the spot". Concretely: first list five items each for absolute terms, promise-style wording and attitude landmines, and have the AI stop at generation time — not making the rep memorise them, but making the AI not say them.
Two: hand "answering accurately" to the knowledge base, and keep "judging correctly" with people.
The root cause of AI customer service answering wildly is not the model, it is that it does not have your company's knowledge. If your return rules, shipping-charge standards, warranty boundaries and campaign definitions live only in a veteran employee's memory and a few scattered Excel files, then any AI hooked up to it can only guess. Organise these into a searchable foundation (we call it the "answer layer") and the AI can answer accurately; deciding red lines and whether to hand over to a human still rests with people.
To be honest
I do not think AI customer service can solve your compliance problem for you. The rules are set by the platform and enforced erratically, and nobody can promise that "using it means you will not be fined".
But one thing is certain: rules will always get finer faster than you can train your staff. Rather than reviewing after every incident that "that sentence should not have been said", put the rules in front of the system — making the AI say one fewer wrong thing is cheaper than ten explanations afterwards.
Who we are: SavantCat, delivering knowledge bases and AI customer service for small and micro businesses.