Have you ever been through this: returning something bought online, you search the whole app and cannot find the human agent entry; you send "human agent" to the AI customer service more than a dozen times and it just loops "I understand how you feel"; you finally get "your refund has been registered", and then the human agent says "the intelligent customer service reply is for reference only".
>
From today (1 September), all of this finally has a national standard. China's first national standard focused on the collaboration mechanism between human and intelligent customer service, "Customer contact services — Requirements for collaboration between human and intelligent customer service" (GB/T 47746—2026), has formally taken effect.
1. How many times have we been let down by AI customer service?
Two real cases first; you will very likely recognise the feeling.
Mr. Li was selling a sound card on a second-hand platform. The platform told him that turning on "AI hosting" would improve the sell-through rate. He agreed. The platform's AI customer service then cut the 1,000 yuan listing to 400 yuan and sold it — the buyer haggled, the AI nodded, and the seller never said a word.
Mr. Ma, a food delivery rider, was fined 50 yuan by the platform and appealed to customer service. The AI customer service readily agreed to a refund, but as soon as a human agent stepped in, they said "the intelligent customer service reply is for reference only". His appeal failed twice, and the AI behaved like a broken record, never answering the question directly.
These are not isolated cases. The China Consumers Association's analysis of complaints in the first half of 2026 shows that after-sales service problems account for 26.79% of all complaints, the largest of any complaint category, and the most prominent issues within it are precisely false promises by AI customer service and difficulty reaching a human agent.
One media outlet tested 20 platforms and found that reaching a human agent took up to 660 seconds, that is 11 minutes. "Human agent speedrun tips" have even started circulating on social media — apparently saying "I want to complain to the regulator" gets you a live person in 5 seconds.
Since when does reaching a human agent require a script and a strategy guide?
2. Why do companies hide the handover to a human so deep?
The answer is not that the technology cannot do it. It is the arithmetic.
An e-commerce AI service provider did the sums for CCTV: a human agent costs more than 3,000 yuan a month, while their AI customer service annual package sells for 10,000 yuan and saves over seventy percent of the cost. In a first-tier city, adding social insurance, training and desk space, one live agent costs as much as 80,000 to 120,000 yuan a year.
Faced with a gap of dozens of times, the instinctive corporate response is to use AI as an "invisible filter": block what you can, because every handover to a human avoided is money saved.
More insidious still, the digital economist Liu Xingliang pointed out another layer on CCTV: some companies deliberately have AI customer service circle around a user's complaint, draining their energy and patience and indirectly raising the rate at which cases are dropped.
In other words, in some situations AI customer service is no longer a tool that solves your problem but a barrier that talks you out of pursuing your rights.
3. Three hard provisions in the new standard, each hitting a real sore point
The standard that has just taken effect sets three rules for AI customer service:
First, the entry point for handing over to a human must be clearly visible and must not be hidden behind layer after layer. In emergencies involving personal or property safety, the system must route to a human automatically; for older people and similar groups, human service must be given priority.
Second, five scenarios must transfer to a human automatically. The number of interactions reaching the threshold, the user explicitly refusing intelligent service, matters involving personal or property safety, risk scenarios, and an intelligent response timing out — in these situations the AI must not keep the matter to itself.
Third, what the AI says, the company must stand behind. This is the most important one. Operators can no longer refuse to honour a promise on the grounds that "an AI answer does not represent the company's position" or that "algorithm-generated content has no legal effect". Companies are responsible for the content of intelligent customer service replies.
One thing also needs to be clear: this is a recommended national standard. It is not directly a basis for penalties, and a company that does not follow it will not be directly fined. But it will become an important benchmark for the industry and a strong basis for your own claim — if an AI customer service lets you down from now on, you can cite the standard number and complain through 12315 or 12345, and the regulator has a ruler to measure with.
4. What does this mean?
For every ordinary consumer: being pushed around by AI customer service no longer has to end in "never mind, I will live with it". You have something to cite and a channel to complain through. AI customer service can no longer wear you down with looping scripts until you give up.
For businesses: AI customer service has gone from a "cost-saving tool" to a "vehicle of responsibility". Using AI as a shield again will only bring complaints, regulatory attention and lost reputation. A compliant "handover to a human as the fallback plus human-machine collaboration" has become a required answer, not a bonus point.
For founders building AI services: this is a watershed for industry compliance. Low-quality AI customer service that only loops scripts, hides the handover entry and writes cheques it cannot cash will be eliminated; those who can offer a compliant solution with "handover to a human as the fallback, human-machine collaboration and traceable responsibility" instead get a window of opportunity. Policy is using a national standard to educate the market on behalf of people who take service seriously.
5. Has an AI customer service ever wound you up? Let's talk in the comments
Honestly, while writing this I was sent round in circles three times by one platform's AI customer service myself. At the time it just felt frustrating; looking back, that frustration was the product of a precise cost design — companies use AI to block people, users use scripts to break through, and the two sides wear each other down against the wall that is AI customer service.
The new standard has put a crack in that wall. It is not a cure-all; a recommended standard still needs regulators and the industry to push it into practice. But at least the sentence "what AI customer service says, the company must stand behind" — nine characters in the Chinese — hands a key to everyone who has ever been wound up by an AI.
If an AI customer service has let you down, or your company is agonising over whether and how to adopt one compliantly, tell us in the comments. Your real experience is more convincing than any statistic.
If this was useful, give it a like or a "wow" so more of the people who need it see it — after all, who has not been wound up by an AI customer service?
#AI-customer-service-standard #consumer-rights #artificial-intelligence #human-machine-collaboration #compliance
Sources: China National Radio of China Media Group; the China Consumers Association analysis of complaints in the first half of 2026; the GB/T 47746—2026 standard document issued by the State Administration for Market Regulation and the Standardization Administration of China; the Caibo Institute of Smart Governance (a drafting body for the standard). Key figures were cross-checked against more than two independent sources.