Starting September 1, the "Requirements for Coordination Between Human and Intelligent Customer Service in Customer Contact Services" (GB/T 47746—2026) took effect. It is China's first national standard devoted specifically to how human and AI customer service should work together. The standard falls under the National Technical Committee for Service Standardization and was led by the China National Institute of Standardization, with more than thirty organizations taking part in drafting — telecom carriers, appliance makers such as Gree, Haier and TCL, and insurers such as Great Wall Life Insurance. All of them handle high call volumes and were among the first to put chatbots on the front line.
The standard covers four areas. First, it draws a line on which tasks go to whom: high-frequency, standardized queries go to AI, while complaints, disputes, emotionally charged conversations and complex business matters go to humans. Second, it requires that the option to transfer to a human agent be clearly visible, not buried under multiple menus or discouraged through a maze of steps. Third, it requires that conversation history and customer information be carried over automatically when a handoff happens, so customers do not have to repeat themselves from scratch. Fourth, it sets evaluation dimensions for service quality, covering response speed, semantic recognition, handoff efficiency and satisfaction. Two more provisions matter more directly to consumers: older people, people with disabilities and similar groups should be routed to a human agent by default, and any commitment involving price or refunds must be confirmed by a human. Companies are also held primarily responsible for what their AI customer service says — they can no longer wave off a promise as merely "algorithm-generated."
A Threshold Priced Into the Ledger
The timing of the standard's rollout tracks the direction complaint data has been moving. According to the China Consumers Association's first-half figures, consumer associations nationwide accepted 985,928 complaints, down 1.01% year over year, resolved 567,926 of them, and recovered 447 million yuan in economic losses. After-sales service issues made up the largest share, at 26.79%, and AI customer service was singled out for the first time as a category of its own — false promises, difficulty reaching a human agent, and inaccurate generated content were the three problems consumers raised most.
The numbers are clearer once you put them on a ledger. A human customer service agent costs at least 3,000 yuan a month to employ, more in first-tier cities; a full AI customer service system can be rented for as little as 99 yuan a month. That is roughly a thirtyfold gap. For a mid-sized e-commerce company, switching to a chatbot can cut direct labor spending by more than 1 million yuan a year. Since connecting a customer to a human agent is pure cost and talking them out of it is pure savings, burying the handoff a few menus deeper becomes a rational design choice — nested menus and longer queue times, the very things customers complain about, show up as positive numbers on the cost sheet.
Li Ning, a professor at Tsinghua University's School of Economics and Management, called this approach "a mistaken form of cost-cutting and efficiency gain," arguing that "on the surface, customer service costs go down, but the underlying problem is not actually solved." The cost has not disappeared — it has simply shifted from the company's payroll to the customer's time. And because the extra hour a customer loses never shows up on any company's financial statement, nobody is held accountable for it.
A Xinhua News Agency report published before the standard took effect described the experience of a consumer surnamed Wu, who got nowhere after repeated attempts to communicate and said it "felt like punching a pillow" — it took him more than an hour just to reach a human agent. A line circulating on social media put the tension even more sharply: "AI customer service is not bad at talking, it is bad at getting things done." Chatbots' language ability has long been good enough; what is missing is the authority to actually solve problems and someone who is accountable for the outcome.
How a Recommended Standard Grows Teeth
GB/T 47746—2026 is a recommended national standard, so a company that ignores it is not breaking the law — which is why many see it as toothless on paper. Its real leverage is not in penalties, but in evidence. Liao Huaixue, a senior partner at Taihe Pengxiang Law Firm, frames the logic as "reasonable reliance": if a company places AI at the front door of its official customer service channel, consumers have reason to believe it speaks for the company. Once that reasoning holds, disclaimers like "the AI's answer does not represent the company's position" or "algorithm-generated content carries no legal effect" become much harder to sustain. With the national standard's text in hand, consumer associations mediating disputes and market regulators conducting inspections now have a concrete yardstick for judging whether a human-transfer option was "clearly visible enough" or whether a promise "should have been honored," instead of arguing what counts as reasonable from scratch every time.
A recommended standard usually turns into a real constraint through three channels: being written into the technical requirements of government procurement and tendering, being cited by regulators during targeted inspections, and being treated as industry custom by courts and mediation bodies. How quickly those three channels move will decide whether the actual experience changes after September 1. Rough math: if a platform makes the human-transfer option visible right at the front page and syncs information automatically, cutting the time needed to resolve a single dispute by half, labor costs rise back up to match — some portion of that saved 1 million yuan has to be paid back. Companies understand this arithmetic better than anyone, and it is exactly where the resistance to implementation comes from.
The technical direction has shifted too. Over the past two years, customer-service bot improvements have concentrated on semantic understanding and multi-turn dialogue — in short, sounding more human. The standard moves the focus to coordination: recognizing when to step aside, handing over full context, and letting the human agent who takes over understand what happened within seconds. That is closer to an engineering problem, and how well it is solved depends on whether companies are willing to spend money on system integration.
On the standard's first day in effect, most platforms' customer service interfaces showed no visible change. The real test will come at the next spike in complaints, when the after-sales share in the consumer association's report gives the answer.
Sources: Xinhua News Agency, China Consumers Association complaint analysis, CocoLoop, National Public Service Platform for Standards Information; the standard's number and lead drafting body are checked against the official standard release, and complaint counts, resolution figures and after-sales share are drawn from the consumer association's bulletin.