China just told AI customer service it cannot promise you a refund
China spent the week writing rules for the AI it already runs — who has to answer you, what the state gets to see, and where the power comes from. All three landed within days of each other, and all three are the kind of plumbing decisions that outlast model releases.
China's first national standard for AI customer service, GB/T 47746—2026, took effect on September 1, and it targets the two tricks consumers hate most. The "transfer to a human" button must be easy to find, handoffs must carry the conversation so nobody repeats themselves, and anything touching price, discounts, refunds or compensation needs a human's final confirmation — the bot can explain the rules, collect documents and open a ticket, but it cannot close the deal. The China Consumers Association logged 986,000 complaints in the first half of 2026, with after-sales the largest category at 26.79%, and named bogus AI promises and unreachable humans as the new hotspots. The economics explain the behaviour: a basic AI客服 system rents from 99 yuan a month against more than 3,000 yuan for a human agent, and some companies reportedly grade staff on an "AI interception rate," with obstacle designs deflecting up to 70% of transfer requests. The standard is recommendatory — no direct enforcement — but it is explicitly written to be the yardstick for administrative inspections, consumer-association reviews and court rulings, which is exactly how a soft rule becomes a real one.
The MIIT says its national compute platform now monitors China's AI capacity as a single system. Announced on September 13 at the China Computing Power Conference in Langfang, the platform has all 31 provincial-level monitoring systems connected, more than 2,000 compute products listed, over 10,000 registered enterprise users and more than 300 large models attached, on a base of billions of accumulated compute-monitoring records. China Academy of Information and Communications Technology chief engineer He Baohong described the result as a national compute baseline that can finally be counted, sensed and tracked; MIIT chief engineer Zhong Zhihong promised a hub-region-edge architecture, millisecond city-level access and a compute standards roadmap next. The read: Beijing cannot direct, price or ration AI capacity until it can see it, and this is the instrument that makes the "East Data, West Computing" slogan schedulable. It is a dashboard, not a dispatch office — participation runs through enterprise registration, not seizure — but no Western government keeps an equivalent ledger.
"Compute-electricity coordination," first written into the government work report this year, is now in the 15th Five-Year energy plans. Inference wants to sit near users, and users are in the east, so the plan splits the country: the west hosts national compute hubs paired with big renewable bases, while the east co-plans distributed compute with distributed generation, microgrids and virtual power plants. Wuhu is the template — an ultra-high-voltage line from Shaanxi imports western power, a new 220 kV substation serves the data-centre cluster, the country's largest solar-plus-rail installation generates close to 10 million kWh a year, and an aggregation platform has sold more than 100 million kWh of rooftop green power to data centres since its first trade last October; 130-plus compute firms there now schedule workloads around price signals the way factories watch weather. The target to remember: virtual power plants delivering over 50 GW of regulation capacity by 2030, more than two Three Gorges dams. The United States is litigating data-centre grid connections in state capitals; China is redesigning the demand side of its grid around them.
What to watch: whether the compute standards roadmap Zhong Zhihong promised turns the platform from a register into a scheduling desk.
The rule is simple: a bot may explain, only a human may promise. Should every country copy China's split, or is it too late for chatbots this polite? Tell us in the comments.
