Consumer association data shows 986,000 complaints accepted in the first half of 2026, with AI customer service a new hotspot. The national standard GB/T 47746—2026 took effect on 1 September: these 5 situations require AI to hand over to a human, plus three things small businesses should self-check.
Ms. Wu, in Xuzhou, Jiangsu, sent more than a dozen messages into a chat box, and the same line came back every time: "Dear, I understand how you feel" — followed by a standard menu. After struggling for more than an hour she finally reached a human, and the return was settled in two or three sentences. She did not know whether to laugh or cry: it was exhausting.
That is not a joke. It is the opening of a People's Daily field report published on 11 September.
If you run a shop or a small business, or your company has just put AI on customer service, this article is written for you: AI customer service is not something you finish by buying a system — it has 5 "hand over to a human immediately" switches, and if one is missing, complaints start rising right there.
A term to translate first: AI customer service means a bot built on a large model that answers customer questions automatically. It can cover the workload of dozens of staff, and a basic edition may rent for only around a hundred yuan a month.
1. Two sets of numbers show how widespread this is
The first set comes from the China Consumers Association: in the first half of 2026, consumer associations nationwide accepted 986,000 consumer complaints (down 1.01% year on year), and "after-sales service" was the largest category at 26.79%. What is surprising is that the problems consumers reported cluster around false promises by AI customer service, difficulty reaching a human agent, and inaccurate generated content — in other words, AI customer service became the new hotspot for complaints in the first half of this year.
The second set comes from a reporter's own search: on a third-party consumer complaint platform, the keyword "AI customer service" returns more than 4,800 related complaints. The recurring grumbles are just a few lines:
- "I searched the whole app and could not find the human agent entry"
- "I sent 'human agent' more than a dozen times and all it says is 'I understand how you feel'"
- "The AI customer service promised a voucher and the human said it does not count"
False promises: the AI agreed offhandedly, and the company would not honour it afterwards.
2. Why are customers so angry? Three real failures
First, layer upon layer of barriers to a human. Mr. Zhou, in Shenzhen, bought a product by mistake on a gaming platform; under the platform's own rules, a refund can be requested as long as the item has not been claimed. But he could only get AI replies, and was repeatedly bounced between a mini program, a service account and the app. On the phone and in the online chat, asking for a "human agent" still produced the same looping response.
Worse than an answer that misses the point is being unable to reach anyone in an emergency. This June, Ms. Wang's smart door lock failed and the family of three was locked outside — including a baby just over six months old, waiting to get in for a feed. She called after-sales immediately and got a mechanical AI customer service recording, with no way to reach a human. The family waited outside for 50 minutes and finally had to call a locksmith.
Second, the AI promises offhandedly and the company refuses to honour it. Before Mr. Gong paid a 2,000 yuan deposit on a used-car platform, the platform's official AI customer service answered clearly: the vehicle price includes the title transfer fee and the plate registration fee. But at the actual transaction, the salesperson demanded another 1,000 yuan for the transfer. He argued with the screenshot in hand and was told: "An AI answer does not represent the company's position."
Third, a general-purpose AI says the wrong thing in a critical situation. According to Science and Technology Daily, Mr. Li was selling a sound card on a second-hand trading platform and was told that "AI hosting improves the sell-through rate". He agreed, and the platform's AI customer service cut the 1,000 yuan listing to 400 yuan and sold it.
Why do failures like these cluster? An industry insider's arithmetic is telling: a human agent costs at least 3,000 yuan a month once salary, social insurance, training and desk space are combined; a basic AI customer service system rents for as little as 99 yuan a month and can do the workload of dozens of human agents — a gap of about 30 times.
Saving money is not wrong in itself; what is wrong is turning a "service ledger" into an "interception ledger". Chen Yinjiang, deputy secretary-general of the Consumer Protection Law Institute of the China Law Society, puts it bluntly: when cost reduction becomes the first goal, AI customer service stops being just a "service assistant" and is also given the job of intercepting requests — add a few more menu layers, loop the responses a few more times, wear the consumer's patience down, and fewer people pursue their rights while the complaint rate on the report looks better. Industry insiders say some companies have shifted the focus of performance reviews from "problem resolution rate" to "AI interception rate", and some "obstacle designs" turn away seven out of ten requests for a human.
Su Haopeng, professor at the Law School of the University of International Business and Economics, puts the arithmetic back where it started: on the surface customer service costs fell, but the problems were not solved — only postponed and covered up, converted into repeat enquiries, spreading public criticism, lost returns and brand damage. That hidden cost is far higher than the labour saved.
3. From 1 September, the rules changed
On 1 September 2026, 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), formally took effect — and it addresses exactly the chaos described above.
GB/T marks a recommended national standard. "Recommended" does not mean useless: it is an important reference for regulators carrying out administrative inspections and service quality assessments, and an important basis for consumer claims and judicial decisions. Companies that do not meet the standard may be required to rectify.
Wu Yansong, president of the Caibo Institute of Smart Governance, who took part in drafting the standard, sums up four key points:
1. A visible entry point — hiding the human agent entry behind layer after layer is prohibited;
2. A clear division of labour — complex disputes, financial disputes, risks to personal or property safety, and situations where the user explicitly asks for a human must be transferred promptly; key matters such as price, refunds and compensation are confirmed by a human in the end;
3. Smooth switching — when switching between human and machine, the system must pass on the full interaction record and must not require the consumer to repeat themselves;
4. Clear responsibility — companies are responsible for the content of AI customer service replies and cannot shift responsibility by saying "the algorithm generated it automatically".
Liao Huaixue, legal counsel to the China Consumers Association, explains the last point even more plainly: "This kind of 'blaming the technology' does not stand up in law." An AI customer service deployed by the company itself is part of its service chain, and answers given on its official platform represent the company's position; consumer protection law requires operators to give truthful and clear answers to consumer enquiries — and that obligation does not change depending on whether a human or an AI is answering.
4. The most important part: 5 situations where AI must hand over immediately
The clause small businesses most need to remember is 5.2.2.6. It lists five categories of scenario where transfer to a human must happen automatically; hitting any one of them means the call should be transferred:
1. The interaction reaches the failure threshold and still cannot be understood — after several rounds it still has not answered the point, so hand over;
2. The customer explicitly refuses intelligent service — as soon as they say "human agent" or "stop using a bot", hand over;
3. The conversation involves information security — account numbers, verification codes, identity information and the like;
4. Emergencies or risk scenarios involving personal or property safety — the mother locked out of her home above falls exactly into this category;
5. An intelligent response times out — the example given in the standard text is exceeding 3 minutes.
Two pitfalls that are easiest to fall into:
- How many rounds is the "failure threshold" exactly? The standard gives no single number. Each company needs to set its own figure based on its business and write it into its internal SOP, so that whoever applies it uses the same ruler.
- A customer's tone getting worse is not the same as those five categories. Sentiment detection is a separate trigger path in the standard; do not confuse the two — if you want to be safe, configure both.
5. Start today: three things that cost nothing
First: move the "human agent" entry to the first layer. The goal is to reach a person in two clicks or fewer. This is the easiest one to meet and has the most direct effect.
Second: configure automatic transfer for all five categories above. When you are done, write "how many rounds before the failure threshold, how many minutes before timeout, who takes the transfer" as a three-line SOP and stick it next to the customer service desk.
Third: keep evidence. Screenshots of the configuration, trigger logs for human handover, script and SOP documents, training records — if a dispute does arise, being able to show "this is what we did at the time" is enough. A company self-check can go through the checklist item by item; a complete checklist runs to 61 items (48 "shall" + 4 "should" + 9 "may", including 5 deal-breakers). Clear the "shall" items and the deal-breakers first, then fill in the "should/may" items.
One more tip that is most practical for a small business: for anything involving price, refunds or compensation, have the AI say only "let me pass this to a human to confirm", and never let it make a promise. One careless sentence can cost you a sale and a customer.
There is nothing wrong with AI customer service itself. The job it should do is catch the questions that repeat every day, like "what time do you open" and "can you issue an invoice" — that is where it saves money. What actually goes wrong is treating it as a wall that blocks customers.
Who we are: we help small and micro businesses turn "the questions customers ask every day" into AI customer service that answers correctly on the first try, built to meet the national standard and to pass compliance registration.
How far we can take it: a single store or small business can be live in as little as one week; we hold the national standard self-check list, and you can go through it against your own setup at no cost first.
The problems you may be facing: customer service answers that miss the point, negative reviews piling up, and experience leaving with the veteran employee.
#AI-customer-service #SME #service-quality #compliant-operation #AI-adoption
Has an AI customer service ever sent you round in circles? Or if your business has adopted one, how have customers reacted? Tell us in the comments.
If this was useful, give it a like or a "wow" so more people running businesses see it.