As new regulations come into effect, why is it still hard to find human customer service?
The reporter actually tested 20 mainstream platforms and mini programs. Industry analysis and assessment indicators have led some companies to "stop" switching to manual labor.
Reporter Li Si Intern Tang Zekai
Starting from September 1, the "Customer Contact Service - Collaboration Requirements for Artificial and Intelligent Customer Service" (GB/T 47746-2026) will be officially implemented. This national standard clearly requires that the manual entrance be clear and eye-catching, the human-machine switching is smooth, no repeated inquiries are made, and the enterprise assumes the main responsibility for AI responses.
It has been more than a month since the new regulations were implemented. Have the problems that AI customer service has troubled consumers been solved? Recently, reporters from the People's Voice Express column of Liberation Daily and Shangguan News conducted actual tests on 20 mainstream platforms and mini programs.
"Artificial" that cannot be found
The new regulations require that manual entrances be "clear and eye-catching." However, among the 20 mainstream platforms tested by reporters, only a few platforms have added a "human customer service" icon to the customer service interface. Almost all other platforms require you to talk to intelligent customer service first, and then look for manual customer service in the conversation.
Different platforms have obvious differences in access to manual customer service. QQ Music only needs to be entered once to automatically transfer; Quark Network Disk requires three consecutive inputs of "Switch to Manual" before a not-so-obvious "Contact Online Customer Service" button pops up above the input box. Some platforms require more triggers: for example, the KFC mini program requires five times, and the Douyin App requires ten times.
The triggering conditions are also different. After different levels of takeaway members click "Switch to manual", some can directly find manual customer service, while others need to continue to enter multiple times; in the Luckin Coffee applet, sometimes they only need to enter once to access, and sometimes they still cannot access after entering ten times in a row.
Such uncertain and repeated conversations make consumers particularly crazy. Consumer Mr. Li had an order and needed to contact the official customer service of a certain platform. After a long conversation with the intelligent customer service, the order was never completed. "I made it clear that I wanted to switch to artificial intelligence, but the other party still kept using AI to reply to me, which is very annoying."
Some platforms simply use intelligent customer service to pretend to be "manual". On October 5, a reporter asked about account login issues on the Hello App. After entering "Transfer to manual", a manual customer service button pops up at the top of the dialog box. After clicking it, it displays "We have received your question and are verifying the relevant information for you and will reply to you later." The reporter thought that the person talking to the customer service at this time was human customer service. However, it didn’t take long for the reporter to discover that the other party would reply with a long message of more than 20 to 30 words every time. The reporter was a little suspicious, so he asked the other party if it was a manual customer service, and the other party replied that it was "bicycle-specific customer service." Helpless, the reporter continued to try to enter "manual customer service". After eight consecutive times, the manpower customer service showed up. In the meantime, the other party kept replying with "I understand that you want to contact human customer service. I can continue to assist you with this problem." After several times, the other party directly replied with "The bicycle customer service hotline is 9517517" to guide the user to leave the page.
Some food delivery platforms will also trigger fake "manual customer service". During a takeout ordering process, the reporter tried to modify the meal content in the order, and the intelligent customer service took the lead in handling it. After the assistance failed, the reporter entered "transfer manual" three times to have the system "transfer the order to more professional customer service personnel for follow-up processing." At this time, an "Online Customer Service" logo pops up in the dialog box, number 2792. However, subsequent communications contained long passages of text and repeated content. The reporter quickly realized that this was a second layer of disguise. Enter "switch to manual" again and finally get through to manual customer service.
The so-called "exclusive customer service" and "online customer service" are actually agents pretending to be "artificial customer service". In order to deal with this routine, many consumers share "manual transfer tutorials" on social platforms. Ms. Li, who works on the Internet, told reporters that she had tried to use emotional words to trigger the algorithm, and even exaggeratedly described the consequences if there was no human customer service access, before finding a human customer service.
An uncountable “promise”
The new regulations require human customer service and intelligent customer service to be considered complete services. Wu Yansong, director of the Caibo Intelligent Governance Research Institute, who participated in the drafting of the standard, said in an interview that this requires the system to completely transmit interaction records when humans and machines switch, and companies need to be responsible for the content of AI customer service replies.
However, the false promises of AI customer service are still a major problem that troubles consumers. Not long ago, due to misoperation, Mr. Chen exchanged the 100GB permanent storage space in a certain network disk App account for a 24-hour fast download privilege coupon. "There is no eye-catching secondary confirmation pop-up window on the redemption page, and the redemption is successful with just one click." Mr. Chen hurriedly contacted customer service and wanted to withdraw as soon as possible. AI customer service said that the coupon can be withdrawn if it is not used, and also gave a detailed operation process. Mr. Chen followed the instructions, but could not find a way to cancel it. After multiple rounds of dialogue, the AI customer service changed its message and said that the privilege coupon does not support direct cancellation of the redemption operation. This left Mr. Chen a little confused. He asked many times to switch to manual service, but the reply he received every time was, "The manual customer service is busy. To save time, please submit a message to describe the problem in detail." Finally, the reporter saw the platform’s reply on the Black Cat Complaint Platform: “The space is redeemed for a high-speed download coupon, which cannot be withdrawn after redemption (please refer to the redemption rules for details). Your coupon withdrawal reply triggered by smart customer service is not the withdrawal method for this privileged coupon.”
Mr. Chen couldn't figure it out: he was asking about privilege coupons, so why didn't the revocation answer given by the AI count? What made him even more helpless was that from beginning to end, he could not really talk to anyone.
Coincidentally, on October 1, Mr. Cheng, who used Jimeng Agent to generate videos, fell into a similar dilemma. Because he was dissatisfied with the generated video, the official assistant proactively promised to issue two reissues of 600 points each, totaling 1,200 points, and promised that subsequent episodes can use the advanced VIP deduction-free generation channel. As a result, the account did not receive points, and paid points will still be deducted normally for subsequent episode generation. He contacted the online manual customer service, but the other party claimed that "AI technology limitations" and "feedback has been recorded" refused to fulfill the AI promise.
"Don't the new regulations make it clear that companies need to bear the main responsibility for AI responses?" Mr. Cheng's question is also a question in the minds of many consumers. “How can the saying ‘AI doesn’t count’ be a reason for the platform to shirk its responsibilities?”
Visible “costs”
Are these problems exposed in the actual measurement due to lack of technical capabilities, or are they intentional on the part of the company?
Xiao Li, who has worked in the customer complaint team of a company, told reporters that after the company started using AI customer service provided by a third-party supplier last year, its labor costs were reduced by more than half, and the working hours of the previous three shifts were also adjusted to 9 a.m. to 6 a.m. "AI customer service can completely handle some basic issues. Enterprises will definitely give priority to using AI customer service."
She also introduced that the company will set trigger conditions for manual customer service based on user group portraits. When it is found that there is an intention to complain in the conversation, confirmation of transfer to manual, or problems that cannot be solved by AI customer service, the transfer to manual service can be triggered. But why do users on some platforms still fail to trigger "convert to manual" after inputting it multiple times? Xiao Li bluntly said: "It should be a matter of strategy."
She told reporters that some companies have assessments for both AI customer service and manual customer service, mainly focusing on three types of indicators: efficiency indicators (self-service solution rate, manual transfer rate, one-time solution rate), service quality indicators (answer accuracy, intention recognition rate, knowledge coverage rate) and user satisfaction indicators (satisfaction, complaint rate). "Labor transfer rate" is listed as an efficiency indicator, which means that the lower the labor transfer rate, the higher the efficiency. This directly explains why companies have to "stop" switching to manual labor - it is not a technical issue, but a KPI orientation.
Of course, the cost account behind this is more intuitive. As a customer, the reporter consulted a merchant that provides AI customer service. The other party said that the basic package including 5 stores and 350,000 traffic only costs 3,999 yuan a year, which is suitable for an average of 50-100 inquiries per day. In its brochure, it clearly lists configuration methods such as "human-machine collaboration assisted reception" and "no reply", emphasizing that the risks of AI are controllable. In other words, switching to a manual strategy is a configurable item from start to finish, rather than a technical problem. This means that in the actual test, "it takes three inputs to trigger manual transfer" and "it still fails to trigger three times and you can only fill in the form", which are intentional obstacles rather than a lack of technical ability.
In this regard, Yu Xun, a professor and doctoral supervisor at East China University of Political Science and Law, believes that some platforms use AI to intercept user complaints as a core assessment indicator, and use the time and cost of rights protection to encourage consumers to voluntarily give up complaints. Similar behavior by enterprises may constitute intentional delay and is suspected of violating Article 48 of the Consumer Rights Protection Law. Anyone who intentionally delays or unreasonably refuses a consumer's request for repairs, returns, refunds, etc. shall bear civil liability.
"Technology is not a bottleneck. Opening up data links requires investment in R&D resources. The old customer service system architecture is solidified and transformation requires costs - this is a problem that companies are unwilling to invest in, rather than a problem of capabilities." Yu Xun also pointed out that some platforms link customer service performance with "keeping users in the AI layer", which confuses the problem interception rate and problem resolution rate, which is putting the cart before the horse.
Implementable “national standards”
In response to these problems, Yu Xun introduced that the new national standard actually places clear constraints on many key aspects of AI customer service scenarios. Scenarios involving complex disputes, financial disputes, personal and property safety risks, and users' explicit request to transfer to manual processing must be transferred in a timely manner, and key matters such as price, refund, and compensation are finally confirmed by manual processing. Enterprises must be responsible for the content of AI customer service replies and cannot shirk responsibility on the grounds of "automatic generation by algorithms".
Yu Xun added: "It is necessary to distinguish between two situations: the AI customer service output content deployed by the operator should be regarded as the operator's intention, and the operator should fulfill the contract. However, if it is a general AI service (such as an independent AI assistant, a chat robot), it is generally not directly regarded as a formal commitment of a specific operator, and mainly plays an information assistance and reference role." This means that the commitment of the AI customer service embedded in the platform should be regarded as an expression of the platform's intention, and the platform cannot shirk responsibility by saying that "AI does not mean it".
Although the new national standard is a recommended standard, it is not mandatory in itself and cannot alone be used as a direct basis for administrative penalties. However, Article 511 of the Civil Code stipulates: When the contract quality requirements are unclear and there are no mandatory national standards, the recommended national standards shall be followed. This provides a superior legal basis for the application of national standards in contract disputes. Moreover, when market supervision departments handle consumer complaints, this standard can be applied in conjunction with the Consumer Rights Protection Law, as a reference standard for determining whether an enterprise constitutes "intentional delay." In relevant litigation, this standard serves as a relatively clear technical yardstick, which can reduce the difficulty of providing evidence for consumers' rights protection to a certain extent.
How to eliminate chaos in AI customer service and make AI better serve people? Yu Xun suggested establishing a combined mechanism of “credit files + administrative supervision + judicial adjudication”. The binding force of credit files lies in the fact that continued disregard of standards and a large number of user complaints will bring about pressure from public opinion and administrative constraints. Relevant information will be recorded in corporate credit files and disclosed to the public, which will force the platform to complete system rectification.
"Convenience switching to manual labor is not entirely a burden. In the short term, it does increase labor costs, but simply positioning customer service as a 'cost center' itself is a strategic misjudgment." Yu Xun said that companies should release human agents from low-value labor, let AI take over high-frequency repetitive problems, and let humans shift to high-value links such as pre-sales conversion and after-sales order collection. The new national standard is designed according to this logic, and enterprises should regard it as an opportunity to optimize their service architecture.
After all, only when consumers can easily compare the ease of manual access across platforms will the service experience itself become a competitive factor.
