Wu Linan remembers the excitement that surged throughout the company when ChatGPT first came out, "From the CEO to the investors, we have meetings every day to discuss how to integrate with the product." Put into a 4-month R&D battle with the team. The goal is clear: integrate the GPT large model into the existing Sobot intelligent customer service robot system.
At that time, the fire of ChatGPT had not yet spread to the country, and the "AI substitution theory" was far less anxiety-inducing than it is now. However, "customer service" has become one of the industries "will be subverted by ChatGPT" in the predictions of "Harvard Business Review", "Forbes" and other media.
Indeed, from the perspective of inputting questions and giving answers, ChatGPT is really like a customer service. Customer service, an enterprise service track, has been transformed by AI from 2014 to 2016. Natural language model (NLP) technology was applied to the mature "cloud call center" at that time to form an intelligent customer service system.
When you are shopping online, the pre-sales and after-sales services of consulting customer service have the participation of intelligent customer service. Many standardized questions and answers have been taken over by robots, which have become an important supplement to manual customer service. Behind the intelligent customer service system embedded in various scenarios, there are many years of "involution" in the track by enterprise service companies such as Wisdom Teeth Technology, Tencent Qidian, Udesk, Meicha, and Kuaishangtong.
Then, after the emergence of generative AI represented by ChatGPT, what changes have occurred in intelligent customer service?
On April 4th, Wisdom Teeth Technology's Sobot x GPT large language model fusion intelligent customer service solution was launched. According to test data, after the intelligent customer service system + AIGC, the direct answer rate of customer service increased by 15%-35%, and the answer accuracy rate increased by 3% %-32%, the operational workload of the enterprise in the customer service link can be shortened from 5 days to 0.5 days.
Has human customer service really been completely subverted? Wu Linan believes that "it is empowerment rather than complete subversion." There are still many scenarios that can be reformed in the customer service link, and AI is an important technical force.
Operational workload reduced from 5 days to 0.5 days
Wisdom tooth technology was "ignited" by GPT from the moment ChatGPT was released. Wu Linan found that the whole company, not only his technical classmates were paying attention, but also CEO Xu Yi and investors were all trying out this chat robot.
As the CTO of an intelligent customer service research and development company, Wu Linan can hardly fail to notice the development of artificial intelligence. In addition, the CEO Xu Yi is also a product manager, and his sensitivity to new technologies is a "self-contained buff". When the GPT model evolved to 3.0, they followed up. "At that time, everyone thought that it might only be used to run some models, and it could not be used for civilian or commercial purposes."
No one expected that on November 30, 2022, ChatGPT would be directly promoted to the public in the form of a product, and the underlying large model would be iterated to GPT 3.5. 'Crazy', everyone is very excited, thinking how this new model can be combined with the company's existing intelligent customer service robot."
Since then, Wisdom Teeth Technology has almost entered the state of "full staff GPT", holding meetings, training, and brainstorming. In December, Wisdom Teeth Technology's intelligent customer service system "+GPT" was formally established. Xu Yi took the lead, and the algorithm group, product group, marketing department and other departments cooperated, and the research and development began.
Even externally, ChatGPT, which is good at multiple rounds of dialogue, is the first thing people think of as a customer service robot.
Also in December, the technology blogger "Algorithm Pig Liye in Microsoft" personally made ChatGPT a guest of Swarovski's pre-sales customer service. This dialogue robot can give answers related to brands, products and prices, but the information is not enough Comprehensive, this is related to the fact that the ChatGPT database is only updated until 2021 and is not connected to the Internet.
The conclusion of "Algorithm pigs are established in Microsoft" is that it is not enough to let ChatGPT directly serve as customer service, but if it can grasp more data, such as orders, products, selling prices and other information, effective answers will be greatly improved.
The technology blogger's prediction has been verified in Wisdom Teeth Technology, and there are more accurate improvement data.
On April 4, Wisdom Teeth Technology Sobot x GPT large language model fusion intelligent customer service solution was launched, and brand customers began to test the "+GPT" intelligent customer service robot.
In the following month, Wisdom Teeth Technology's customer test data showed that after the Sobot intelligent customer service system + AIGC, the customer service's direct answer rate increased by 15%-35%, and the answer accuracy rate increased by 3%-32%, while the company's customer service link increased by 15%-35%. The operational workload can be shortened from 5 days to 0.5 days.
The data comes from Wisdom Teeth Technology
Efficiency improvement is the most obvious upgrade that AI brings to the customer service industry. The context understanding, sorting and summarization capabilities of large language models such as GPT and the human feedback pre-training technology behind them can successfully match the entire link of the intelligent customer service system "knowledge base configuration-question learning-robot answering", thus bringing improvements in the entire link. effect.
A large amount of knowledge and precise questions and answers are the prerequisites for the existing intelligent customer service robots to answer consumers' questions fluently. What cannot be seen is that enterprises invest a lot of manpower when operating intelligent customer service robots.
Wu Linan explained, "For example, medication guides in the medical industry, company product manuals, product function documents, etc., are all materials for the knowledge base and should be sorted out as comprehensively as possible."
These alone are not enough, and these materials need to be manually processed into a "question-answer pair" form, such as to classify questions, to sort out questions and standard answers to similar or related questions, and upload as many question-answer pairs to the system as possible.
Customers have thousands of faces. Once the questions asked exceed the coverage of the correct answer, the intelligent robot will appear "mentally retarded" performance of getting stuck or answering wrong questions. Therefore, during the operation process, enterprise personnel have to collect new questions in a timely manner, and then manually add or optimize answers.
"The workload is very heavy." Wu Linan took out a set of data. When a certain company expands and edits 180 standard questions with similar or related questions, it can add 1,800 questions and answers. "This requires at least 3.5-4 days of manual work. .”
When AIGC models such as GPT participate in the bottom layer of intelligent customer service, changes have occurred.
Enterprises can directly upload various document materials to the intelligent customer service system for backup. Some standard content does not need to be manually split into question-and-answer pairs. The material will also be summarized, summarized, and reorganized to form an answer that conforms to the language logic and contains the core information.
Wu Linan said that from the large amount of customer test data they have obtained, it can be clearly seen that the direct answer rate and answer accuracy rate have improved, "and the time spent manually on the 'question and answer' link has been shortened to half a day, because they only You need to mark whether the answer is accurate or not, and add unreserved questions.”
Empowerment rather than complete subversion
In addition to wisdom tooth technology, a large number of smart customer service companies are also on the road to +AIGC.
According to a recent report by the consulting firm IDC, Tencent Qidian "will integrate the AIGC capabilities of Tencent's large model to achieve linkage with marketing automation scenarios"; on April 21, Udesk released the original heart engine 2.0 based on the GPT general model Provide a unified customer service platform with intelligent support; also in April, Kuaishangtong announced that it will launch AIGC products in the vertical medical industry based on ChatGPT technology, and comprehensively update product functions and enterprise service models; another intelligent customer service company Meichao also "officially launched Enter the "ChatGPT" era.
For all intelligent customer service companies + AIGC, efficiency improvement is unavoidable and the most practical direction for enterprise services.
The customer service industry has experienced 30 years of changes in China, and has gone through a path of technological change combining communications, big data, and artificial intelligence. One theme is to improve efficiency. From telephone customer service 30 years ago, to online customer service in the mobile Internet era, to today's intelligent customer service robots, the link between enterprises and consumers has been shortened from door-to-door service and business halls to only a PC screen and a A mobile phone can solve most of the problems.
Businesses and consumers have also tasted the sweetness of cost savings in this process, behind which is the technological upgrade of the customer service industry.
In the past ten years, the development of mobile Internet, big data, and AI technology has gradually transformed the traditional customer service call center into an intelligent SaaS model. Through the cloud server, the customer service system has completed the comprehensive integration of multiple customer service channels such as phone calls, emails, and text messages, and the artificial intelligence natural language model (NLP) has also begun to be applied to the intelligent customer service track.
It was also during this period that a large number of capital and entrepreneurs poured in, and a large number of intelligent customer service companies such as Wisdom Teeth Technology, Meicha, Udesk, and Ronglianyun were born, basically realizing the ability of intelligent customer service robots to respond to questions from consumers and customers. Answer independently, and even realize complex operations such as business processing.
In the past ten years, Chinese people have gradually become accustomed to the mode of being handled by customer service robots in various scenarios. Of course, you can still feel that it is robots who serve you. When they are uncertain, human customer service will appear.
When the AIGC that can speak, is good at drawing, and specializes in generating AIGC appears, will the machine-like feeling of customer service disappear? Can AI really completely make human customer service off work?
In Wu Linan's view, "AI is still empowering the customer service industry, but it has not yet reached the level of subversion." The reason is that the changes AI brings to intelligent customer service mainly occur in the improvement of operation efficiency. The most intuitive way is to answer directly. increase in rate and answer accuracy.
Wisdom tooth technology product application process
In addition, when evaluating customer service efficiency, customer service managers need to analyze service quality, organize reports and even analyze changes in user (buyer) emotions during the customer service process. The ability can come in handy again, this function has no problem in terms of technical feasibility, and we have also commercialized it.”
It can be seen that even if the intelligence of the customer service system has reached a new level, people are still an important part of customer service, and they must calibrate, supplement and even evaluate the service quality of the answers generated by the current AIGC. What AI replaces is the most basic and complicated work of sorting out knowledge bases and pre-set questions and answers in the customer service link; machines can provide data to analyze customer emotions and judge consumers' emotional changes, but how to improve after the data comes out still requires the final analysis. People who understand humans.
Xun Pengjun, vice president of Wisdom Tooth Technology, also mentioned the new problems brought about by the application of large models such as GPT to the intelligent customer service industry. The reply time is longer; in addition, another common problem of the general large model is that too much corpus is used in training, and the customer service robot is likely to have the company’s competing product information when replying to the question; there is another thing that needs to be paid attention to is data security , Enterprises must ensure that sensitive information is removed during data transmission.”
The existence of these problems also makes AIGC large models such as GPT, for intelligent customer service systems, only an empowerment rather than a complete subversion. This is also the reason why Wisdom Teeth Technology finally chose the more challenging "+AIGC" instead of building an AIGC-style smart customer service "from 0 to 1".
At present, the bottom layer of the intelligent customer service system of Wisdom Teeth Technology is not only guaranteed by the Sobot engine that has been continuously iterated for 9 years, but also combined with the GPT model. , for the customer service industry, although AIGC has not yet brought about a subversive revolution, the potential for transformation and upgrading is huge. , such as facilitating conversions and closing transactions."
At least for now, AI customer service robots have not completely replaced human customer service. People still can’t help but press the “transfer to manual” or “0” button when the robot gives a message in the dialog box or the phone receiver. , you are too eager to hear direct and understandable answers from human customer service. Replacement may actually happen when you no longer perceive that the other person is a robot.