Hard Krypton Exclusive | Backed by three rounds of investment from Li Zexiang, a PhD graduate from Zhejiang University has developed the world's first vision-based after-sales technical customer service robot

According to Hard Krypton, visual model enterprise MicroLink Intelligence has recently completed an angel round financing of nearly 10 million RMB, with investors being Li Zexiang and Lu Qi. The investment entities are Dongguan Clear Water Bay Phase II Venture Capital Partnership (Limited Partnership), Ningbo Institute of Intelligent Technology Co., Ltd., and Beijing Qichuang Chuangtan Phase II Venture Capital Center (Limited Partnership). This round of financing will be mainly invested in research and development.

Founded in 2023, MicroLink Intelligence's founder Ning Dongdong is 31 years old this year. He pursued his doctoral degree at Zhejiang University, engaged in machine vision research, and is one of the early practitioners of deep learning-based machine vision in China, with rich experience in image processing.

Ning Dongdong told Hard Krypton that he originally followed Li Zexiang to engage in product R&D in the image field, and found in the investigation that the customer service of the consumer electronics industry was very primitive. However, Ning Dongdong did not act rashly. In order to fully understand the e-commerce customer service chain, he investigated three different entities: e-commerce platforms, merchants on the platforms, and tool vendors selling AI customer service.

Ning Dongdong's first stop was to work as a customer service intern at JD for 3 months. He found that JD's internal system was very perfect. "The logistics, finance and e-commerce departments are fully connected, and customer service can directly modify the detail page of self-operated stores or adjust the logistics address."

Later, he went to work as a pre-sales customer service for an electrical appliance merchant on the platform. The situation here is different, which is completely in the manual stage. The platform hardly provides any tools, and you have to make phone calls to contact the logistics or maintenance personnel, leading to low efficiency. "JD's customer service can solve 300 complex after-sales problems a day, while merchant customer service can solve less than 80 a day."

However, the two internship positions have common features: low income, about 3000 RMB, and high staff turnover. "For example, when I was interning at JD, there were nearly 100 people in my cohort, and only about 5 people remained a year later." In addition, the shifts are day and night reversed, which is very harmful to health, and the workload is very heavy, requiring high psychological quality of employees.

The entire e-commerce industry has more than 5 million manual customer service staff. Ning Dongdong introduced that during big promotion events, the platform requires messages to be replied within 30 seconds. Previous AI technology enables robots to reply to some simple questions, but this only accounts for less than 20% of consultations, and manual review is still required. "Some companies even need to add a post called AI customer service trainer to configure question-and-answer pairs in order to use robots, which costs more than 10,000 RMB a month."

Ning Dongdong analyzed that there are several serious problems in the whole industry: first, long training cycle. The customer service staff has high mobility, most of the new recruits have no experience, and their training takes 1-2 weeks or even a month. Second, high labor cost. At least 4 people are needed from 8 a.m. to 12 p.m. "A mid-sized enterprise with annual sales of about 500 million RMB needs 50 customer service staff. With social security, management and site costs, the average amortized cost per person is nearly 10,000 RMB per month." Third, great brand after-sales risk. New customer service staff are likely to cause brand after-sales problems, "A post on Xiaohongshu may destroy a brand."

Starting from these problems, MicroLink Intelligence has developed the industry's first "technical customer service" dialogue robot specially for e-commerce scenarios, focusing on solving the problems of high technical threshold, difficult fault diagnosis and high labor cost in the after-sales link of consumer electronics and complex hardware products.

Specifically, its product can achieve several points: first, it breaks through the limitation of traditional customer service robots that can only process text, has strong image and video understanding ability, and can even interpret complex electronic product technical drawings and structural diagrams to quickly locate hardware or usage problems. Second, it can think independently, actively diagnose and ask follow-up questions, and has deep logical reasoning ability. When faced with users' difficult descriptions, it can actively ask users questions and guide troubleshooting like technical experts, realizing fully autonomous Q&A and independent reception in complex after-sales scenarios without frequent transfer to manual work. Third, it supports multi-channel integration and aggregated operations across mainstream e-commerce platforms, seamlessly embeds into the existing workflow of merchants, and automatically processes after-sales technical consultation.

In terms of commercial implementation, the company is backed by Li Zexiang's consumer electronics ecological chain enterprises, and has a high-quality customer base of more than 270 consumer electronics brand enterprises. Ning Dongdong said that in the long run, there will be two ways: directly selling tools and selling services. "If we sell tools, the software license fee for a company with 500 million RMB revenue is only 50,000 to 100,000 RMB, and it will take too long to achieve 100 million RMB revenue. But if we take over his customer service and directly deliver results, the 2 million RMB annual team expense he originally paid will be paid to us instead, the customer unit price will change from 50,000 RMB to 1 million RMB, and we can also help customers save 1 million RMB."

The following is an excerpt of Hard Krypton's interview with Ning Dongdong:

Hard Krypton: How should we understand the "fully autonomous after-sales" you mentioned? What are its advantages?

Ning Dongdong: Most of the traditional customer service staff are new employees who have been on the job for 1-2 months, with slow response, easy to answer wrong knowledge points, and the post-2000s customer service staff have great emotional pressure, and their replies tend to be perfunctory. However, AI customer service has the complete knowledge base of senior employees, responds quickly, the knowledge base has strict update logic and will not answer incorrectly, and can always maintain enthusiasm.

The so-called "full autonomy" means that from the moment the user enters the session to leaving, the whole process is received by AI without any human participation in the middle, unlike other manufacturers that require humans to review all the dialogues next to them.

Hard Krypton: From a technical perspective, how do you achieve fully autonomous takeover?

Ning Dongdong: The agents of other competitors only call the model once, ask a question and reply with a message, and have no multi-modal ability to understand images and videos. But users rely heavily on images and videos in after-sales scenarios, which requires multiple rounds of follow-up questions and answers to solve problems.

In the previous customer service difficulty pyramid, pre-sales is the easiest, such as checking parameters, after-sales is the second, such as returning and exchanging goods. Technical after-sales is the most difficult, because it requires troubleshooting and solving heavy usage problems. Most founders in the industry come from e-commerce or NLP backgrounds, lack of vision background, and are all crowded in the pre-sales red ocean.

But our team is from the machine vision field, and the difficulty is reversed for us: technical customer service is like doing math problems, there is not much complex emotion, only fixed solutions, whether it is return, troubleshooting or teaching operation. Technical customer service and ordinary after-sales are areas that consume a lot of labor, and it is also a "no-man's land" that no one has developed. We can just use our multi-modal capabilities and image processing experience to cut in.

Hard Krypton: The technical after-sales of different industries vary greatly. How does your product achieve universality? What about the image and video recognition capabilities?

Ning Dongdong: Our generalization ability is very strong. Although product knowledge is different, we can realize it through automatic database construction. The standardized part is the whole after-sales process: upload structural diagrams, usage materials and historical dialogues, and these three types of data can be imported to build a knowledge base. At present, we are mainly focused on the consumer electronics industry.

In terms of recognition ability, videos have more coherent semantics than images and are easier to understand. Our understanding of videos and images is similar to that of humans after testing. We can see clearly what humans can see; if the picture is not taken clearly, the AI will prompt the user to aim at the correct position and take it again. At present, we can solve 80% of after-sales technical customer service problems in experiments.

This system has very high technical barriers. First of all, the founder needs to have a deep understanding of images, and the whole system is completely different from pure dialogue systems. Secondly, it is necessary to combine self-developed models with the multi-modal models on the market. Finally, it is necessary to understand the user's product structure and local part drawings. For newly invented consumer electronics products that the large model has never seen, how to combine large and small models and architectures to make it understand the local part drawings is very difficult.