Dai Zonghong, former co-founder of Lingyiwanwu, starts a new business: A 20-person team can handle AI customization projects equivalent to those of a hundred-person team.

Text | Zhou Xinyu

Editor | Su Jianxun

The domestic market, ToB, and customization - each of these terms hits the pain points of current AI startups.

However, this is exactly what Dai Zonghong, the former co - founder of Lingyiwanwu, is doing now.

On March 24, 2025, after leaving Lingyiwanwu, he founded a new company called "Jidian Qiyuan". Nevertheless, when entering the B - end customization market at present, what Dai Zonghong wants to tell is a completely different new story from traditional B - end customization.

The manifestation of this story in Jidian Qiyuan's business is: With 7 or 8 customization projects running in parallel, the size of Jidian Qiyuan's engineering team remains at around 20 people, and they have never missed a delivery deadline.

The key lies in using AI to automatically execute the "dirty and tiring work" in the customization process.

Traditional customization requires AI companies to form an expert team of dozens of people. Through pre - project manual interviews and data collection, and then manual data analysis and modeling, finally, the business process is replicated in a digital form - this process is equivalent to creating a digital copy of the real business manually.

Based on the "copy" of the business process, enterprises can conduct solution deduction according to their own set business goals and finally find the best path.

What Jidian Qiyuan wants to do is to hand over the processes of early expert visits, data collection, analysis, modeling, and depicting the business process to AI.

To this end, Jidian Qiyuan has built an AI operating system that can automatically depict the enterprise business process:

• The foundation of the system consists of multiple base models and large vertical industry models. Based on various types of original business data from enterprises, the underlying models can automatically sort out and understand the production element nodes of the business.

• The built - in production element toolchain of the system can model each production element one by one to form business nodes.

• The RL toolchain in the system, based on the digital twin of the real business process, builds AI software such as large enterprise models for business customization through reinforcement learning.

After adopting the AI operating system, Dai Zonghong told us that it only takes one day for enterprise data collection and governance, plus the construction of the enterprise business process, "After our customers review it, there are no errors."

Now, this AI operating system has been put into operation in enterprises from different industries such as steel smelting, environmental governance, and new energy management, helping enterprises achieve goals such as cost reduction, efficiency improvement, and supply chain optimization.

For example, when an enterprise's business goal is to optimize the supply chain cost by 15%, they only need to connect their supply chain data API to the AI operating system, and they can get their business model in a short time. After setting the 15% optimization goal, the business model will provide a path to achieve the goal based on business knowledge.

Recently, Jidian Qiyuan completed its angel - round financing, with the financing amount exceeding 100 million RMB. Investors include Sinovation Ventures, Nut Capital, Nine - Fusions Ventures, Puhua Capital, Yinqiao Fund, Zhengyang Hengzhuo, Zhongke Chuangxing, Zhuichuang Ventures, etc. (listed in alphabetical order).

Of course, this entrepreneurial journey is accompanied by many doubts - Dai Zonghong told Intelligent Emergence that during the financing process, some investors' concern was that Chinese B - end customers have a very low willingness to pay.

Indeed, since the upsurge of large models, the confidence of domestic AI companies in industry implementation has quickly been shattered. The long payment cycle and limited customer willingness to pay are long - standing problems.

Not doing customization but focusing on standardized products has gradually become a consensus. Robin Li, the founder and CEO of Baidu, once said bluntly: In B - end business, try to avoid thankless project - based work and try to launch standardized products.

As an experienced player in AI ToB, Dai Zonghong has a different view on this assertion. Before this, he served as the director of AI Infra at Alibaba DAMO Academy and later became the AI CTO at Huawei Cloud.

During his time at Huawei Cloud, he was involved in hundreds of AI customization projects. This made him realize that "it's not that the domestic customers have a low willingness to pay, but that their payment preferences are different from those overseas. Overseas enterprises are used to paying for tools, while domestic enterprises are used to paying for results."

This means that there is still a huge demand for AI customization in the domestic B - end market.

△ Dai Zonghong. Image source: Provided by the enterprise

 

After starting his business, Dai Zonghong spends nearly 40% of his time on customer visits. He found that these customers are not willing to use the standardized products on the market because they cannot be directly integrated with their existing businesses. "They still prefer to pay for end - to - end customized solutions and are willing to offer a high budget."

Rather than being anxious about competition and financing pressure, Dai Zonghong is more worried that without more players entering the market, customers' confidence in B - end customization will decline.

During the conversations with customers, he realized that the unsatisfactory results of previous AI customizations have damaged some customers' confidence in AI technology and directly affected their AI budgets. "Most AI companies in the past only deployed a large model for enterprises, but what customers want is not the large model but the results."

Therefore, in the customer acquisition stage, Jidian Qiyuan is not in a hurry to directly let enterprises deploy its AI operating system. Instead, based on the customers' business goals, it directly delivers a set of solutions run by the AI operating system.

Dai Zonghong summarized to Intelligent Emergence that overall, competition is a good thing for the B - end market. "Once there are more successful cases, customers' confidence will increase, and it will be easier for us to acquire customers."

The following is a dialogue between Intelligent Emergence and Dai Zonghong, edited for clarity:

Large language models have not shown me opportunities for widespread industry implementation

Intelligent Emergence: How would you describe Jidian Qiyuan's business?

Dai Zonghong: I hope to provide a technical platform called an "AI operating system". Based on this technical platform, it can bridge the gap between the industry and AI, enabling the core business links of the industry to quickly leverage existing AI tools and means to complete the AI transformation of the entire industry and ultimately achieve the improvement of business value.

Intelligent Emergence: It's a bit abstract. Can you give an example?

Dai Zonghong: For example, from the original data of various systems of a certain customer, our model and platform can automatically learn this set of data. This learning process is completely automated.

Using a few machines, we can automatically write out the entire business process of the enterprise in about a day and a half.

In the traditional customization field, if you want to depict an enterprise's business, it usually requires interviews with various levels of personnel, technical interviews, and data sorting and governance. By the time the results are out, half a year to a year has passed.

But now we can clearly sort out the entire business in one day. And after a detailed review by the customer, there are no errors.

Intelligent Emergence: Summarize the differences from traditional customization?

Dai Zonghong: Previously, almost all enterprises mainly focused on customization and manual model building when serving the industry. They couldn't avoid the problems of data governance and business process learning.

But now our technology can leverage the complex data of enterprises to independently learn and understand these things, and then reproduce the enterprise's workflow. This is a set of technical systems for unsupervised data governance that we have developed.

Intelligent Emergence: What are the benefits of this set of technical systems for customization?

Dai Zonghong: First, it can be generalized. Second, it can penetrate into the core business links, not just as an office assistant. Third, it is actually related to the first point, which is that it can form a scale, rather than just helping one or two leading enterprises complete the transformation.

Intelligent Emergence: In 2023, during the wave of large models, you became a co - founder of Lingyiwanwu and were among the first batch of people to engage in large - model entrepreneurship. Did you see opportunities for AI customization at that time?

Dai Zonghong: From ChatGPT itself, I couldn't see opportunities to empower all industries. Its more prominent capabilities are language and logical abilities. Only when the model has in - depth reasoning and thinking abilities can it understand the business process of enterprises and assist in decision - making.

It wasn't until the release of o1 that I thought the implementation of large models in all industries had become a matter of optimization degree rather than feasibility.

In fact, when I was at Lingyiwanwu, we also tried to do work similar to ReAct (Reason + Act, a decision - making mechanism for building Agents) and achieved certain results.

Although it wasn't as amazing as o1, we still made great progress. This was the result we achieved on our own without the guidance of o1 and o3 because we had been exploring the underlying technology at this level.

Intelligent Emergence: Why did you choose to leave Huawei and engage in the large - model field with Lingyi in 2023?

Dai Zonghong: At Huawei, I was able to participate in and even lead some work on using AI to empower all industries. I had many practical opportunities and was exposed to a large number of enterprises and scenarios. So this experience was indispensable for me, and some prototypes of my current entrepreneurship were also formed at Huawei.

However, at Huawei, it was difficult for me to delve into the technical details of large models. At that time, Lingyi already had a preliminary technical foundation and a certain amount of capital. Finally, I could clearly see that I could do a lot there. So after talking with Kaifu and Xuemei, I joined.

At Lingyi, I systematically observed and learned a series of large - model technologies. Previously, I was only an AI assistant, but the experience of being deeply involved in every indicator at Lingyi was very different.

Intelligent Emergence: Which technological developments became the impetus for your entrepreneurship?

Dai Zonghong: In fact, I'm not a technical expert in large models. The reason I started working on large models is that I think they can help me achieve the above three points.

I think in - depth Reasoning ability is a key node, which was when o1 and o3 (OpenAI's reasoning models) were released in 2024. The reason I launched the "Jidian Qiyuan" project at this time is also because large models had reached this level.

I think I don't need to do a lot of work on large - model training myself. I can just stand on the shoulders of others.

A 20 - person team has completed the work of hundreds of people in traditional AI companies

Intelligent Emergence: What is the structure of the "AI operating system" you are currently working on?

Dai Zonghong: It's a platform that integrates models and applications. The bottom layer of our system is a group of large models, including basic large models and large vertical industry models, which serve as tools for understanding business data and building models.

The upper layer of the system uses the enterprise's own business data for modeling and finally forms a digital twin environment that combines the virtual and the real. The models built are free of hallucinations because they are based on the enterprise's business data, so they understand the enterprise's business chain.

Intelligent Emergence: What is the method to achieve "using AI to depict the complete business process"?

Dai Zonghong: An enterprise's data is very complex, with different sources and storage formats. The first thing we need to do is to automatically and deeply mine this data, including exploring its potential manifestations, and be able to conduct automatic data analysis.

Then we need to mine the so - called "complete set of production elements" based on the existing data and then build the business based on these production elements.

Next, we need to automatically model each node. Automatic modeling also relies on data. For some nodes without natural data, we need to automatically supplement the data.

In short, it is to present an enterprise's real business process in a modeling way.

It sounds like the whole process is very simple, but it is very complex to achieve automation. Because there are tens of thousands or even millions of nodes in some enterprises, and it is quite challenging to automatically learn all the relationships between these nodes, including development and logical relationships.

Intelligent Emergence: How do you ensure that the supplemented data conforms to the real business logic?

Dai Zonghong: We don't need it to be precise; we only need it to be correct. Because later we will actively improve the accuracy of the model through reinforcement learning.

In essence, in the AI 1.0 era, most of the work people did was to improve the accuracy of the model through labeled data. This required a large amount of manual work, and the uncertainty and instability were very high.

Now we no longer use manually labeled data. We need to form a natural reinforcement - learning environment through some ingenious designs, and continuously generate data through the actual production environment to strengthen our model.

In essence, this process can achieve the same effect as manual data labeling.

Intelligent Emergence: Why use AI to model an enterprise's business process? What is the value for enterprise customers?

Dai Zonghong: Enterprises can use the depicted business process to conduct business deduction and decision - making.

You can regard the replicated business process as a digital conversion environment and a mirror image of the real business process. It is relatively easy to conduct various deductions in this digital environment.

Deductions based on the digital environment will also bring many possibilities for the realization of enterprise business. For example, in addition to helping enterprises deduce the effects of different decisions, it can also achieve cost reduction, efficiency improvement, production capacity improvement, production capacity balance, supply chain optimization, etc.

Intelligent Emergence: Does this process require a large amount of manpower like traditional customization?

Dai Zonghong: The degree of manual intervention is almost zero. It's like when a customer comes, as long as we get the permission for the corresponding data interface, the entire process of the model's learning and understanding is completely automated.

Intelligent Emergence: Does the final effect of this system achieve the same result as traditional manual customization?

Dai Zonghong: This system actually replaces not some internal functions of the enterprise but the experts of AI companies to help an enterprise make some scenarios intelligent.

According to the traditional customization process, we need to send a large number of experts to help the enterprise with corresponding analysis and modeling work. Now we use large models to replace these AI enterprise experts to help enterprises build all - element models.

If relying on people, 10 people can only build 10 models, and 20 people can only build 20 models. If there are 10,000 production elements in an enterprise that all need to be modeled, there will need to be 10,000 models and 10,000 people.

But with this system, we now rely purely on computing power rather than manpower.

Intelligent Emergence: Many ToB customization enterprises eventually end up with a bloated staff size. How many people are there in Jidian Qiyuan?

Dai Zonghong: We are currently advancing seven or eight projects simultaneously, but the total team in charge is only 20 people. Previously, it might have required hundreds of people.