完成B+轮融资，收入过亿的FancyTech说：每个阶段都需要再找一次PMF丨涌现36人

Source: Intelligent Emergence

Emergence is a key phenomenon in the generative AI wave: When the model scale expands to a critical point, AI will demonstrate human-like intelligence, being able to understand, learn, and even create.

Emergence also occurs in the real world - the silicon-based civilization is on the verge of emerging. Entrepreneurs and creators in the AI field are using their wisdom and minds to light up the long journey to achieving AGI.

At the juncture of the alternation of old and new productive forces, "Intelligent Emergence" launches a new column "Emergence 36 People". We will record the new thoughts of this stage through conversations with key figures in the industry.

Text | Deng Yongyi

Editor | Su Jianxun

As the boss of an AI startup company that has been established for four years and just caught up with the large model wave, Kongjie, the CEO of FancyTech, faced a crucial decision last year:

Should more human employees be added to the AI project?

FancyTech is an AIGC content platform. Using its self-built image and video models, FancyTech can help KA brand customers in the fast-moving consumer goods and retail industries generate and produce product images, videos, and other materials with AI.

This company first became well-known when Zhu Xiaohu praised it in 2024 as a model of an invested enterprise in finding PMF (Product Market Fit), saying, "The product effect is particularly good and can be monetized immediately."

△ FancyTech customer backend 

△ FancyTech operation interface 

In terms of business model, FancyTech does a standard To B business. This means that unlike ordinary AIGC tools that write poems and draw pictures, FancyTech takes the real money of customers and needs to deliver more satisfactory content results for customers. And "satisfaction" is a very personal standard.

"It is difficult for pure AI to 100% meet the requirements of customers. Even if you can achieve 70% or 80% with AI, customers will not be satisfied and will not be willing to pay for it," Kongjie told "Intelligent Emergence".

For example, FancyTech disassembles the materials in a product video according to different parts of the product and assigns them to different studios to complete. For example, Studio A is good at character movements, and Studio B is responsible for AI changing clothes, etc.

But for brand customers, even if each detailed item can be 30% or 50% more efficient with AI, when these "detailed items" are combined, the brand still needs to send its own people to complete the final 20%. This will greatly reduce the customer's order success rate.

Therefore, FancyTech chose to have human employees coordinate the entire process, which comes from Kongjie's work experience.

Before founding FancyTech, Kongjie was the first-generation head of Tmall Luxury Pavilion and the product operation head of Mobile Taobao. "We did not let the brand and the content producer directly connect. Everything is guaranteed by FancyTech for project delivery. From the customer's perspective, we are equivalent to the content supply side of AIGC, directly delivering results, so that more merchants are willing to cooperate with us."

For the venture capital circle, the "AI + human" model sounds not very attractive, but FancyTech still withstood external pressure, reduced the team size, began to increase the proportion of human employees in the delivery, expanded high-end brand customers, and introduced more third-party creators into the ecosystem.

By the second half of 2024, FancyTech has established a model similar to a "front store and back MCN":

In the front, Fancy meets the brand material needs of brands and merchants. The account manager disassembles the needs and subcontracts them to multiple creators / studios on the FancyTech platform;

Creators complete content creation based on the FancyTech technology platform. During the process, FancyTech provides a variety of material creation capabilities in the e-commerce field, including the underlying large model and the vertical model of each element;

FancyTech finally completes the project coordination and the connection with the customer to ensure the project effect of the delivery.

From the perspective of commercialization figures, FancyTech's reform has achieved good results.

In 2022, FancyTech achieved a revenue of tens of millions of yuan, and in 2023, it exceeded 50 million yuan. In 2024, this number doubled again. FancyTech has also expanded this model to more than 10 countries, including the United States, South Korea, Singapore, Malaysia, Thailand, Indonesia, the United Arab Emirates, Saudi Arabia, Brazil, etc.

"Intelligent Emergence" has learned that FancyTech has recently officially completed a tens of millions of US dollars B + round of financing, led by Zhilin Capital and the old shareholder GSR Ventures, and LightSource Capital served as the exclusive financial advisor.

For more AI startup companies, the pressure from the market is becoming more and more realistic and urgent: In the gap of the rapid iteration of the base large model, how to find PMF and achieve self-hematopoiesis?

FancyTech was founded in a year when VC investment entered a low tide. Therefore, from the first day of its establishment, FancyTech pays special attention to cash flow and running through the business closed loop.

Kongjie said that the idea of previously being a pure AI-driven company "may be a detour". Focusing only on a single-point technological breakthrough is difficult to establish a closed loop in business. Startup companies need to find PMF (Product Market Fit) again at each stage.

"No matter how deep you do it, a large factory can immediately have the same products and effects as you. There will always be someone who raises more money than you and then subverts you," he said.

Now, working hard to win over one KA customer after another has become Kongjie's daily routine. When this round of financing was completed, he also intensively shared the good news of signing contracts with customers such as Yuzhe, NINE WEST, and United Overseas Bank (UOB) in Singapore on his WeChat Moments.

Kongjie also shared with "Intelligent Emergence" the practical experience from domestic to overseas in recent years, including many attempts and abandonments. We have sorted out these shares:

It is not feasible to only make pure AI tools, as they will be overturned by large models

"Intelligent Emergence": Your company was established in 2020, just before the arrival of the large model wave. Your products have undergone many changes in these years. How would you summarize what FancyTech does now?

Kongjie: The entire industry has experienced three stages of changes. At the earliest stage, everyone was working on large models; in the second stage, they turned to developing AI applications, either training small models by themselves or directly connecting to other people's interfaces. However, the commercial prospects of these methods are relatively limited.

Now we have explored the third step, which is platformization. We have established a commercial platform for AI content, aggregating a large number of AI creators and studios at home and abroad to make deliveries, and we jointly serve customers.

"Intelligent Emergence": How to understand this AI content platform?

Kongjie: We have cooperation with many AI studios and individual creators at home and abroad.

They may be small studios or AI enthusiasts who will do some small model training, including Workflows, Lora, etc. They are very focused on their respective vertical fields. For example, some specialize in making AI images of babies, and some focus on jewelry display, being able to control the details of how to wear a ring on the hand very well. We integrate these excellent creators into the platform.

"Intelligent Emergence": What is the specific service process like?

Kongjie: We have built an editable platform. Different from ComfyUI, which can only be used by professionals, our platform is more general and easy to use. Creators can upload Workflows, Lora fine-tuning or Fine-tuning models based on open source.

We have accumulated a lot of experience in the To B field, because many large model companies, including overseas companies, prioritize doing To C. We have accumulated professional knowledge, working methods, and basic materials in various industries.

From China to overseas, we have found that the best practical way in To B is to directly deliver results rather than providing tools. Our customer orders are relatively high, and many studios will directly deliver results to customers based on our platform.

"Intelligent Emergence": What part of the customer's work is replaced?

Kongjie: Previously, customers needed to find agencies and shooting companies, do lighting, actual shooting, hire models, and edit and trim after shooting. Now these links are no longer needed.

"Intelligent Emergence": Where is the contribution of AI in the entire process?

Kongjie: In every production link, AI can improve efficiency by 30% to 50%. The accumulation of these efficiency improvements can ultimately bring satisfactory results to customers. If it is only a pure AI function, it may only solve 50% to 70% of the customer's needs.

Businesses will think that although you have improved efficiency here, the overall result is still not good, and I still have to do it myself. This is why AI cannot be implemented now.

"Intelligent Emergence": Can you give an example?

Kongjie: For example, we have a real estate customer in the Middle East. The real estate market there is very hot, and they start pre-selling before the house is built. Previously, there were only blueprints, but now they need to produce content including the exterior and interior renderings of the house, furniture and living scenes, etc.

The first step is for Fancy's local business team in the overseas market to negotiate with the customer. The second step is to disassemble the entire project into different parts through our platform, some for making pictures, some for making videos, and some for making dynamic effects. The third step is that we will find producers who are good at the corresponding fields on the platform and assign the disassembled tasks to them.

In order to improve their efficiency and competitiveness, we will provide various Workflows and suitable models on the platform to enable them to produce more quickly and effectively. These Workflows and models come from the contributions of other creators on the platform.

"Intelligent Emergence": You are like an MCN.

Kongjie: For example, compare Disney and SHEIN. Disney has popular IPs every quarter, but these IPs are basically not made by themselves. Instead, they find content that is relatively popular in a small range, buy it, amplify the content through movies, test peripheral products through Disneyland, and finally amplify sales through global agents. They earn the difference between the final large-scale sales and the initial purchase of the IP.

SHEIN discovers fashion trends overseas, connects small factories in Guangzhou to make sample clothes, tests the market response on the website, and expands production for the popular ones. Previously, these small factories could not receive international clothing original design orders, but this can be achieved through SHEIN's mechanism.

Similarly, previously, small content producers in China could not receive content production requests from large overseas customers, but this can be achieved through our AI content platform. Previously, it was the production supply chain going overseas, and now it is the content supply chain going overseas, greatly improving efficiency through AI.

"Intelligent Emergence": It sounds as if it has not much to do with the development of AI technology. In fact, it is an iteration of business thinking. For example, in your workflow, humans are still a very key element.

Kongjie: Although this AI wave is very hot, very few can be truly commercially implemented. Previously, everyone hoped that just like the past SaaS, making a pure AI tool could be used, but AI is difficult to solve 100% of the problems. Even AI Agent is only to assist humans.

After countless attempts, we found that only a small amount of manual work plus a large amount of AI to complete the delivery can satisfy customers. Therefore, we must be platformized, and from the training of AI to the people who actually deliver, there must be various external personnel to join.

"Intelligent Emergence": If it is open to third parties and an ecosystem is established, how to manage the overall workflow?

Kongjie: To ensure the result of this delivery, our staff is responsible for connecting with the customer's sales, and the project management is completed by Fancy through review and adjustment.

The advantage of this is that with the development of AI technology, our efficiency will become higher and higher. It may be reduced from, for example, 20% manual work to 15% or 10% now.

Nowadays, startup companies are very worried every day. If GPT-4 or GPT-5 comes out, or OpenAI and ByteDance make an Agent today, will their efforts be in vain?

"Intelligent Emergence": The best practical way is to directly deliver. When did you realize this?

Kongjie: It was around the second half of 2024.

"Intelligent Emergence": What happened in the middle, or how was it explored?

Kongjie: It is a gradual process. At the earliest, we achieved rapid commercial implementation. Later, we found that doing everything by ourselves was too heavy, and external roles must be added.

During the process, we also received a lot of criticism. Others said that you are not a pure AI, and why do you have so many manual workers? We don't care.

Previously, we also trained our own models, and the effect was good, but it was very vertical. It may only be able to do a certain business in e-commerce, and this cannot be further expanded in business.

After expanding the overseas market, we found that the demand for AI content varies in style in each country and region. Therefore, third parties, such as individuals or small studios, must be added. Everyone can train together based on the open source model.

Finding PMF at each stage

"Intelligent Emergence": In the past two years, the progress of the large model layer has been very rapid. How does FancyTech decide whether to train its own model or consider the issue of input-output ratio?

Kongjie: When the large model first came out, we did not train our own model. We were more focused on studying how to use it.

From about the second half of 2022 to 2023, we began to train our own image and video models. We also reaped the benefits in the early stage because at that time, if you had a model, it would be different. We also seized our own characteristics. Because we do To B, our effect in physical restoration and control is relatively good.

Starting from the second half of 2024, we found that only our own model is not enough to meet all customer needs, and the problems it can solve are limited. Therefore, we began to weaken our own model training and leave it to the creators on the platform. They may only train one part of it respectively.

For example, some creators are responsible for specializing in the control of fingers, and some can make the effects of hair, clothes, and shoes very well.

"Intelligent Emergence"