Actual Records of Pitfalls Encountered by Agents Going Global | Global Insights for Overseas Expansion
Author | Ma Shuye
Editor | Lu Zhen
Source | Jingxiang Studio
“I once spent five months developing a large model that no one used.”
In April last year, Liu Tianqiang was looking forward to winning a large order from a fast - fashion retail company in the United States and developing a B - end product for them that uses AI to generate product - on - body images. His team worked on it for a full five months. At the critical moment of product iteration, instead of the final signing notice, Liu Tianqiang received the news that the client company had been acquired and the project was aborted.
Liu Tianqiang is one of the entrepreneurs in the Agent overseas expansion wave in 2025. Most of these overseas - bound entrepreneurs have overseas study backgrounds and have grown up in domestic or overseas tech giants. In order to secure US dollar investments, they registered their companies in Singapore or the United States early on. The Agent products they launched are all in English, and their promotion is also targeted at overseas markets. They try to make themselves look like an ordinary overseas startup.
This complete “overseas - oriented” approach brings many benefits to startup teams. “Nineteen out of twenty companies are going overseas.” Liu Jiaming, the founder of CrePal.ai (AI Video Agent), summarized the current situation of domestic AI Agent teams in this way. For the same AI product, the overseas market can bring five times more revenue than the domestic market, which is due to the more mature payment awareness of overseas users. For AI Agent products where each output incurs token consumption and costs, it is almost inevitable for startups to directly develop overseas - version Agent products.
Manus, which moved overseas just three months after its popularity, is a representative example. Going overseas to “seek funds” is a consensus among startups. With domestic investment tightening and investors focusing on the team, Agent startups that can't attract top - notch technical talents and are unknown also need US dollar funds with deeper pockets.
However, when entrepreneurs who grew up in the East Asian cultural circle and were trained in domestic Internet giants try to “go overseas,” they hit a wall due to the completely different cultural environment, user consumption habits, and more complex policy constraints.
Liu Tianqiang never thought that in his first year of leaving Amazon to start a business, he would develop a large model that no one used. Different from the domestic culture that values efficiency, AI is too innovative. Foreign companies even take weeks to sign a confidentiality agreement for a contract. In the end, without time for emotional fluctuations, Liu Tianqiang quickly downsized his team, and everyone switched to an “unpaid” mode to continue iterating the product.
Liu Tianqiang described overseas AI entrepreneurship as “repairing the plane while flying.” Changes may happen at any moment. The overseas market is not just a place of “unrestricted freedom.” It is both a gold - mining field and a swamp. Going overseas sounds wonderful, but it is not as easy as expected. We talked to several overseas - bound Agent entrepreneurs to see what pitfalls they encountered and what experiences and lessons they gained.
01 Big Clients Are Not Saviors but the Beginning of a Nightmare
In 2011, Liu Tianqiang came from Harbin, Northeast China, to study at Boston University. He “played around” in his dormitory and developed a computer vision API. Later, this casual project was acquired by Amazon, and Liu Tianqiang joined Amazon, managing a team of hundreds of people.
In November 2023, with his expertise in the visual field and Amazon background, Liu Tianqiang started to develop an AI design Agent for the fashion retail industry. In April last year, a friend who had worked with him at Amazon approached Liu Tianqiang, hoping that he could develop a product that uses AI to generate product - on - body images. The client was a large fashion retailer in the United States. This retailer launched hundreds of new products every day and spent $5 million a year on shooting product - on - body images. If AI was introduced, it could save money and speed up the process dozens of times.
The opportunity was too tempting. If successful, Liu Tianqiang would not only get a multi - million - dollar order and high - quality clients but also achieve technological breakthroughs with the client's exclusive data.
Liu Tianqiang's team developed a model - on - body image Agent for the fast - fashion field.
Image source: Respondent
Liu Tianqiang first tested the popular open - source solutions in the market at that time, including OOTDDiffusion and IDM - VTON, but the results were terrible. Since the open - source models in the market were trained with public datasets, mostly front - view pictures of women's clothing, they could not meet the client's complex requirements, which involved green - screen backgrounds, men's and women's clothing not included in the public data. Just when Liu Tianqiang was hesitating whether to train a vertical model for the fashion field, the client offered to provide all the product images from 2018 to 2023 for training.
Next came rounds of testing. There were deformed fingers, disordered clothing textures, and distorted back views of characters. They had to start over again and again.
In September, just when Liu Tianqiang's large model was taking shape, the project was interrupted due to the acquisition of the client company. Liu Tianqiang spent five months developing an AI model that no one used.
This experience made Liu Tianqiang feel the differences between domestic and foreign companies. Domestic fast - fashion companies advocate using machines to replace manual labor. Even small overseas - bound e - commerce teams of a few people use ChatGPT to generate rough product - on - body images. However, overseas fast - fashion companies are still hesitant about AI. Besides the low acceptance, there is also a big difference in efficiency. It took American clients several weeks just to sign a confidentiality agreement, which is unimaginable in China.
In December last year, Liu Tianqiang got in touch with a middle - level manager of a well - known overseas - bound fast - fashion enterprise with the trained model. At that time, AI could not directly generate high - definition large - size pictures suitable for e - commerce display. To improve the cooperation experience, he also partnered with a domestic photo - editing company to handle the “last mile.” However, it was the photo - editing company that Liu Tianqiang invited to cooperate that finally won the order. Liu Tianqiang lost because he only had the model and did not adapt to the work process of fast - fashion companies.
These two failures gave Liu Tianqiang new insights. Compared with domestic fast - fashion companies that are more receptive to AI tools, overseas fast - fashion companies are more cautious. Besides improving efficiency, they value full adaptation to the work process more. Rather than needing an efficient AI tool, they hope to have a complete solution that integrates AI tools into every aspect of large - scale fast - fashion enterprises.
Following this idea, Liu Tianqiang's huhu.ai (design - type Agent) has finally been iterated to a stage where large enterprises can use it right away. Since April, Liu Tianqiang has been expanding his client base in Europe and the United States. Now he is cooperating with many large fast - fashion companies in Europe and the United States and has managed to survive with difficulty.
Liu Tianqiang, who chose the overseas market from the very beginning of his entrepreneurship, is still experiencing “spiritual overseas expansion” again and again. Compared with truly native overseas startups, these AI gold - diggers with dreams still have a lot to learn.
02 Ranked First on Product Hunt's Weekly List, but the Conversion Was Less Than 2000
Making their Agent noticed is the first hurdle that AI overseas - bound entrepreneurs need to overcome.
Liu Jiaming worked as an AI product manager at Tencent for several years and then jumped to Xinshi Universe, leading the development and implementation of the AI Agent creation platform MindOS in the North American market. Currently, Liu Jiaming is working on an AI Agent for video generation, CrePal.ai. One of the difficulties he encountered in the early stage of entrepreneurship was product promotion.
“In China, promoting Agent products means doing content marketing on social media. The main channels are Xiaohongshu, Video Accounts, and Douyin. But in foreign countries, especially in the United States, doing a good job in Google SEO (Search Engine Optimization) is very important.”
Google Chrome, which has a huge user base overseas, almost controls the fate of startups. Doing a good job in SEO, that is, making one's own website rank high when users search for certain keywords on Google, has become a compulsory course.
However, since Google uses more than 200 ranking factors and the weights of these factors are constantly changing, Liu Jiaming had to start learning from scratch. In the past two months, the Dr value of CrePal.ai on Google (a key indicator for Google to evaluate whether a website is of high quality) has increased from 7 to 22, tripling. Behind this is Liu Jiaming checking his business against Google's hot search terms every day, updating the pages. In the fast - paced and innovative AI field, he has to take his time in promotion.
Previously, Liu Jiaming developed an exploratory Agent product. To gain popularity, he placed his hopes on Product Hunt. This is a highly active community for technology practitioners in North America. Even OpenAI's o1 model chose to conduct its first - batch internal testing on Product Hunt.
In Liu Jiaming's imagination, as long as he got a good download ranking on Product Hunt, it would trigger discussions among core North American technology practitioners. However, the development of events poured cold water on him.
For Product Hunt, Liu Jiaming led his team to spend a month contacting various KOLs on overseas social networks, hired interns to canvass votes, and continuously monitored the ranking. Even though they ranked first on the weekly list, Liu Jiaming found that a month of careful preparation only resulted in less than 2000 downloads. “It was much less than our estimate,” Liu Jiaming said disappointedly.
“My understanding of overseas channels is too shallow.” Liu Jiaming later found that Product Hunt is more suitable for products that can be experienced on the web. At that time, his product was an app that only supported the iOS system.
In addition, promoting products through influencers overseas is far from as simple as in China. Compared with the simple contract spirit of “getting paid to do the job” in China, overseas KOLs need to be “convinced” by the product.
As a startup team, even if Liu Jiaming is willing to pay for promotion, overseas KOLs are more picky. “If your product design can't convince them or give them a sense of surprise, they may even refuse to cooperate.” However, once the product can really help the KOL and his fans, entrepreneurs will also receive in - depth free testing and enthusiastic promotion from KOLs.
Different from the situation in China where one only needs to consider how much to spend on promotion, overseas, entrepreneurs often need to explain to KOLs how the product functions will help their fans before product placement.
This unique “fan culture” in the overseas AI circle is also changing entrepreneurs' promotion strategies.
Liu Jiaming found that giving exclusive rights and interests to KOLs' fans during product internal testing has become the most cost - effective method. When fans can get exclusive discounts by clicking on the KOL's exclusive link, “I'm his fan, so I can get the discount.” Fans are happy, and this form of giving benefits to fans also motivates KOLs to promote the product more actively.
Liu Jiaming's team. Image source: Respondent
However, with the continuous emergence of new products, it is becoming more and more difficult to stand out overseas.
“The development of AI coding has greatly reduced the cost of product development. For consumers, there are more new products, and they can't keep up. Naturally, the cost of product promotion has also increased.” In the three years since Liu Jiaming went overseas, the overseas AI Agent product track has become more and more crowded. He has witnessed the growing number of fans of overseas KOLs and the rising promotion fees.
“Now, to make an Agent product as popular as Manus, a startup needs to prepare at least $2 million per quarter.” A person engaged in overseas promotion of AI products said that in the highly competitive overseas market, startups need to be well - funded to stand out.
As a small team with limited marketing budget, Liu Jiaming is more cautious when choosing markets. “Find some markets that are more suitable for your AI Agent, preferably high - ROI markets. After penetrating these markets, then expand to other countries.”
In addition to market screening, Liu Jiaming no longer pursues rapid large - scale expansion. The high promotion cost overseas prevents him from pursuing high - volume sales like in China. Instead, he contacts users who liked CrePal.ai from the early stage and have already earned income on YouTube through AI - generated videos.
03 The High Wall of Compliance Review Blocks AI Entrepreneurs
In addition to promotion costs, overseas - bound entrepreneurs have to face stricter compliance requirements than in China and incur higher compliance costs accordingly.
Chen Kaijie, the founder of Macaron ai, graduated from Duke University in the United States and developed MidReal (an AI interactive story platform). He registered his company in the United States early on and fully transformed it into an American company. Even so, Chen Kaijie still hit the review wall.
Before the emergence of Character.AI, Chen Kaijie also wanted to develop a similar AI application. In his initial idea, he noticed that most users never left their mobile phones, and AI needed more text data to improve the output effect. Therefore, if the app was allowed to read everything happening on the mobile phone in the background and let the AI analyze the user's life, the AI could post dynamic content at any time.
To Chen Kaijie's surprise, during the internal testing of the Agent, the product first sparked discussions among North American teenagers. In a school in North America, Chen Kaijie's product already had several hundred initial users.
This product, which requires users to open their life data to AI analysis at any time, was welcomed by the young generation who wanted to be different from their parents. “They think that there is nothing in their lives that can't be shown because they are all very real.”
However, this product, which had been polished for several months, was stuck by regulations when it was about to be launched.
Chen Kaijie and his team deduced technical details. Image source: Respondent
In Europe, where privacy regulations are strict, Chen Kaijie hit a wall first, so he had to choose not to launch the product in Europe for the time being. Later, the product also failed to get permission from the US Apple Store. Although AI - related products had not been launched on a large scale on the iOS system at that time and their scenarios were not as clear as those of meeting apps, it was difficult to estimate the impact of AI collecting user information. The platform still had strict requirements.
Chen Kaijie's team tried their best to fight for it, but then they found that the AI's information collection in the background also identified voiceprints, which belong to biological information and require higher permissions. Even if a third - party was willing to provide permission, the app needed to ask the user for confirmation to run in the background every 10 minutes, which was a fatal blow to the user experience.
“Technical implementation is just the first step. Compliance is the biggest challenge.” Chen Kaijie found that the data - collection methods that could get permission in China were considered illegal overseas. In the end, this product that had been developed and internally tested had to be taken off the shelves.
After nearly two years of overseas - bound entrepreneurship, Chen Kaijie, born in 1997, has developed a hobby of growing green plants. He joked that he “has reached a certain age.” Although he has never experienced the “hundred - regiment war” in the mobile Internet era, as a startup team in the Agent overseas - expansion market, Chen Kaijie feels the pressure of intense competition.
The water in the overseas market is getting deeper and deeper.
“Silicon Valley investors used to receive thirty or forty BPs (business plans) a week, but now they receive seventy or eighty a week.” Liu Tianqiang also felt in his communication with investors that with the large - scale layoffs in Silicon Valley's technology giants, more people are flocking to the Agent track. According to a study released by Stanford University in August, the number of software developers aged 22 - 25 has declined by nearly 20% compared with the peak at the end of 2022. At the same time, investors are tightening their purse strings. “Many fund companies would rather put their money in the bank to earn interest than take risks to invest in startup Agent companies.”
In this way, in addition to the high costs of efficiency, promotion, and compliance, even the reason of easier access to investment for overseas expansion is not so convincing. Next, there are more pitfalls waiting for AI overseas - bound entrepreneurs, and they will sail into an increasingly deep sea.
This article is from the WeChat official account “Hangzhou Qiantang Enterprise Overseas Expansion Service Base”, author: Jingxiang Author.
