AI Proem Podcast

Show Notes
Hi all,
Christian Hu from Alibaba’s Qoder joins the podcast again to talk about the fierce domestic workplace agent competition. For context, Anthropic and OpenAI have decided not to allow access to their models and products in Greater China; thus, the competition for models and agents is largely driven by domestic players. For workplace agents — coworker-style products — the most popular ones at the moment are Tencent’s WorkBuddy, Alibaba’s Qoder, and ByteDance’s Trae. The other agents we see coming from the labs are mostly coding agents. With that, I’ll hand over the floor to Christian, who graciously found ~40 minutes while on a business trip to talk to us about the landscape and Qoder’s own strategic shifts.
This episode was recorded on the day Qwen3.8 Max launched.
For more, check out the podcast lineup here and explore the episodes!
Chapters
* 00:00 — Southeast Asia Expansion and Market Dynamics
* 02:43 — The Shift Towards Business Logic in Coding Agents
* 05:29 — Industry-Specific Applications and Vertical Agents
* 11:22 — Global Strategy: Lessons from Market Differences
* 14:00 — Partnerships as a Key Element in Market Strategy
* 16:41 — The Future of Coding Agents and Market Trends
* 19:22 — Pricing Models and Inference Economics
* 22:08 — Open Source Trends in AI and Market Competition
* 27:34 — Building vs. Buying AI Models
* 29:38 — Future Directions for Coding Agents
* 32:14 — Final Thoughts on Market Opportunities
Transcript
(AI-generated, for reference only)
Grace Shao (00:00)
Hi Christian, friend of the pod, you’re back on. Really excited to have you here. Where are you these days?
Christian (00:07)
Yes. I just landed in Singapore last night.
Grace Shao (00:12)
Okay, perfect. Cause we’re gonna talk about your Southeast Asia expansion. But since we last spoke, competition among coding agents has intensified quite a lot in China, particularly — the market has changed a little bit. Where are you seeing the whole market going, and how do you think Qoder fits into all of it?
Christian (00:32)
Yes, it is a war, you know, between every major tech company in China. Because it’s a war, nobody wants to lose. I think the logic behind this coding agent war is that most companies believe that a coding agent shall be the fundamental path to AGI — artificial general intelligence. So nobody wants to be behind in this race.
But the most interesting thing behind the war, or this race, is that something is changing. When we look back to the last twelve months, everyone is talking about the models, everyone talking about what kind of model will be the most competitive advantage for your coding agent. But now, for most leading agents, they are trying to be more focused on the business logic and workflows. Just like what Qoder is doing — we want to be more into the business logic of our customers. And even for some individual users, they are trying to do something for business. They have one-person companies or one-person workflows. So they need the coding agent to do more about the business workflow, not just code generation.
So that’s the shift behind the war. For at least most of the companies, they are trying to raise not just for a coding assistant, but want to be dominant as a desktop assistant for employees or maybe some individual users. And maybe in the future, they want to be the digital employees for the industry. I think that’s maybe the ultimate race for the coding agent.
Grace Shao (02:28)
And you’ve got players like WorkBuddy from Tencent that’s been doing really well, right? Exactly to your point—
Christian (02:33)
Yes.
Grace Shao (02:33)
I think they’ve been plugging in, kind of operating as an assistant on desktop, very intuitive and user-friendly. You’ve got ByteDance with Trae pushing in a similar direction. So how do you feel about that? Where is Qoder’s differentiating point or offering here?
Christian (02:51)
I have a few things to share with you. I think for the past year, Qoder has accomplished very remarkable business performance. In terms of ARR or revenue, we are the leading one — we are number one. And maybe we are more than the combination of number two to number four. That’s a huge achievement for Qoder in business growth.
And as I just said, Qoder from day one is more focused on building an agentic platform for developers and AI builders. It’s not just a coding tool or coding assistant — we want to be a platform for the next generation of agentic power. So the skill marketplace, the plugins, the connectors, the MCP connectors — all of these features are now shining. All of these features are bringing advantage to our business. That’s the truth of what’s happening in China and even in overseas markets.
Grace Shao (04:04)
That’s really interesting. And so when we caught up in Hangzhou recently, you were saying that Qoder appears to be broadening from just coding into industry-specific agents. That was something quite fascinating, because from the start of the conversation, you say coding capability is the fundamental foundation for all of these agents, but people are moving towards these vertical agent use cases. So tell us a bit more about that. What kind of industries are you guys targeting? What is your thinking behind this new strategy?
Christian (04:34)
Okay, I can share an example. We got a very important customer case with Xiaopeng. Xiaopeng is one of the leading electric vehicle producers in China, maybe one of the best. For Xiaopeng, for the company now, Qoder is not just a coding assistant or code generation tool. Qoder has become an agentic driver for the restructuring of their workflows and business logic. They’re trying to educate their developers and their engineers, maybe even the HR department, to use Qoder to renew their business logic and their workflows.
For example, for the legal department, they’re trying to use Qoder to reduce the legal review cycle from six days to one day, or even less than one day. And it’s not just Xiaopeng — I think there are a lot of similar cases. So for Qoder, it’s not a shift to pivoting to vertical applications or specific industries. From day one, Qoder wanted to be a platform. The platform means we want to be a tool — we don’t want to be just a coding assistant competing with other coding tools. We want to be more embeddable and compatible with the customers’ business logic and workflows. So it should be more vertical. But it’s not for us to do the vertical things — it should be let the developers, the customers, and even some individual power users develop domain-specific agents on top of Qoder. That’s the philosophy for Qoder.
Grace Shao (06:51)
Okay. So the thinking is you guys are offering the tool, but really it’s still on the user to build out the tailor-made agent for themselves. Not that you’re specifically pushing—
Christian (07:01)
Yes, that’s the truth. Actually, we got some organic creation of skills by the power users. They are creating some skills, some creating some plugins on our marketplace. We are going to open the marketplace to all the business users so that the users can deploy different kinds of skills or plugins from the marketplace. That will be the market for Qoder. Maybe in the near future, we could be a marketplace for agents, a marketplace for skills.
Grace Shao (07:49)
Interesting. Okay. Well, this leads to something else you talked about. You said that you guys were building an ecosystem around Qoder. So tell us a bit more about what that means, and how should we understand or expect how Qoder might change as a product in the next few months.
Christian (08:05)
Okay. I think we can change the view of Qoder — from a coding assistant to an agentic platform. I can give an example: we are doing something together with Microsoft. That may be an example. We want to collaborate with Microsoft to make Office more usable and more accessible on an agentic system just like Qoder.
You know, in the past, Microsoft Office is Microsoft Office, ChatGPT is ChatGPT — they’re quite different products. For the users, they have quite different experiences on two different kinds of products. And now we want to be one. For Qoder, for the agents, they need the agent to know how to use Office, how to deploy Office, how to use the best features from the Office suite — the toolkit — to help the users do presentations, do documents, process data. So that’s maybe the very interesting thing for the users in the near future.
This kind of collaboration is happening everywhere. Qoder is trying to partner with partners from collaboration software, finance software, even legal software, HR SaaS providers and vendors. We are doing things like that to be more compatible and more useful in the near future.
Grace Shao (09:55)
Very cool. Let’s take a step back. I want to talk about your business expansion, because I think when we talked about a year ago, you guys were really gung-ho about going to the US, going to Japan. I know you’ve been spending a lot of time in Japan yourself. But recently it seems like you are pivoting, or at least putting more priority and focus on Southeast Asia, and now you’re in Singapore yourself. Tell us a bit about the thinking behind your global strategy, and which markets you’re currently focusing on, given that you’re the head of GTM on the international expansion side.
Christian (10:26)
Yeah, you know, I actually got a lot of lessons from the last maybe twelve months — about ten months for my global journey with Qoder. Everything is different in different markets. I just came back from Paris. I think Europe is quite different from Japan. Most people think Japan and Europe share something in common — they are pretty slow in AI adoption, they care more about compliance, they care more about trust. But things are also — you can still find something different between Japan and Europe.
In Japan, the buying cycle is pretty long. Maybe six months or maybe even twelve months is very common. But in Europe, the cycle maybe is not too long — maybe one month or maybe one week. Because they are trying to catch up with the AI wave in Europe. But there’s still some obstacles in Europe. They care more about the law, the compliance, the GDPR, the AI Act. That’s the truth.
So for every AI marketer, or anyone doing good marketing in Europe, you need to care very much about the legal — the law, the act, the things changing and happening in Europe.
But for the US — I spent almost the first quarter this year in the United States. I met a lot of AI developers, AI startups, AI founders. The story is quite different from China, from Japan, from Europe. In the United States, speed is the most important thing. Speed means you are innovating. Speed means you are upgrading your product. Speed means you are accountable. You’re telling the users that you are accountable because you are innovating, because you are upgrading your product day by day, conversation by conversation. You need to upgrade your product.
Now I come back to Southeast Asia. Singapore is the first stop for Qoder to be a global product. And now I came back to Singapore, I will go to Vietnam and Thailand in the near future. For me, Southeast Asia may be the mixture for my go-to-market strategy, because in this place, in this region, you’ll find you have competition against some American AI vendors or AI producers. You will also find some Chinese competitors. That’s the mixture. This is a very competitive market, but it also has very high potential in this region.
Grace Shao (14:00)
Really fascinating, because you’ve done a world tour and given the high-level vibes of each area. And I can totally see that it’s also quite fascinating you say, like, ending up in Singapore. In some ways, it’s almost the most competitive for a sales role, because you know you have all the options in the world and nothing is actually off limits, and it’s really a price war as well. It’s quite fascinating compared to maybe other areas of the world that will be leaning towards certain companies or certain countries’ technology, given maybe geopolitical concerns or compliance reasons, regulatory reasons, whatnot. And other areas might be purely driven by price sensitivity. So anyway, fascinating. Thank you so much for that. But why then? Why are you guys now doubling down on Southeast Asia after your big global world tour?
Christian (14:51)
Yeah. Okay. I guess I forgot one of the most important things in our last conversation. Partnership is the key element and the key part of our go-to-market strategy. Because we believe that local partners, a local ecosystem, is the best way to get a connection with local community and local industries. Around the world, we have different kinds of partners. For the past twelve months, the most important thing I did was to find partners as many as possible. That’s the strategy for our global markets.
Okay, let’s go to Southeast Asia. Actually, I don’t think Southeast Asia is the most important one. Maybe I think for now it’s too soon to nominate which region will be the most important. But Southeast Asia should be a very important part, because Southeast Asia is a fast-growing market. It’s not too much affected by geopolitical concerns, and it’s fast-growing, so for every major AI company, there are huge opportunities, huge market potentials.
And for Qoder, we are growing very fast. We are evolving every day, every conversation. People in Southeast Asia — the developers really find that it’s very interesting and they can get much from Qoder’s evolution. Because for most users in Southeast Asia, to use closed-source agents from the United States or some other place is maybe too expensive. Maybe it costs too much for most users or most developers in Southeast Asia. But if they want to try Qoder, they maybe have the best position and best time to catch up with the evolution of AI and coding platforms. That’s a good starting point for most developers in Southeast Asia. And Qoder is evolving, so they will be evolving every day. I think that’s a good point for both Qoder and the developers and industry in this region.
Grace Shao (17:30)
Interesting. Let’s take a step back and look at just the overall industry at this point and the trends that are taking off. So the underlying coding models are improving quickly, and we’re seeing that it’s increasingly becoming commoditized, or at least the price is coming down, right? As all of that is happening in the background, how does that actually affect the tools that are built on top of these models, such as your product Qoder?
Christian (17:56)
Yeah. I believe the token, or maybe some of the large language models — I think the cost of token may be zero in the near future. Most of them. Maybe in twelve months.
Grace Shao (18:13)
Wait, how would it become zero though? What’s the thinking behind that?
Christian (18:17)
It should be. It should be. Maybe it’s science — it’s about science and engineering, but it should be zero cost. Looking back to the last twelve months, if you want to buy one million tokens, it may cost you about twenty dollars. But now it’s about half a dollar, just in twelve months. Maybe in the next twelve months, the cost may be quite near zero.
But I think in the future, there’ll still be some very expensive and exclusive models — some frontier models. There may still be expensive and exclusive for some users, for some industries. But most models will be very inexpensive, and even cost zero for most developers. And I think maybe ninety percent of the daily tasks can be accomplished by these kinds of models — the public models or maybe some cost-effective models. I think that’s the future.
Grace Shao (19:22)
Okay.
Christian (19:23)
So on this kind of zero cost, the agent layer — I mean the context layer, the agent layer, and the business layer — will be the most important one for the builders, for the users. So that’s actually what Qoder wants to do in the future.
Grace Shao (19:43)
I see what you mean. But coding agents can be very token-costly, right? So how are you then thinking about the inference economics, usage limits, or even your own pricing models when you are selling your products to users?
Christian (19:57)
Yes. I just said we want to charge our users — we won’t charge by token consumption. We will charge [differently].
Grace Shao (20:03)
Mm.
Christian (20:04)
We won’t charge our users by token consumption. We will charge.
Grace Shao (20:10)
Okay. So it’s not directly token consumption. If token costs come down, your cost of your product also comes down with it. Okay.
Christian (20:17)
Yes, actually, the shift is happening. We want most of the profit of Qoder to come from the agency layer — the workflow layer, the agentic layer, the context layer, the memory layer. That’s close to your daily work and operations, not just token consumption.
Grace Shao (20:42)
How do you view the subscription model? We still see that dominate coding agents today.
Christian (20:47)
I think for today, if you are just a coding assistant, you will be dominated by the large language model. That’s the truth. If you’re just a coding assistant, you should be dominated by the large language model. For now, the best model determines the best coding tool, because you’re too close to the large language model. If a lab has a very powerful logic model, the large language model can easily change to be a coding assistant. You just need an interface — you just need to type your language, you can be a coding assistant for every user. So if you are just a coding assistant or just a coding chatbot, you will be dominated by large language models.
But if you are an agent — an agentic workmate, an agentic layer that empowers users to regenerate their workflows, to connect with their collaboration layers, connect with their ecosystems — you will not be dominated by large language models.
Grace Shao (22:08)
I see. Okay. I want to pivot and kind of look at the model layer right now. I know you don’t work in the model layer, but you’re very familiar with the ecosystem. So help me understand. Chinese labs right now have actually used open source very effectively, right? And I think they’ve built up a global reputation and taken over a lot of the market share. And obviously, there’s a lot of discussion on whether they will continue to open source some of their best models, or maybe become more and more similar to US frontier labs, what we’re seeing with what they’re doing. What do you think of this trend? Will they remain open, or do you think they’ll gradually start closing up some of their best models?
Christian (22:51)
I think I can give you two angles. For the angle of the regulators — for the government — they want the open source strategy. They want the AI wave to be a new engine of growth for the national economy. That’s actually the national strategy for the regulators. So open source should be the trend, should be the future.
And from the angle of the market side — the market players. Actually, for most players in this industry, in the AI industry, open source is maybe the only strategy, or maybe the only path, for Chinese players to surpass their counterparts in the United States. I think maybe the only path.
Grace Shao (23:53)
Why is it the only path?
Christian (23:54)
For now, we can’t see the advantage from chips. We have no advantage about chips. We have no advantage about human resources, even some capital. We have no advantage. So maybe open source is the only path for Chinese companies to surpass United States companies, to be dominant in the industry — in the applications. We have some figures to prove that — Xiaomi’s language model and Zhipu’s language model are the most used around the world, right? They are the most used models, not a Claude, not an Anthropic model, not a Google model, in terms of usage. So actually, open source is maybe the only path, or maybe the only way, to compete against the counterparts in the United States.
Grace Shao (24:57)
Question on Qwen then, just because you guys just released another very large model — it’s not getting as much attention. I feel like the headlines are being taken all by K3 right now. But Qwen3.8 Max is claiming to be just as good as Claude. What’s your view on that? And it’s available on Qoder right now, right? So how do you view the model?
Christian (25:22)
I’m not a model expert. I can’t give you an expert view about which one is better, which one is not the best. For developers, for users, you can try Qwen3.8 Max — the preview — on Qoder. I think maybe positive. It should be positive. And we believe that, as I just mentioned, the model gap is closing. In the near future, the model gap is closing. So for Qwen3—
Grace Shao (25:53)
Actually, why is that? Why do you think model gaps are closing, closer and closer?
Christian (25:58)
That’s my belief. I’ll give you an analogy about this. I think the best analogy is education. You know, for education — university education — we’ve still got some top universities, right? The Ivy League, the best in the world. And maybe some of the best two or three universities in China, they are at the top. But most universities and most education in university, I think they’re universal. I don’t see too much difference in the education in most universities around the world. People can get educated, people can learn the skills, people can learn the knowledge in most — I mean ninety percent of universities around the world. That’s not the biggest difference.
There should be some frontier models, there should be some ecosystem models in the future. But mostly — more than ninety percent — will be the same. Almost the same.
Grace Shao (27:05)
So you’re saying basically it comes down to talent. Talent is strong. So therefore, more and more talent going into the industry, thus it’s catching up. But why wasn’t it catching up a couple months ago? Why was the gap, say, six to nine months prior to that — it was claiming to be nine to twelve months. All of a sudden now people are saying maybe it’s only one to two months. Where does this number come from? I’m always fascinated by these claims.
Christian (27:28)
I’m not a scientist, I’m not an engineer. I just have the philosophy, I have the belief.
Grace Shao (27:34)
Yeah. Okay. Well, let me ask you another question that’s kind of taking over the broader industry right now. The broader debate right now is about whether enterprises should be fine-tuning privately and deploying their own models, or even — you’re seeing more and more so-called tier two, tier three smaller internet companies in China leading the way in already building up their own models. Obviously, some people are saying they have the edge of having a very strong open source ecosystem in China, so the barrier to entry is easier. Anyway, from your point of view, do you believe this trend where companies are actually going to start building their own models or hosting their own models and fine-tuning on top of it? Does that make sense? Or should companies continue to do what they were doing, which is, you know, paying for the most frontier models from the frontier labs?
Christian (28:20)
Okay. I think for most companies, even some large enterprises, it doesn’t make sense for them to fine-tune their private models. I can’t see the sense, I can’t see the business logic behind this. Just like the analogy of education — I don’t think you need a private school for your own children. I don’t think most people need a private school for your own children, just for your two or three children.
Maybe ten percent of enterprises have very complicated, very sensitive applications, sensitive data. They need to deploy their own private models, fine-tune their own models. But I think most companies don’t need to do that. Instead, I think they should invest more in the context layer, just as I mentioned. They must invest more in the business logic, in the business, in the workflow, in employee management. That’s where they should invest. Because that will benefit their business growth, benefit their management, benefit their governance. I don’t think it’s necessary to invest in fine-tuning their own large language models.
Grace Shao (29:54)
Makes sense. Not every company should be working on the R&D of this, but really should be putting more energy into managing their own data, context, memory, expertise. Okay, well, I have a last question for you, which is, stepping back and looking at the whole industry right now, where do you expect the coding agent market to evolve over the next year? Because you alluded to a little bit throughout our conversation, but where do we see this whole industry going? Are we going to continue to see standard products that are going to be the main driver in coding agents, or like you said, model labs will swallow everyone’s lunch, and then it’s really all just putting up an interface on top of them? Are we going to see more enterprise-grade, specific use cases and systems built? How do you see this direction?
Christian (30:48)
Yes. I don’t think there will be too many coding agents in the near future. I think some of the coding agents, even some existing coding agents, will be swallowed by large language models. I believe that. Because language models can easily build a coding agent. So there should be some — maybe two, maybe three, or some number of coding agents still in the market. But most of them will be swallowed by large language models. And there should be some new agents — business agents, workflow agents, or some domain-specific agents in the future. We’ll have more domain-specific agents in the near future. Legal agents, human resource agents, even some agents for you — just like podcasters or bloggers, we’ll have some new agents for you.
Actually, agents will be just like digital employees. That’s the future. You will have more digital employees. You can ask your digital employees, just like you can ask your teammates, you can ask your friends — but it’s digital friends — to help you do anything you want them to do. That’s not just a coding agent, that’s your working agent.
Grace Shao (32:14)
Very interesting. All right, Christian, I have one last question actually for you, which is a question I ask everyone that comes on the podcast. You’re not unfamiliar with this. What is one non-consensus view you hold now? Or has anything changed since we last spoke?
Christian (32:28)
You mean from last time? What’s changed?
Grace Shao (32:30)
No, just any non-consensus view. Even what you said earlier was against what a lot of the industry is saying already, but do you have any view you think is very non-consensus or against consensus right now?
Christian (32:48)
Okay. I have something, maybe something shocking for many AI builders. I don’t see too much potential in the United States market. I mean for Chinese AI—
Grace Shao (33:06)
You mean for Chinese companies going global, whose first stop is often the US? Okay, very interesting. Go on.
Christian (33:09)
Yes. Actually, I want them to quit. You can invest in the US market, but I don’t think they can benefit, or get real margin and profit from the US market. But it’s a different story between consumer AI and enterprise AI. For consumer AI, there may be some possibilities for Chinese companies, because we still have some advantage — we have cheaper manufacturing, cheaper human resources. That’s the advantage for consumer AI companies to do business in the United States. But for enterprise AI, I think it’s quite difficult. Quite tricky and quite — it’s just not a friendly strategy to do business in the United States.
Grace Shao (34:11)
Where should they be going? Southeast Asia right now, Singapore, Japan, like you guys?
Christian (34:16)
Southeast Asia is much easier for them to do business, compared to the United States.
Grace Shao (34:24)
But the thinking from a lot of these founders is often that Southeast Asia, frankly, doesn’t have as much purchasing power either, or willingness to pay, especially on software.
Christian (34:32)
But they are growing. They are growing. We must invest in the future. And one thing I want to note — there’s much more potential in Europe.
Grace Shao (34:41)
Interesting. Okay. I think Europe has been a bit overlooked. People have kind of — some people have written it off a little bit, frankly, just given the recent years of slower innovation. How do you view the European market?
Christian (34:54)
I think Europeans are catching up. They are awake, they are working to catch up.
Grace Shao (34:59)
Mindfully.
Christian (35:01)
Yes, they’re working to catch up. That’s what I got from the Paris Summit, very intensively. And I believe — I think from the economy side, you must look at the market from the economy side, I mean the macro side. The economy in Europe — they are struggling. That’s my personal opinion. The economy in Europe, for many countries in Europe, they are struggling. So they need a new power, they need a new engine to restart the economy. I think AI should be the best engine for new growth in Europe.
Grace Shao (35:40)
Very interesting.
Christian (35:41)
So what kind of technology? Whose technology do they want? I don’t think they have the time to catch up just building their own AI labs, their own AI infrastructure.
Grace Shao (35:54)
Their whole stack, yeah.
Christian (35:55)
Yeah. They need help from China, even some other places. So I think that’s a huge potential for Chinese companies to do business in Europe.
Grace Shao (36:05)
Interesting. Okay. Very differentiated. Very cool. Okay. Thank you, Christian. Thank you for your time.
Christian (36:11)
Thank you.
Grace Shao (36:12)
Enjoy your trip and travels.
Christian (36:15)
Thank you so much. I enjoy it. Bye bye.
AI Proem is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.



Get full access to AI Proem at aiproem.substack.com/subscribe

In this episode, I speak with Ziwei Chen, product marketing lead for Accio Work at Alibaba.com, about Alibaba’s effort to turn decades of sourcing data and commerce know-how into an agentic business platform for small and medium-sized businesses. Accio began as an AI sourcing engine, but Accio Work is designed to support a broader workflow, from product strategy and supplier selection to store operations, marketing and growth.
The most interesting distinction is between AI that tells a business owner what to do and AI that actually does the work. Ziwei explains how agents can research products, compare and vet suppliers, draft inquiries, follow up on missing answers, update Shopify listings, generate structured data and coordinate campaigns across existing tools. Alibaba’s advantage is not only its supplier network, but the category knowledge, transaction context and direct communication layer built around it.
We also discuss where human judgment remains essential. Accio Work can narrow a supplier list, negotiate across variables such as MOQ, lead time and materials, and prepare an order, but the user still approves purchases and typically takes over the final supplier relationship. That balance matters because commerce is not only a workflow problem: branding, product taste, trust and long-term supplier relationships remain difficult to automate.
Finally, we explore why Alibaba built Accio as a separate, more open product; how it works with third-party platforms rather than replacing them; its subscription and usage-based model; and the future of agentic commerce across B2B and B2C. Ziwei’s non-consensus view is a useful one: not every problem needs AI, and domain expertise becomes more valuable, not less, when powerful tools are widely available.
For more, check out the podcast lineup here and explore the episodes!
Chapters
00:00 Introducing Accio Work01:01 From Alibaba.com to agentic commerce05:18 What an “agentic business team” does06:27 The commerce workflow and human control11:11 “Do it for me” versus “tell me what to do”18:32 Connecting fragmented commerce tools22:36 Alibaba’s sourcing-data advantage26:03 Supplier quality, matching and verification31:58 Accio versus general-purpose AI tools33:36 How agents communicate with suppliers37:09 What Accio offers factories and suppliers39:29 Designing AI for non-technical SMEs42:42 Business model and monetization43:54 The future of agentic commerce48:57 Why not every problem needs AI
AI-generated transcript for reference only
Grace Shao (00:00)
Hi, Ziwei. Thank you so much for joining us today.
Ziwei Chen (00:02)
Hi Grace, so good to see you.
Grace Shao (00:07)
I’m excited to talk about Accio So I just came back from Hangzhou like a month ago and I met some with some of your colleagues on the ground. Was very impressed by the product and thought was just really intuitive. So let’s get started. Tell us a bit about yourself and your role at Accio and what Accio is all about.
Ziwei Chen (00:24)
Well, hi everyone. my name is Ziwei Chen. I work at Alibaba.com as the product marketing lead for Accio Work. prior to Alibaba.com, I actually spent a few years in the world of developer marketing where I was kind of driving good market plans for software tools that are meant for developers who are building AI products. So now it’s kind of completing the whole picture for me to now move on to the other side about actually growing the products. built by those developers to the end users. So that’s kind of my kind of AI journey coming from the development side now to the end user side, driving the growth for Accio Work among our small and medium-sized businesses around the world.
Grace Shao (01:01)
And tell us, what does Accio do actually? So I think a lot of people are not familiar with it and just how it fits into the whole bigger, I guess, Alibaba ecosystem as well.
Ziwei Chen (01:09)
Yes, absolutely. So maybe I’ll take a quick pause and take a step back just to talk about the overall broader picture. So Alibaba grew as a large enterprise. we were founded in 1999, and our kind of main North Star was to make make it easy to do business everywhere. So that has always been kind of our focus of focusing on B2B and focusing on small and medium-sized businesses who typically don’t have that resources that large enterprises have. So throughout the years since 1999, we have been really just focusing on what we can do, either there’s a new product. And new services to make it easier for them to start a business, to launch a business, and to grow a business, maybe something to exit a business. So along that line, Alibaba.com is kind of the bread and butter where we first started, focusing on B2B sourcing. So actually, I will kind of almost break down our development or our growth journey into three eras and then show you where Accio fit. I will call it the digitization era, the AI co-pilot era, and then the agentic team era. So the first digitization is when we are first founded with that platform of Alibaba.com where on this digital platform we are connecting the sellers or what we call the suppliers, the factories, with the buyers, our buyers, maybe sellers on the other side, into one platform. So now they’re not limited by time zone, they’re not limited by the location. And then in that process, we what we have learned over the years is that things can get overwhelming. There just so many suppliers, so many products, right? So then when all of these foundational AI capabilities came about, we were like, okay, this is the moment. This is the moment where we can introduce AI capabilities to make it easier for our buyers to find the right products and the right suppliers. So the name Accio actually comes from Latin, which means summon. So the idea is we help you summon the right products, summon the right suppliers, summon the right information. So that’s when we first launched Accio as a sourcing engine back in November of 2024. And that went really well because a lot of our users now, even without any experience in sourcing, without any experience in physical products, and find it really quickly with confidence. But then what we had learned over time is that First, sourcing is only one small part of a broader business journey, right, for a lot of our users in the US, Europe, and around the world. So now we’re thinking, okay, what else can we do to expand on that? So that led to kind of the last phase for Accio Work. So when I compare Accio Work with Accio, a few things that stood out. The first is the focus from sourcing only to sourcing plus. The other thing is about the agentic part. So earlier when I talked about the three phases of digitization as a platform, right? The website. And then the second era is called the AI co-pilot. So that means is that you are still on the driver’s seat, AI is in the passenger seat. It’s giving you advice, it’s doing some little stuff for you, but you are still making all the decisions. You’re still wearing all the hats. And that’s what Accio did back then. For example, it can find suppliers, it can recommend messages for you, but that’s kind of it. So now we’re moving into what we call the agentic team era. where actually we’re gonna get things done for you and get more types of work done for you. so that’s kind of where we are really sort of moving into this phase, where truly kind of it’s like the spirit of the agentic commerce world, where you’re not only using AI as a passenger seat, but you’re actually arranging a team of agents to get things done and maybe let the teams work among themselves, what we call A2A. So that’s kind of the overall context of where we have been starting from Alibaba.com. To Accio the sourcing engine and now to Accio Work as the Agentic business team.
Grace Shao (04:45)
That’s very comprehensive over you. I appreciate that. So just to help listeners understand, 1688 and Alibaba.com are the sourcing kind of platform within Alibaba. And then obviously everyone knows Alibaba for the Taobao and commerce side of things, but that’s actually a merchant-facing product versus the Taobao and T-mall that are consumer-facing. with that kind of context, okay, so if I had to ask you to describe it in a very short sentence or like say two sentences. How would you actually describe the product today? It’s just like, who is it for? What does it do?
Ziwei Chen (05:18)
Absolutely. I think I would Start with three words that summarize what it is we will call it the agentic business team. and I’ll break it down each but one of them that eventually lead to the who and the how. So the agentic part means these are agents that get things done for you. It doesn’t just give you the recommendations, but it can write emails for you, send messages for you, publish things for you, right? And then the second part is business, right? So we are dedicated to the world business, specifically e-commerce as a really strong emphasis. an all all Aspects of business starting from the front development side to the growth side to the operations, everything. And then team part kind of aligning up to that as well is that you can build multiple agents to work among each other. So tip from a format perspective, we have a desktop app that you can download, also a web app as well. You can access the same thing on your mobile devices, on your web as well. So all these things are connected where you are creating agents on this platform. telling the agents what to do and then you can go to sleep, you can go to work, you can go to your events, right? All of these things, go to your shop, right? Do all these things while the agents are in the background running all these tasks assigned by you.
Grace Shao (06:27)
So exactly your point, like you can do a lot of things, and that’s what I found the most fascinating thing. So, you know, we know a lot of products on the market these days that can do a lot of the, I guess, step three, step four in selling. So what I mean by that is they can help you manage this SEO, the content management, the marketing, and the like, you know, plugging into Mailchimp and sending out your emails. But your edge really isn’t just the agentic AI part, what I thought is really your supply chain network, right? so of course Across the entire workflow right now. Like walk us through like the different steps. Cause I remember there was like a slide your colleague showed me. It was like there’s four different phases of commerce building a business. you guys can go across it and how much of it is actually truly a genetically run right now and how much of it is still, I guess, needing direction from a human to really verify and process in the pro yeah.
Ziwei Chen (07:17)
Break it down to the two parts. I’ll talk a little bit about the overall flow first. I think it can kind of follow a typical business like life cycle where you’re starting from the strategy and the planning part. Let’s say what product, what’s the portfolio, what’s your branding, right? What’s your competitor, what’s your positioning or pricing, all of that. And then into the actual product development and sourcing part of that. And then you’re going into your store management, especially your e-commerce and different marketplaces. And then eventually the last one will be for marketing and growth. it can be for your B2C. Your traffic, right? Your conversion. Also, some of our users are not only doing B2C, but also B2B to wholesale. So, how do you deal with that B2B, like management or relationship, all of that? And then, so that’s across the different core stages of run launching and running a business. And then Accio Work supports all of that, but not individually, but it connects to that all together. So that entire ecosystem play, right? It’s really where it stands out, where we call the sourcing plus. So it’s not just about SEO, right? It’s not just about how to improve, let’s say, the images on your store, but how do you do that starting from the moment you have an idea, bring it to life, bring it to your store, and then drive it to more users, and then maybe iterate over time all the way back to what other products you can do, or how can you make your current product better. And along this way, I think when it comes to which areas that are truly agentic, or kind of where does that relationship between human and agent do? I think that there are a few ways to go about it. So I think for example, all of the critical decisions or area still remains a human to be to approve. For example, you can develop or create an order. The agent can create an order on behalf of you. So you can just click and then go into the checkout page, but it will not check out for you. So you have to go in and confirm that or from a security perspective. That’s one example. The other example is about your own vision as a founder or as a brand owner. So let’s say it’s the position of your brand. Some of them can be subjective. Maybe it’s your personal story, your personal style of the brand. It can also be kind of that. More objective or critical thinking that after learning about different perspectives, this is where we decide to do. I think running a business is it is a lot about the science part, but it’s also the personal, the entrepreneurship, right? So a lot of these kind of personal decisions still plays a role. So for example, maybe the agent can develop all different kinds of images and videos on behalf of you, but how do you d determine what’s Good, it can be through A/B testing, right? But it can also just be this is kind of the aesthetic you are going for that you think what matters the most to your audience. So that’s a second example. I think the last one, taking another pause, I know that when we talk about human, it’s not just about the users, it can also be about the developers, right? It could be the people in the Accio Work team. And then so where we are seeing is that we put we will love for Accio Work, we are specifically targeting small and medium-sized business operators who are not. The strongest in AI tools. They’re not developers by training, right? That’s not where they want to dedicate their time to. So what our product team and engineering team and algorithm team have been really focusing on is what are the things that we can do it for you. So take the burden off of you. For example, in this broader ecosystem of everything that Accio Work can do, it does require a lot of connect connections to different channels and platforms right can be about marketplaces having different tools i think that ecosystem is where it really plays a really great job but then who is creating those connection points you know you can have your account but then from our perspective where our team spend a lot of our time is to build those connections so you can do a one-click activation to connect axial word to your email to your CRM to your Shopify store so that is one of the strategic areas where we decided to take more of a burden onto ourselves so our audience can just click and then log in and then start to connect the dots for themselves.
Grace Shao (11:11)
Very cool. I want to double click on the partnership and like how you guys built the plugins definitely a bit later. I do want to ask about the use cases because so here my brand. So when I went to visit you guys, there was like really it was really interesting. So on one hand, it was very obvious on how SMEs would use this, like you said, like you know, if you already have a very strong intent on how to what you want to purchase, what you want to sell, so then that whole process is very clear. But one Example one of your colleagues showed me was so fascinating. Basically, like they were running an event, someone was running an event using Accio. They went on the Accio like webpage and then basically uploaded a picture of like the banner, the physical banner like size of like an outdoor space and said, I don’t know how large this space is actually. I don’t know what looks the best with this background, but I need a banner to run this event. And then Accio became like a thought partner. And it was really interesting because it was not like just a general like AI like thought partner. It was actually someone that had a lot of know-how and industry know-how from creating these things, from manufacturing. So they would literally say, it seems like blah, blah, blah, the weather is like this and that. You might need this material because that kind of texture is more durable in the sun or durable in the rain. And I just thought that was so fascinating. and I just want to hear from you what use cases are most common and then beyond that, what are some like I guess not so mainstream use cases that you’ve seen that are quite fun and interesting.
Ziwei Chen (12:34)
I think when it comes to use cases, you know, we can break we can slice the pie in a lot of different ways. One way is about the specific type of like actions for a business owner to do, but I’ll take it a different approach to talk about I think just in general when it comes to agentic tools in general, I think there are two types of use cases where air comes into play. The first is like do it for me. And then the other is tell me what to do. So the do-it for me part is that I know exactly what to do. I know the specific steps. Either I’m just busy or I’m just one person, there’s not enough hands. So this is where AI does a really great job as a follower, but it can really scale for you, right? It can be, for example, When it comes to assessing different suppliers. If I’m just up one person, I can only s assess so many, but now the tools can do it for you. So what we are seeing in similar cases, it can be using Accio to find, evaluate, vet, and compare suppliers or products. It can be you up creating and updating listings, especially for those sellers. with categories that have multiple SKUs, multiple kind of products where I want to make an update to all of my marketplaces, all different colors, all different sizes, right? It’s really doing taking that repetitive work off their shoulder, but making sure things are consistent across different channels. can be creating content, right, on either social media, it can be about blogs, all of that. That’s one area. The other area that kind of really aligns well with what you’re saying is that tell me more. Like I don’t know what I don’t know. Sometimes I don’t even know what’s the right question to ask, or maybe I do know the right question, but maybe the LM itself did not have that training to start with. And this is where we are seeing a really strong aha moment from our users. And when I say users, it’s not only just people who are new to e-commerce or new to physicals or like products. That’s definitely the area where we have seen the most appreciation. But it can also be for even successful sellers or operators entering into a new category. And it happens a lot, right? Maybe extending your product line, maybe you successful exit once one brand entering into a new one. They all they they truly understand. the cost or the costly mistakes and how important that could be. So in both cases, what we are seeing is open-ended questions on this is kind of what I’m thinking, this is what I need, or just this is my store. Tell me what’s wrong, tell me what’s missing. And then lear first learning from actual work based on what we have seen in the space about key areas, key specs, key steps or things to flag. And also, I’ll tell you what’s wrong, or I’ll tell you what you should pay attention to, and then let me fix that for you. And I’ll show you a few examples. The first example about banner for sure makes a lot of sense. we had one, a similar case, but with a different flavor to that. We had a almost like a con like a product consultant. So he would take out new projects all the time. Very common, even as someone who’s experienced in the field, to enter into new categories. And he was helping a client with the bronze. plaque outside the building where you know there’s a building name or the history and all of that. And then so he it was a retro case where he’s already spent the time and the cost to pay for a designer to kind of put his client’s need into this kind of sizing and whatever. so he took the time, spent the money, sent it to the factory, and then realized that actually the font size did not work for readability purposes. So while he actually replicated that need into axial work. One of the first things that actually were told him is that depending on the technique for the for the font, is it going in or out from that plaque? here are his recommended size or the height of each each of these tags to make sure that you are following the best practices. And then he was like, Wow, I wish I had that when he was working on that project. So he can save not only the cost, because he had to redo the thing, repay the designer to fix it, but also lost weeks in between. So there’s lots of examples around that. They’re small, but they can be costly, right? Especially you’re busy, it’s your own money, you’re bootstrapping into your business. So that’s one example. The other one, it can be we recently had a had a user who was just who has a Shopify store and was just trying to see like what what can I do? What can I do with Accio Award? So he’s not someone who knows exactly this is a tool I’m gonna use for X, Y, and Z. He just pasted his Shopify link. To Accio Work and tell me what’s the low-hanging fruit. And Accio Work actually dig out a bunch of like very specific areas for improvement, especially when it comes to SEOs and GEO. For example, for all his images on his store, he was lacking the image alt tags, which is a really important thing for SEO purposes. There were thousands of images. And he was like, he knew that was kind of important, but he never really paid attention. But Axe Work was like, if with with this, now you can improve your ranking. And then in about a few hours, Accio Work did all of these tagging for him in the back end, on Shopify. And then the other thing was that his product was baby Like baby beddings or a baby clothes. So there’s a size guide. But then what actually work called out is that his size guide was just an image. It does not have structured data. So LLMs cannot read it, which means that he’s not getting any traffic when people, when new parents are searching about sizing for their babies, right? If it were baby clothes. So then he actually told me that he he started this task right before a drive, and it was an hour drive on the road, and then actually worked basically using that and One hour, recreated his size guide to make sure there’s structured data behind it. And then so far, he’s already seeing like new traffic from like LLMs or like AI, like what we call now the AEO, right? for the first time. So these are all really great examples of what I call like teach me what to do, or teach me kind of what I’m missing out on, where we are seeing lots of like aha moment from our users as well.
Grace Shao (18:32)
That’s incredible. I think I’m like absolutely in shock because I think a couple of years ago during COVID, I was like playing around with the idea of drop shipping things. I wanted to build this like pet shop store. But just like not having that kind of guidance is actually really daunting to start something new when you’re not in this space. And if you just had that kind of thought partner, that would be incredibly helpful. Actually, I want to ask you about the partnership thing you mentioned earlier. You talk it did kind of touch on, you know, there’s these plugins that are built in. at SEO will help the vendors actually distribute across different distribution channels. So walk us through that. Does that not kind of cannibalize your own business? Or are you is that not like, you know, you guys are not in the same category, you don’t see that? Like, how does that relationship work?
Ziwei Chen (19:16)
A few thoughts here. I think first thing first just putting Putting down our Alibaba.com hat and just think from the user’s perspective. I think one of the biggest learnings that we have had by looking at our e-commerce sellers in this space is fragmentation. It’s truly, and when I say fragmentation, it’s not just about multi-marketplace. Cause you know, some people have like a Shopify store or Etsy store, that’s one type, but also think about you have your storefront, you have your B2B sourcing side, you’re speaking with manufacturers and different messaging channels, and then maybe you have your own like CRM. System, you have your own, let’s say, analytics system, right? And the all of each of these might have its own very professional SaaS solutions already, or maybe they’re hiring a consultant for that. So I think a lot of their time is really spread thin by just connecting these dots. So I do think that’s even from a North Star perspective, connecting a dot, putting them into an ecosystem is ultimately what’s gonna drive benefits or impact to our users, which is what we want, right? So I think that has always been we feel confident in that direction. And on the other hand, I think just like how there are so many different startups or even statute companies focus on each one of these, just like how Alibaba.com, our bread and butter started from sourcing. So I think yes, we can choose to do everything on our own or which is gonna spread thin. Odd, right? Or we can choose to collaborate with them as well. Especially a lot of times maybe our users are already so familiar with an existing tool. So making them give up on and then transfer is one option, or maybe in the transitioning phase, right? Or maybe just as a collaboration to say, bring your tools into Accio Work with just one click and then you’re login, and now you can connect all the dots together. We are definitely seeing that, for example, we have a user who has their own kind of email, like like solution system where he was struggling partially because of setting up campaigns and take lot of time but also you know he’s using Accio Work for research for messaging and now he has to translate all of these learnings right into a different tool. So he just integrated that into Accio Work and use Accio War as a central hub of say I have a new product now establish a new email campaign establish a new SMS campaign and then this is my key message you go To those sites on behalf of me. So I think that consistency of I have a decision, I want to apply it across different platforms, is where it’s gonna bring the benefit to our users, while without making them transfer or give them everything in their ecosystem that they have built out over time. And I think in that process, what we are also benefiting or learning through those partnerships with those channels are also just like learning opportunities on what’s what’s the future of AI when it comes to different kinds of areas. I think a lot of those traditional SaaS solutions still have are also experimenting. So I think that also creates a really win-win collaboration like experimentational type of like collaboration of let’s see how what kind of new agentic behavior that maybe actual work introduced through our our plugin to the new platforms and now they are learning what else they can improve or enhance or ex or or in integrate into their own ecosystem as well. But ultimately, because the our sellers or users are using those tools back and forth, it’s a shared type of traffic that has that synergy in general.
Grace Shao (22:34)
Very fascinating. So so it’s.
Grace Shao (22:36)
It’s really the user experience first, but like Alibaba second in that sense. So let’s talk about Alibaba then. so it’s no secret ATT sits on top of like Alibaba supply network, you know, guys got transaction history, sourcing know-how, all the data in the world from e-commerce over the last two, three decades. how do we understand that? Like is that like huge edge for you guys? do you think you guys have leveraged Alibaba’s ecosystem to build your things out? How do you see that relationship?
Ziwei Chen (23:02)
I think in general, I would say that having our own like our own supply chain and our own learnings and the industry know-house is definitely like the biggest part about our remote. or the starting point of our remote. I don’t think it’s the only thing, but I think that sourcing plus storyline is something that we definitely want to keep in mind to differentiate from the rest. obviously at the same time, we also understand that Alibaba.com is only one of the marketplaces that our users might be sourcing from. So we’re also trying to make it more open in the sense that it actually where you can also search for suppliers and like products even outside of Alibaba.com’s ecosystem as well. but I think how that looks like into how that translates into kind of our own product experiences. So far we have briefly touched upon that industry know-how because we know all of these Category best practices. So no matter what kind of categories essentially we have experience in, you’re gonna get that intro that recommendation or that guidance along the way. That can also include negotiation and supplier communication. So for example, because right now through Accio Work, we have our AI Auto Chat function, meaning that now you can train the agent to conduct or multiple agents to conduct that sourcing process. Task on behalf of you. So instead of you having to stay up late, depending on time zone, asking or answering questions, you have 20 agents, right? Each speaking with the supplier, negotiating. And then when I say negotiation, it’s not just about the bulk order price. It can be about the sample cost. It can be about the lead time. It can be about your MOQ. It can be about different material flexibility. So all of these kind of How do you approach negotiation? How do you approach establishing credibility as someone new in the space? How do you nurture that relationship with the supplier? All of these longer term things. So it’s not just like you find a supplier, you place an order and you’re done. So that the overall growth journey is something that we see continues as our users’ business grow or kind of evolve over time. That’s one thing. I think the other thing again is kind of that. synergy or consistency across platforms where you made a change to your product because it happens quite often where you are updating your products, maybe because of your learning and the competitive space, sometimes learning from Accio Work, or it can be your suppliers are recommending the latest techniques to the production, like you know, capabilities that you’re incorporating that. So how do you make sure your communication on the production end is reflected to your marketplace? is reflected to your social media, is reflected to your community, to your newsletter. So I think that sourcing plus, kind of that type of like experience or that consistency is also where we see we take pride in as kind of that holistic experience or evolution for our users as well. and then so that’s something that we’re definitely taking pride in, keeping keeping it as our mode while expanding on How do we translate that know-how into every aspect of our users or our operators business lifecycle?
Grace Shao (26:03)
Yeah, actually I’m gonna double click on that. So Alibaba itself has cited one point five million verified suppliers, you know, more than something like four hundred million SKUs behind Accio. these are wild numbers. this is not even including other suppliers, right? Like you just mentioned. How do you maintain that data quality and actually how do you ensure there’s fairness in terms of like how you actually provide your users the right supplier network or the supplier contact because I don’t understand how the back end ends. you know, when I’m looking for say a hair clip, what comes up? How do I understand what goes behind the interface?
Ziwei Chen (26:40)
I think I’ll answer it in two different layers. So I’ll separate Alibaba.com where our supply chain network as one layer and then talk about how where actual work fits in. but starting from the Alibaba.com as our own network, it has we have established our own very comprehensive system for that verification. for example, for all of our verified suppliers, that verified batch means it’s being verified by a third party where they have submitted their like certificates, all of the requirements and the registries and all of that. That’s one example. And the second example is that for example for all of those like Trade Assurance or all of these things that we’re constantly evolving to make sure all of these things are available for you to make sure you know you can choose and also including a lot of so that’s kind of on our own network side that we have our own system. And then where Accio Word comes in because it’s a it’s a different name, right? It’s not called Alibaba.com So it offers a few different flavors. The first thing first is that in Accio Work, there isn’t any sponsorship or ads in place. Well in Alibaba.com, yes, there are those different placements, but Accio Work purposely decided not to do that to make sure you are getting recommendations based on the best matches of your requirements. So let’s say depending on where you are, sometimes you might be in an exploration phase. Let’s say, I’m interested in in building what was the other thing I looked for the other day? It was like Wooden bits for like board games. And I was like, I don’t even know like where to start. And then Accio can tell you. So here are the five key areas you should consider. Let me walk you through how to make each of these decisions first. And then let me match it with the right supplier. So don’t rush, let’s take it slow. Or it might be, I don’t know exactly what’s the model, what’s my location, what’s my MOQ, what’s my price training, etc., and then match it for me. So you’re what you’re gonna see in Accio Work is let’s say you have a list of five things, and then for each of these suppliers, are they three out of five, are they four out of five? And what else might be a recommendation? And this can be hard requirements. Let’s say it must be this model. it must be they must have expert experience in certain places. It can also be about preferences. Let’s say maybe you already have lots of factories in a certain region, and then you just prefer that this new factory for this new component is in the same region. So it’s easier for consolidation for shipping. So you can be a preferences. And those are areas where a regular label, right, doesn’t or tag doesn’t cover as well compared to an AI for LLM to process that for you. So that’s one that’s one area. And then the other area is that so that’s the first thing that actually baked in is that no ads just based on what fits the best, based on your matches. The other thing obviously is kind of having that third triangulation of data. So we do have this ability for you to vet a supplier on Accio Work, which means that what it’s gonna do, it’s it’s not only gonna search for in the back end for alibawa.com, like what kind of certificates they have, but let me go on a custom public records. Let me go in, let me go on social media, double check, do they have a presence? do they have they attended trade shows? So all of these additional data you don’t get from the platform, actually was pulling in. Compare notes to make sure you are making a confident decision on that end. so that’s the second thing. And the third layer again is that even with all of this fixed data, or sometimes could be outdated, or you just want to verify whether triple or triple check it, and that’s where that automatic communication comes in. Where now let let the agent directly speak to the suppliers on behalf of you, ask for photos of certificates, right? As for images of the factory if they don’t have any, triple check on all of these areas. So all of these three layers are how Accio Work is building a different experience to make sure you are making a more confident decision.
Grace Shao (30:28)
Okay, so why then was it so necessary to build a separate agent on top of Alibaba.com and 1688 instead of you know just adding a chat interface or conversation interface on top of I believe a lot of the Alibaba products that right now all have an AI chat bot it embedded in their product interface, but why was it so necessary to build a separate Accio?
Ziwei Chen (30:50)
Yeah, good question. I think A few things. the first is that maybe back to our North Star when it comes to that open ecosystem. I think having a separate brand or even of an identity is a must-have to make sure our supply chain is not limited to Alibaba.com and also our use cases are not just limited to sourcing. So that explains everything that Accio Work is doing by incorporating new supply chain options, partnerships, incorporating new connectors or services or like solutions together. I think that’s kind of the biggest thing itself. And I think the other side of the things is that when it comes to branding, to some extent, although we do have lots of existing Alibaba.com users adopting Accio Work for that new value add as well. But we are, Accio Work is really reaching a brand new group of audiences around the world. A lot of them are like net new, aspiring entrepreneurs. Who now realize that opening your business on your own is possible. So I think there’s a brand new type of like audiences that we’re reaching, with its own kind of the AI flavor of like a branding as well. So we’re building a community that is also beyond Alibaba.com.
Grace Shao (31:58)
That’s interesting. I have a bit of a a spicy question for you. Then what’s the difference really if I were a solo entrepreneur using ATSIO versus maybe I just get Claude then? I get Cowork.
Ziwei Chen (32:08)
That’s a great idea. I think a few thoughts here. The first thing first, the simple answer is that Accio Work is the only agentic platform now that has the official connection to Alibaba.com. So all the other AI tools, maybe they can search for some of these suppliers by using the browser extension and all of that, but they’re not gonna get all the back-end data about those suppliers and not gonna be able to directly communicate with the agents, cannot chat directly with them and also cannot connect the dots all across the board. I think that’s a the simple, the simple way out. But I think also because all of our team’s energy is super focused on this field. And I think in this world where everything is nothing, right? So by us putting all of our resources and energy into just supporting e-commerce, small and medium-sized businesses, we are making more progress, more in-depth progresses on just that ecosystem specifically. And I think in that phase, the connectors with the plugins are just one of the areas. The other one is the just the general industry best practices. And what I’m referring to is not just kind of what the platform has learned, but also the people in the ecosystem. we’re building a community, we’re sharing skills, very like or even sometimes building different agents and now you can learn from each other as well. So I think it’s not just about The data is about a community, it’s about a best practices as well. So people are learning from each other. all of those solopreneurs are aspiring entrepreneurs of how to start something and then grow from there.
Grace Shao (33:36)
So then I wanna ask you about something you mentioned, I think, in passing a couple of times, which is like the agents can help you speak to the suppliers. It was quite fascinating. how does that work in the back end? Like our are at this point are just like agents talking to agents and then making deals happen and then you know, you have a product that’s purchased for you and the next thing you know it’s already out there on a website. What does the human need to do still? So tell us about that.
Ziwei Chen (34:01)
Yeah, so maybe let’s we can walk through a pretty classic source and experience. Let’s assume that you kind of already have a pretty good idea about what you’re looking for. So you’re gonna go into actually we’re gonna talk about let’s say maybe this is a category, this is my specs, the color, the material, etc. give me recommendations and then you can it will give you some top recommendations you can select and then actually were the agent will develop a pretty professional inquiry for you to make sure it’s speaking. The industry language to set you up for success as someone who’s an experienced series buyer. And then you can say, select these five, what’s like top five, and send this inquiry directly to them. So then the back end, it’s connected through the Alibaba.com. So the agents are sending those information. That’s the first thing. But what else that’s gonna turn on is what we call the AI Auto Chat, meaning for each chat conversation, the agent is actively engaging in conversation during that time. And then what we have seen As the most common workflow, which actually aligns very well even in the pre-AI era, is that people start broad, they shortlist, and when it comes to the final one or two, they actually go in for the full-on investigation, right? So that’s what we’re seeing for Accio Work for our AI Auto Chat as well, is that the agents does the best, brings the most value in that early short listing phase. So let’s say for example, I have a really highly customized product. It’s really hard for me to tell or some really unique needs. It’s not enough for me just to go on a website to evaluate that. And I have a list of five questions for every single supplier. In the past, I have to ask each of these five questions, maybe two five suppliers max, right? But then not all of them are gonna answer the question in the right order, or maybe some are missing something. So you are going you’re the you’re the Excel sheet, right? You go into this chat and say, okay, question number one, number Number three, number four is done, but not number two. The other one is only as another whatever, right? So the agent, what it’s gonna do is say, okay, you gave me a list of five questions. I will make sure every single supplier responds to each one of them. If they’re missing one, I’m gonna chase them and then I’m gonna save those answers into a table so you can make apples-to-apples comparison. And usually halfway, this supplier says they cannot accept this, that supplier says they cannot do that, or they’re out of order for something. You end up with one or two, and then Typically now I can jump in to say, okay, now I feel good about it. auto-chat. You can take a pause, let me take over the communication just to make sure the final collaborations are on are I feel good about that. and then I can place an order, etc. And in that process, I do want to highlight that obviously AI can automate a lot of things for you, but in many cases, just like in other industries, it’s a lot about relationship, right? Unless You’re paying for what we call like ready-to-ship, like standardized product. Let’s say I’m just buying some balloons for a party or buying some bracelets for an event. Otherwise, for highly customized product, you’re building a relationship. So at the end of the day, it is recommended for you to have some interaction with the suppliers because you might be working with them on your next iteration and new product as well.
Grace Shao (37:09)
Very interesting. So it’s really just helping improve the lead quality and then shortening the negotiation time in many sense. And then but the ultimate actually decision making is still like in the hands of the human. that makes a lot of sense. So I want to turn the conversation around. We talked a lot about how it’s helping businesses, right? Helping the vendor, the seller. but the other end of it is the suppliers. And I believe that at a lot of suppliers are actually on at SEO and also turning on auto chat and whatnot. So what basically support do you provide suppliers and help them in selling their products to the vendor in the between or like the brand owner.
Ziwei Chen (37:41)
Yeah, absolutely. so we do have a version or specific plugins that are meant for the factories and the suppliers who are who can also choose to be on Accio Work for all of their for their side of the sales as well. and I think even pre-Axial World, a lot of the more AI tech savvy factories are already developing their own chatbot for services as well. So I don’t think that’s new to some extent. But I what I think a few of the really key areas that we are seeing success, the first One is actually a little bit more about background research on potential buyers because a lot of them maybe they’re overseas, they don’t have all the all the awareness or the ability to search across the world about learning about who is this buyer, are they serious, what’s the scale, right? All of these things. So we’re seeing a lot of our factories or the sellers, our sellers on the platform using Accio Work to To understand their potential buyers. Because a lot of times, again, maybe in the maybe there are buyers who are not as experienced but are very serious. That wouldn’t come across in their messaging, especially maybe English and other first language either, right? We have users in Europe, in in Latin America as well, so and also in the US. So I think by giving more data to the suppliers to get a better picture about the potential buyers, it also helps them. capture opportunities or avoid losing or missing out on opportunity. I think that’s a really critical part for sure. And I think the other part is about that general selling optimization as well. It can be most of the times it will be on the Alibaba.com platform. How do you stand out as a seller by learning about the data, learning about the user behavior as well? So I think these are a few areas that we are seeing a lot of usage from the supplier side.
Grace Shao (39:27)
I see, I see. It makes sense.
Ziwei Chen (39:28)
Excuse me.
Grace Shao (39:29)
Something we talked about offline before recording this is, you know, that a lot of the users, whether they’re the s your sellers or, you know, the actual brand owners, is that AI is still relatively new, right? They’re not the people in Silicon Valley. they’re not working for big tech. And that’s quite different from a lot of the other genetic tools being sold or marketed to these more so called sophisticated AI users. Now, how are you helping them, I guess, whether it’s upskill or understand or not be so intimidated by AI?
Ziwei Chen (39:57)
Yeah, great question. I think a few thoughts here. maybe the different a few different perspectives when it comes to the product development side, on the go-to-market side, on the educational side, et cetera. I think to start from the product side, I think what we have been really emphasizing or de-emphasizing is the is a is a focus. It’s too much focus on the features themselves. So what we have learned is that when we’re packaging them, we’re not like It’s almost like sometimes our users they don’t need a toolbox, they just need like an almost built up tool that can run itself. So what our product team has been really focusing on is not just to build the building blocks, but build the framework of those building blocks. So we’re not making our users to learn to become a Lego expertise like expert right away. so I think from a product end, it’s more about how can we take off the burden of the configuration as as much as possible. That has always been an emphasis, but even more importantly, as we’re getting more newer users in the space. I think that’s one thing. The other thing when it comes to just overall marketing side, what we are also experimenting, exploring now is a lot more when it comes to live events, either it’s in person or online, where we’re taking things slow. And then in this process, what we are trying to emphasize is more the workflow, the use cases as opposed to the feature. Like let me show you how this thing works from what’s the input and what’s the output. And this is all you need to do is to copy and paste and whatever. And then I can explain to you what happens on the back end. Because ultimately what we call the job to be done, right? They’re not adopting a nail, right? They’re adopting a hole like on the wall, things like that. I think that kind of stays consistent. I think the very last thing in this process From an educational perspective, that it’s also learning from my perspective, is that a lot of times the users are not only learning about AI, but they’re learning about just how to launch and run a business. That’s almost even more important than a tool. The tool is an enabler to help you achieve the goals in business as well. So what we are trying to do is to resurface the business know-how. It can be sourcing, it can be about product design, it can be about how do you do SEO or like how do you do like B2B, like pipeline or lead gen? And then let me bake in those automation in the back end. But what you’re taking away is how to do this thing, even in the pre-AI era, but now just making it easier to do. So I think that has been more of an emphasis from a go-to market or like a community building perspective as well, to make sure we’re not overwhelming our audiences. because maybe they’re already overwhelmed by all the AI tools out there in the market as well. So how can we shorten that and let them get to the aha moment faster and easier?
Grace Shao (42:42)
So how do you plan to monetize that SEO then? Are you guys charging subscription? Are you guys gonna start plugging in advertisement?
Ziwei Chen (42:49)
Yeah, so Accio Work is on currently on a subscription plan model, both for personal subscription plan and also for business plan as well. and then all of these typically by a monthly plan or an annual plan. It’s primarily based on token usage. So depending on how much how much work. How many types of things was the frequency of the task that you anticipate on the platform and then you’re upgrading yourself based on the amount of work that token consumes. So typically that depends again on the volume, on the complexity, and also on different models you choose to activate for each of these tasks. And that kind of r it aligns with our vision since the beginning, at least starting from Accio. the sourcing engine to now this the agented platform as well. So it’s mostly focused on usage. we do not currently have a plan again because of that kind of fairness or the quality of results, we’re not dealing with like advertisement from suppliers. most other things is more about how do we empower our our like sellers, like or buyers on our platform, the sellers to the consumers to get their business running using the agentic capability.
Grace Shao (43:54)
Now throwing it forward, where do you see agentic commerce going? Because it’s really interesting. I just interviewed agentic payment solution company last week as well. And then now obviously talking to you guys about agentic supplying, supply network support. where does agentic commerce kind of take us? Like in the future, are we just gonna be like telling the agents to do this whole transaction, this whole activity loop? Or Do you think that human is still needed for a lot of the verification and and and quality control in between?
Ziwei Chen (44:22)
Yeah, I think a few thoughts here. I do think that human in the loop is critical at different stages of a development. So for example, there are when we talk about agentic or agentic automation, right? We’re we’re automating an existing workflow. Yes, there are a lot of best practices, but it really varies, right? So even having your you’re co-developing an automation workflow with the AI tools in the beginning, and that will vary all the time. The more you The more you invest in that co-creation with the agents in the beginning, the better the automation or the workflow will work well for you. And then along the way, what we’re also seeing from our uses is that they’re learning while doing. You’re building a plane while you’re flying it. So there’s always gonna be learnings along the way. Just like in coding, there are ways for you to build the agents to do that automation by themselves. But I think the more complex the considerations are, the better. Better impact a human can bring to that loop as well. So that’s in the middle part, right? And also in the end, I think there’s still we’re still building that trust along the way, like between the human and the agent or the agentic loop as well. So I do think that having the option to or for an agency, for the human to have that agency, is always critical. That I think it’s it needs to stay. but then on the other hand, what I also want to add is that I think the word of agentic commerce. Can vary quite a bit between B2B and B2C. for example, in the B2C world, there’s always that shopping experience. It’s not gonna get taken away, right? Just sh going to it’s browsing that emotional experience is not gonna be replaced by AI. While the B2B world is a little bit more calculated, is more, right? But still it has that you have that relationship part. So I think in both end, I think there’s still human touch. In addition to human decision, needs to go in. But where I’m the most excited about is that because our target audience or our target user sits literally in between. They’re in this loop for B2B agentic commerce, and also they’re in this loop of B2C agentic commerce. So what I’m excited to see is how these two worlds are merging, or our our core audiences sit in this overlap in between. And then I’m I’m excited to see more of that. synergy of what works in the B2C authentic commerce world can get better integrated with the B2B world as well. Because again, like our sellers or our buyers are sitting between the factories, right? And them as a brand owner and then the consumers as well. So I think this is where I think the next iteration of innovation or maybe redefined like r the new definition of workflows of tools that might happen as well.
Grace Shao (47:04)
That’s super fascinating. Cause I think when I’ve been speaking.
Ziwei Chen (47:05)
Yeah, but.
Grace Shao (47:06)
To people who are working in the agentic commerce space, it’s often so very compartmentized, like what you said, like you there’s a B2C world but there’s B2B world, and these two worlds don’t really, you know, co like collide at all because whoever’s talking to a factory is not really telling what’s happening to the consumer and there’s no really feedback. Factories don’t even know where the product’s going and how it’s being branded. But now there’s that kind of ecosystem, I g like communication or the world colliding between on Accio. It’s very interesting. what do you think the world’s still getting wrong about agentic commerce? Because I think there’s still some people who are very reluctant or resistant towards agentic commerce or the idea of it. What do you think is people are having misunderstanding about it?
Ziwei Chen (47:47)
I think one thing that we briefly already touched upon is whether a gender commerce is taking away the agency of human. so I think that’s the part that causes the most concern or hesitance is like is the agent gonna replace me? But I think all a lot of these judgment calls, even that branding, right? that personal story about you being as an operator, you having that guidance on the branding on the story and all of that, it’s not gonna go away. But also it’s not gonna go away, but it’s also critical for you to inform where the agent should go. I think that’s one thing. But the other thing that I feel like maybe where I think it’s also like A of times when just as you said, when people think about commerce, they’re th thinking it still tend to go more siloed or go into just on specific steps. So I think it’s still that orchestration layer when things start to get connected and st start to flow together. So it’s not just about how well each of a step is done well, but how well all of these steps are connected. And then that’s also where that human in the loop or your agency is bringing that flow together. So I think that’s an overall area where I think it’s it’s People might be evaluating the wrong thing as opposed to their making the wrong judgment call about a one specific thing.
Grace Shao (48:57)
Thank you. And then my last question for you is a question I ask every single guest that comes on, as the podcast name is called Differing Understanding. what is one differentiated view you hold or something you think that’s very non-consensus?
Ziwei Chen (49:08)
Yeah, so I think my the first thing that really stood out to me is that not every problem for not not every problem needs AI to be solved, or AI is not a solution for everything. and I think that has implications both on the development side and also on the user side. So on development side, what we mean is that there are solutions or there are things that we can build that is more efficient or more accurate even without AI. So we should not shy away from that. And that should that will impact how people build products. But on the other hand, even for the users, that also means that skills in AI is important, but that’s not enough, or that’s not the only thing that I will focus on in this era AI era phase. Instead, I will still spend time in building your domain knowledge as well, because without that. You’re not gonna be able to learn how to use AI the best because maybe sometimes AI cannot solve all the problems as well. So I think be having a balanced view about where AI sits in the product or in your tool stack is something that will allow you to get the best out of all the AI tools available and drive results for yourself.
Grace Shao (50:21)
I love that it’s like keeping the human side of things in check. But I think it’s only non-consensus of where you sit because sit in the Bay Area. Only people in the Bay Area believe AI is taking over the world right now.
Ziwei Chen (50:31)
Okay.
Grace Shao (50:32)
But thank you so much for your time, Switzue. really appreciated your insights, and you’re very you know, deep understanding of the agentic commerce world.
Ziwei Chen (50:39)
Yeah, thank you. I had fun. Appreciate it.
AI Proem is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.



Get full access to AI Proem at aiproem.substack.com/subscribe

In this episode, I spoke with Patrick Wu, Founder and CEO of Clink, about what payments need to look like in an agentic commerce world. Patrick previously worked on payments infrastructure at Amazon and AWS Payments, and later led global payments at Temu as the company expanded market by market. His view is that payments today are largely built for humans at checkout, but not yet for agents acting under delegated authority.
The conversation started with a basic but important question: if Stripe, PayPal, Airwallex, Visa and Mastercard already exist, why do we need another payments company? Patrick’s answer is that the problem is not just processing a payment. For agentic commerce to work, the system needs to understand who the agent represents, what permission it has, whether the purchase fits the user’s original intent, and whether that authorization can still be verified later. Clink is positioning itself less as a replacement for payment processors and more as a connector across agents, merchants, payment providers, and card networks.
We also spent a lot of time on why payments remain so fragmented globally. Payment habits are local: cards, wallets, bank transfers, convenience-store payments in Japan, and different payment methods across emerging markets. Patrick argues that no single provider is best in every market, which is why AI builders and global merchants may need orchestration across multiple PSPs, local payment methods, and eventually multiple agent platforms.
The most interesting part of the conversation was around trust and liability. If an agent buys the wrong thing, who is responsible? Patrick’s view is that fully autonomous shopping is possible eventually, but the trust layer is not there yet. In the near term, agentic payments will likely work through delegated mandates: user-defined instructions, spending limits, merchant controls, passkeys, audit trails, and chargeback mechanisms. His differentiated view is that the industry may be too focused on creating brand new crypto or blockchain rails, when the more immediate challenge is making existing fiat rails work safely for agents — and getting the ecosystem to align around clearer standards.
Get in touch with Patrick via LinkedIn.
And company website here.
Chapters
00:00 Introduction
00:13 What Clink does and why payments need to evolve for AI agents
01:18 Why today’s payment stack is not ready for agentic commerce
03:07 Why payments remain fragmented across geographies, regulations and user habits
05:48 Clink’s role as a connector across agents, merchants, processors and networks
07:06 Why AI builders may need more than Stripe, Airwallex or PayPal
09:34 Visa partnership and what it means to be an “agent enabler”
15:23 What happens when agents spend a user’s money
21:02 Trust, liability, chargebacks and why China’s wallet model is different
23:26 What agentic commerce actually means today
43:54 Patrick’s differentiated view on fiat rails, stablecoins and protocol fragmentation
Transcript (AI-generated, for reference only)
Grace Shao (00:00)
Hey Patrick, so good to have you on today on AI Proem. Thanks so much for joining.
Patrick (00:05)
Hey Grace, thank you for having me here.
Grace Shao (00:06)
Yeah. To start with, give us the high level. Who are you and what is Clink in a few sentences?
Patrick (00:13)
Sure. so at simplest level, I’ll introduce Clink first. Clink is building the the payment and the bill infrastructure for AI builders and for the agent that they will serve. So payments today we see at design around human being on a checkout page and but we’re building the layer that let authorized agents complete the same transaction safely while helping merchants accept the global payments through the system they already use.
And for myself, I started my payment career path at Amazon and AWS. I learned how to build a payment infrastructure that has to perform at a very high level of reliability and the security. And at Temu, I led a team to solve the problem with payments and the global market expansion. So we expand country by country, market by market, and solving the payment issue locally.
And Clink bring those lessons to a new problem that when we see agent becoming the new commercial actor, but we see the payment stack is not fully ready for that, so we won’t solve that problem.
Grace Shao (01:18)
So what do you mean by the payment stack is not ready for that?
what do you mean by the world’s not ready for agentic payments? And where do you see the gap from existing offerings? Because obviously, when we talked offline, I did challenge you on this. I said, you know, there’s a lot of major fintech players already and payment solutions, Stripe Airwallex PayPal,
right? But where does Clink come in and how does Clink solve this issue,
Patrick (01:44)
Yeah, I think payments are largely solved for human who is present at checkout. so the industry had made it much easier for developer to accept payments like except from card, from our wallet and from like so many payment that local customers prefer. and also the developer can easily launch a product to launch a subscription, send a payment link, right? Those those are like a real progress already made the word.
So but what is not solved on the payment side we see is agents acting under the delegated authority. Right. Processing the charges only one part of that transaction, but the system also needs to know who the agent represents and the what permission that it has and whether the purchase is within the policy, within the intent, like matches the original mandate that the user has given. And then the how the evidence being being persist, right? And if we come back in six months later.
is that transaction still like able to be verified at the point that the agent made the decision. So I think like global merchants still face another unsolved problems like a single provider isn’t no single provider that best at the every market and solve all that agent issue. And then we know worldwide that the agent can can buy product from anywhere. So we see that’s a big gap that the agent e commerce need to
have a both the authentication layer authorization layer and also a practical way to across with all the providers in the market today.
Grace Shao (03:07)
Yeah, actually on that, you did say, you know, payments are very fragmented, right? Like, and but is that fragmentation mostly based on geo geography, regulatory reasons? Is it about payment habits? Like how do we understand that? Why is it so fragmented and it’s not a one solution fits all?
Patrick (03:24)
Yeah, I think that’s a great question and a great observation, right? And to be honest, I think it’s all about like all about that because payment really reflects a local financial system, right? Like you spend twenty dollars in Singapore that you may use a card, you may use your wallet, your bank transfer. But if you know in Japan that many e-commerce orders actually being paid over konbini payment. Konbini is the Japanese word for convention store. If you would like to place an order online and
getting to a company store and the pay when they check out some some goodies. So it’s just a habits and also the regulation and also how the payments being developed in that particular marketplace. So we see that it’s like all different across the world. That that’s a lesson we learned from from building Tammu payment stack that we actually adopted 70 payment method when in the first year. So every country has their own top three payment method.
And every time you add a new pay method, you got some new customers. So it’s kind of very interesting. And then but then the fragmentation is not simply like I I guess you can solve that by building an API and building more like a convenient experience, but we see that as a reflection of local, right? Like habits and migration, all that just as we said. And the the things that we cannot force the users, right? Use got their own votes for what they want to pay.
And that will always exist because that’s how the world works here.
Grace Shao (04:48)
And so thus like where do you fit in then? Like do you have a certain geo gro geography you’re targeting or a certain kind of I guess use case that you’re targeting?
Patrick (04:58)
you mean the agent e commerce or like in general? I think Clink today we try to solve two problems. Yeah. Yeah, we’re trying to solve two problems. Yeah. So one one problem is for today, like the global monetization for the digital servers and AI builders, right? They they’re selling their product, their great application to worldwide, and then the users from worldwide that like to pay them with the local payment.
Grace Shao (05:02)
Just in general because Mm. Yeah, yeah. Sorry, go on.
Patrick (05:23)
So that’s the monetization issue we we try to solve. And also also we’re trying to solve a forward looking issue that when we see the trend that the the actor, the consumer is moving from human to agents, and how agents can leverage the user’s assets, users’ local pay methods, right, and to make that payment, make that purchase, and to make all the information flow and the funds flow fat like try to flow fluently as it is today.
Grace Shao (05:48)
I see, I see. Let’s double click on the business. Help me understand your business a bit more. So one thing that really stood out to me from our earlier conversations that Clink acts more like a connector. how do you describe your role in the whole like payment stat?
Patrick (06:01)
sure, yeah. So I would describe Clink as a connector, as you said, right? And a controller or an aux trader. So we do not want to replace the bank, the car network, a processor, all the commerce platforms, the merchants or other agents, right? So we translate the agent’s intent into a transaction, or it could be the human’s intent and delegate to the agent, right, into a transaction. And the merchant existing system can accept.
apply the rules that the user has set earlier and route the payment as the user preferred payment and also preserve the record for the future validation or whatever verification that comes from. So that will let merchant connect once rather than building a customer payment pass that for every agent, for every processor. So that’s the important part that we won’t be that connector anywhere alive.
And it carries contacts, permissions, all that and solving the issue to connect all different kind of agents, different kinda agent platforms, and also the different kind of mer merchants and all the PSPs and different kind of networks as well.
Grace Shao (07:06)
I see. So I’m I’m gonna challenge you again on this way. I brought it up already earlier, but why would clients then choose Clink instead of an established platform, given that it’s so important for trust, reliability, proven processes in in payment, right? Why would they choose you over another major platform?
Patrick (07:24)
Yeah, we we mentioned some names earlier like Stripe, Airwx, PayPal, right? And there’s so many of them. Actually, I think probably thousands of processors all over the world. So I think sometimes when when you just want to sell one product in a particular market, then a single PSP might satisfy what you want. but I think the trend here today is all the AI builders, we are all selling products. what do I?
So that means like if you do want to of course trip and water can cover majority of the market, but we see in emerging markets and all different places, like different processes have different advantages, right? And if you as a merchant want to maximize that, then you may have to connect with multiple PSPs. Also pricing is another concern that if you want to lower your cost on payment acceptance, then you may have to have multiple PSPs to lower your cost.
So Clink will be like helpful as a connector that we saw the problem cross bond trees that you have multiple markets by default and you have all different kind of local payment methods, especially for example, we have some solutions in India, in Latin America. So like we are beyond one processor, right? So that’s on the payment acceptance side. But also we see that the AI application today requires, well, not that complicated, but does require some sort of bidding.
Right, top ups and subscriptions. And if you want to build from scratch, that’s something extra. And if you stay with existing PSPs, that’s a solution. But the expansion may not be that flexible. So we see that clinks serve as a connector layer, but that’s for the human business, the co like the product side of monetization. But when agents come in, that the problem becomes bigger, right? Because the you you don’t know where the agent platforms from.
It could be Kodas, could be cloud, it could be Gemini or any vertical agent platform that you never heard of. It could be own, like Ermers or Open Cloud. Right. And on the other side, the merchant may, as we we just said, the merchant may have only a single processor, but they may have many. But for that particular issue, then if there are many, who’s doing the routing? And who is connecting the agent to a particular merchant and their processor?
Right, and then we see that a single processor may not be capable of solving all that problem at once.
Grace Shao (09:34)
I see, I see. so tell us like how you’re working with existing platforms too, because this is another interesting point when we talked offline, you were telling me you have actually a lot of partnerships with traditional payment platforms like such as Visa, whatnot. how do you fit into this stack here?
Patrick (09:50)
you mentioned Visa, right? So I wanna bring an interesting fact that Visa described Clink’s role as an agent enabler. So I think that captures very well. So like there’s so many agents out there today, right? The big names we just mentioned, and also lots of self deployed or vertical agent platforms, right? And then we see that all those agents have the need the the requirement or the demand to make payments.
to do transactional for the user that trusts them. So but however that agents may not be able to integrate with our like Visa system. Right? Visa requires you to have a PCI because agent may not be able to touch the credential information. And if for self-deployed then like we don’t know like what is being uploaded and all that, all the security reasons and privacy concerns are there. So as the agent enabler
We process all that credentials, all that sensitive information. So we connect with the visa for all the like pass key verification, the mandate creation, intent verification, all that. So we perform that on behalf on the agents. And also so that that means we turn on or we enable the existing merchants to be able to connect with visa’s network securely or in a certified way. So that’s how we play with it.
existing I mean like schemes like a visa building a network also similar for MASCAR. And you mentioned some other big commerce platforms like Shopify and so I think Shopify is highly complementary for to to to zero, right, for sure. And there it’s already giving the merchants a catalog, a card, a great system. And they’re like a great pro it provides great solution for the merchant to easily
set up your own site and start selling goods. And we’re not going to rebuild all that. It’s it’s just simply not feasible and we see that the ecosystem is super strong. And we want to be add on to the ecosystem. Right? We see that added to like being a plugin for example, then we will turn on Shopify’s merchant to be able to like agent ready. Right. We build that agent readiness into the merchants. doesn’t matter it’s Shopify or WooCommerce or Mikado or so so many like open source
building platforms and commerce systems. We see all of them are being like like have their own adoption in different marketplaces and then the role of Clink is not going to replace all like any of them. And we want to be like add on to them. We solve the problem that the merchants at this point may not be aging ready and we want to make them aging ready. So that comes back to the earlier question that the role of Clink as a connector
Grace Shao (12:23)
Really interesting because basically, actually, you’re not trying to take the payment companies pie, you’re not trying to take the merchants business either. Then who are you charging in between how are you making money? Are you taking a cut from say visa and the transaction fee, or are you taking are are you making the merchants pay you for your service?
Patrick (12:42)
in in general the goal will be having the merchant pay us. But we are seeing different like from our own business models, right? Different monetization paths that if the merchant in our network that we have the influence on the consumer agent side, that we have the potential of charging them the billing management. So the the billing is not just to human billing but also to agent billing. That agent billing includes like
help the agent to check out and process that information, process the aging side of credentials that to allow the merchant to have the order and we capture all the credentials and the intent of all that. So that’s a potential of charging the merchants on that. And also we see the card schemes having some solutions to treat the agent and the everybody on the path being a like
the the the search changing or GEO side of thing that you have the potential of charging a fee for for the influencing capability. So I think in general today we want to build the ecosystem and having more and more users, agents and merchants on board and the monetization path will be clear and across all the actors over the ecosystem.
Grace Shao (13:46)
I see. Okay. Well, there’s often a disconnect between the merchant brand people and how they see, you know, payment versus what the infrastructure is behind the payment. ex explain to us like why there’s that mis disconnect, I guess.
Patrick (13:59)
Well, the the customer sees the merchant becomes the merchant they own the product, right? The price, a promise, and for e commerce is a fulfillment as well. And underneath that the connect the process actually connects the payment and the network and the process the transaction. So and there’s the issuer that decides whether to approve it, to decline it, and there might be fraud, credentials tax and settlement, all that. So those roles matter and when something fails.
So who made the promise, who improved the payment, right? Who processed the pre the payment, who protects the credentials, right? Who should handle the fraud, who’s responsible for that, who owns the liability. So I think the agent e commerce adds another actor and the chain becomes even more important. I think the the biggest difference in agent e commerce is when we look at from physical store to e commerce, right? like the ultimate decision makers do human.
But with Gene Commerce, with the autonomy that the agent is performing, does a potential the decision making is from the agent. Then is that decision making legitimate? So those questions are still remains and not being solved clearly. So all that we see that the clinic’s role is to try to carry that authority and the transaction context and all that across the trend. So again, come back to the definition of connector.
So we connect all the parties and try to pass through the information, the authority, across the chain, across the rails.
Grace Shao (15:23)
I see. We can double click on the safety and liability side of things later. I do find that quite fascinating. But it one thing you just brought up that was quite fascinating is that it agents will have autonomy. However, even though they have autonomy, the money, the feat of money is not theirs, right? So how do we understand that relationship if your agent is out there spending your money? do you always have to give them consent like before they go out? Or
Potentially what we’re gonna see in the future is agents literally just like going out there buying things without you even knowing they’re purchasing things.
Patrick (15:55)
Well that that’s a good question, all right. And I think the step by step would would just describe that agent in the future that like shopping fully autonomously it’s possible, right? But I I just think it’s not there because the trust is not there yet. So all the we we see all the demand is being delegated. So that means you as a human that you have some requirement, you have some demand and you want to make some purchases that essentially the demand is from human.
And the agents help you discover, help you to make some decisions, but maybe not for for the final decision. Right. So we see that as like involving pass. So but but as you just said, after all, the agents spend your own money, right? Spend your asset, see, spend your fiat. So we see that for today, in some case by case, like just one-time purchases.
That the user or the human will set up intent, give enough contact to the to the agent and the sign off on that. The sign off could be like a passkey, like we with a signature. Or it could be like A P two kind of a protocols, right? So there’s many actors in the industry try to solve that issue. But after that I see
Like the agent later on will do the the sourcing and do the patriotization as long as the falling fall under that mandate or instruction that was given, then we approve it. But essentially if we come back, see like where the orange and the context from, that’s from the user. So that’s for like a one-time instruction. But for sometimes if the merchant is trusted, like a digital service you always buy or digital you always top up frequently, then you can set up a recurring mandate.
Saying, okay, for this particular product or this particular merchant. as long as like over two hundred dollars a week or like every top up is a five dollar, that that’s fine. Then the agent does not have to receive or like get your approval every single time. But rather it’s just like a one time setup in the very early. But the setup or the mandate or instruction had to be super clear saying, okay, this is a trusted merchant. And we like it’s it’s fine for the agent to keep top up on that.
Grace Shao (17:53)
It’s like auto renewal but the without like resubscription.
Patrick (17:57)
Yeah. Well
I think the this is a little bit different, right? People are saying like if from merchant side if you give the merchant’s car and then merchant tell the merchant to do auto reload, yeah that that sometimes will achieve the same outcome. But I think the role is different. So on that side is merchant W and on this way it’s like the agent top up, like the actor is different. Right. And I think in general it’s like you trust the agent more or a tr tr or trust the merchant more.
Grace Shao (18:19)
Mm-hmm.
Patrick (18:24)
Right, that we see a lot of issues like disputes or chargebacks from the merchants actually deduct money, debit money from the user without that without the the the approval. And then with the agent being the actor that and also clean in the middle, then like we we actually protect the user. Right when the when the agent yeah, got
Grace Shao (18:28)
Mm mm.
I see what you mean. The ball
the ball’s back in the buyer’s court, kind of like the actual intent comes from that. It’s very interesting while we’re talking. I was thinking about a conversation I had with Ali Baba and Tencent recently. Both of them obviously are exploring agentic commerce. And then for them, you know, the biggest hurdle is how to manage this liability issue with who approves the payment. And obviously they both have payment systems embedded. So on the Baba side with Quinn.
Patrick (18:46)
Exactly. Right.
Grace Shao (19:09)
It’s really interesting. They’re saying that even though they’re doing agentic commerce, actually every time the money is gonna go out of your account, they’re still gonna push out a human verification kind of notification. You still have to like manually take that responsibility and press that button. On the Tencent side, I think they’re exploring the idea of creating a separate ReChat payment account, like kind of like two separate accountings, like two separate accounts.
And then one account will just potentially have very limited amount of money for the agent to play with. So you kind of just gave the agent like you know access to that one account, but not your full account. I don’t know. It’s just very interesting. I I’m kind of going on a rant, but it is funny because, like, on the other hand, it could potentially become like a scenario where essentially you give a credit card to your like unhinged teenage daughter and they would just go rogue and buy anything. yes. but let’s talk about agentic commerce.
Patrick (20:07)
No, no, I was just saying that which just is right, that just like giving a teenager your your child a card, right? And then you in in you don’t know what they’re gonna s spend and but after all you know them, right? But for Asian, I think the issue is like they may not have a trust yet, like fully trusted. So that’s what you mentioned, the Tencent has a weChat card and also the the Q and they they want the user to prove that every single time. And we see that that’s because the
the the different level of or different like I mean the path of payment being developed in in the country is different. So China has moved to the e-wallet and has very few dispute and chargebacks. So that means when comes to this particular situation, I think it’s just we we skip the credit card time. Like we skip the credit card UA and then move into the bank based or U wallet based directly. But if you look at the the
Grace Shao (20:49)
Why is that? What what
Sorry, my question is more like, why are there less dispute because it’s a U wallet? Wouldn’t like cat merchants still charge you if they want to charge you? Like if you’d like subscribe to certain things?
Patrick (21:02)
the the
because yeah.
yeah, so there’s two side of like if the the the payment is one side from the buyer to the seller, right, and then that’s real time transfer. And then we all know that today you have to scan your face in China, right? Or use a fingerprint to authorize the payment. And in those cases like the the dispute or charge back it become nearly impossible or or not necessary. And owning with a card card yeah, yeah, it’s so hard.
Grace Shao (21:23)
I see.
Like fraud is harder.
Patrick (21:37)
And then for for the auto debit thing then indeed and Alipay and WeChat has a compliance check and they revoke a lot of merchants auto debit. And now if you go want to receive the auto debit, it has to be super critical and then describe your business and the models clearly to the payment team and they will approve that. And it’s merchant by merchant of approval today. Unlike previously, you may either two turn that on. So but come back like when we
we see that issue particularly in card system, right? And then because like we we all know the disputes and chargebacks are all fraud, right? And like it’s just like so popular. Well not popular, it’s a common to see, right, in in the United States and in Europe as well. And that’s that’s why Visa has the entire chargeback management system to protect the the rights of the consumers. Right. so when they come to a genetic word that that create the the convenience that
because it is card based or the credential based that the user may not necessarily to approve every single transaction. So that brings convenience. But again that brings the risk, right? But however, thanks to the the chargeback system that has been developed over decades, the the risk can be like manipulated or to be like managed properly. so that’s why we we need the mandate system, we need the instruction to
to validate afterwards where when something really is bad then we come back to see if whether the transaction was being authorized properly by the human. And if it is, then well, probably the chargeback or dispute will not go through. Right. And so that mech mechanism does not e exist in Chinese payment industry today. So that that’s why that’s what I’m saying thinking is like may maybe the reason that they have to use a small amount of like a sub account, subcar.
Or a ask you as a human to approve every single time. And but honestly I think that actually yeah. So I actually Yeah.
Grace Shao (23:26)
I see, okay. That’s so nuance. Yeah, yeah, interesting.
okay, so but let’s talk about agentic commerce. So I think so many people have confusions around what really agentic commerce means. it’s you know, is it just that like basically like you said, you go out there, your agents are out there like going rogue and buying you things, or agentic commerce more understood? Like there is a process right now being built out. Like, help us understand what gender commerce means. What is the use case for agentic commerce right now, and are we actually seeing
it being proliferated in the real like real life right now.
Patrick (23:59)
yeah, yeah. So I think agent cameras is a very big word, right? And then it has many aspects and the dedicated execution. like I think that’s the most important happening right now ‘cause the the as I said earlier, the the the demand is still from the human and then the the product could be recommended by the chatbot or the agent. But after all that the the final decision may still come in from
the the human at most of time at today, right? But moving forward it could be the agent, with your authorization or understand you better, like has more context of you from the lifetime conversations, right? And then make a decision for you as long as the decision actually sits fitting the policy that you have set for the agent, right? So that’s another sort of like a next level of agent e commerce. so I think today
We see more and more like agent being an actor, but the actor could be on the sourcing side, it could be on the queuing side. And it really depends on how you as a user trust the agent. So in general I think agent commerce is very big but in general, like e commerce is a long chain, like doing all the actions from from the very beginning where the intent started and where to to the final the others play placed or the fulfilled. Right. So
in that long pass, any part that agent take kicks in, right, I think it’s agent commerce. But most of the time right now the agent is in the consumer side to provide help sourcing because the efficiency that they search products, right? And they may discover some very rare product that you may never be able to find. Right. There was a joke back at the time that probably as a human you read only first page of the Google search results, right? Super clear.
Right. The best place to hide their body is a second page of Google Search Results. But agent can easily search through twenty pages and get all different products or they they just have more context of you. They know your taste, for example, moving forward. And then they may be able to find like from a very real merchant, right? You you never know. Right. So that that’s a great potential that we see agent can unlock. And I think today that’s
Some of the aha moments I’ve talked to many people on Gene Commerce is the product search or recommendation site that the the the agent gave them some such suggestion or recommendation they never thought of as themselves.
Grace Shao (26:14)
That’s interesting. But I also think wouldn’t it make more sense for enterprise use case given the nature of like you said, sometimes it’s repeat purchases, it’s not always evolving. would would would that make more sense? Like if you’re a construction company, you’re buying like cement every year or whatever for this project, you know exactly the quality, the the the price you want to pay. your lumber company, whatever. You know what I mean? Like, wouldn’t that be much easier for agentic commerce to be in
Patrick (26:40)
Yeah.
Grace Shao (26:42)
integrated, implemented.
Patrick (26:43)
I think that’s a great point. but I I I do want to call that those cases are easier because they’re repeat repeatable, right? They’re bonded and you need to verify. it doesn’t matter the the buyer is enterprise or a pioneer of consumers, right? So it’s rather the behavior or the product itself define that the that use cases is more trustworthy or like you are more comfortable because to r like rebuy laundry stuff for example, right? And then
Grace Shao (27:00)
I see, yeah.
Patrick (27:09)
You don’t like that you always use a single brand and you you probably most of the time don’t want to try something new then then just a repeatable ask agent to to keep doing that. I think that makes it total total sense, right? To have the agent do that delegation, right? And like to to help you execute. And but like if it’s enterprise or like the pioneers of consumers, I don’t I’m not sure. It’s always like a some group, a small group of people.
They’re they’re interested in trying everything new, right? I like all the new stuff they they on board. I think those people are really push the word forward and help us to do the early adoptions and explore all the issues and helping improve the systems. It could be some pioneer of the enterprise, we don’t know, right? Small teams like startup teams, they are always willing to try stuff new. It’s all possible, but we don’t we don’t limit there. But I think you’re right that we use the
a small amount, repeatable and you need to verify purchases, to build the trust between the consumer and the agent. But the consumer could be human, it could be enterprise.
Grace Shao (28:11)
walk us through that case study ‘cause you were telling about it beforehand called Hello Minds. That was quite fascinating.
Patrick (28:16)
yeah, so it was a it was a short story. Actually, it was quick. when we were at Supreme in Singapore, they came to our booth saying, okay, they they see a clear need from their customer, they want to do transactional stuff. But as an agent platform they are not able to do that because they talk to Visa and then it’s just simply difficult for for them to to receive a PCI and all that in a short time. Right. So
then visa recommended us because we as aging enabler we’re designed to to turn on the capability of the aging platforms to do transactional. So and then we we just work with them super easily. There they have our skill pre-installed for their platform and then the the user be able to tell the HelloMind I want to buy this and that and then the system will actually get the user into our aging portal.
and add the card and we will go through the entire visa process for the verification, for device registration, and then it just happened automatically. So I think that’s a great example seeing like out of like all those vertical agents, they see the clear customer requirement, but for the role that they are today, they see a longer path to achieve that by themselves. And they see the they are seeking for for the help and for the experience partnership as Clink.
from the ecosystem to help them. So and then I think it it’s super clear for for Clink as well. We like to help them, right? And help other agents because our goal is being a connector. We’re now building our own consumer side agent. Of course we can do it, but we see it’s just like a payment method, right? Different people, different location, different markets have different preferences. Hello Minds is a great Asian platform over Hong Kong and and APAC area. So that’s kind of
like a possibly a capability that Clink provides actually help all their agent platforms to build their own and to satisfy or serve their customer better.
Grace Shao (30:12)
I see. Okay. Yeah. Cause you did say you’re agent agnostic. So that makes sense what we’re saying here now. but if the feature is, you know, many different agents acting on behalf of users, what does the payment layer need to do especially well then? Is is it in the identity, authorization, routing, settlement, security? Like how how I guess for you guys in the middle layer, what is the core, core offering that you have for everyone?
Patrick (30:36)
yeah, I think you’re right that like a one model may not fit all, right? Like a one base model or one agent may not fit all and the one payment solution may not fit all, right? Payment methods and the acquires, carnet was so many things. So all you mentioned, like identity, association, routing settlement, security, all of them actually matters. Right. So for us like we
We’re not going to like just build a single rear or like being a connector, we want to optimize every single layer, right? To be a connector that brought like try to bring every parties all together and closer and then try to solve the friction along the path for the agent to perform transactions. So in general that’s Klink’s goal at this point. We want to create a smooth word for the agent.
from talking to users and to the placing order on the merchant side and also have moving help moving the the fiat funds from from consumer to the merchant as it is today to have ‘cause we we see that as most smooth because after decades of development, the merchant solution, their commerce systems, their payment stack are being mature.
And there’s no reason to rebuild all of that just for the agent, right? And the return on that investment could be super low. So that’s why we see that as an evolving pass but rather a revolution pass.
Grace Shao (31:59)
look, I think let’s talk about the the most sensitive bit now. I want to talk about the trust and safety bit. So, you know, you do emphasize, you know, you’re building reliability, you’re building safety into the product and everything. But as the consumer, it still seems like really, really unsafe or scary to try out a new payment system. You know, as the average consumer, we will still default to big names like Visa and MasterCard, even if we know we’re paying them much more, right?
so help us understand that. Like what what is what is a trust barrier here and how do you build that up? and I guess who bears that liability? So hypothetically, I’m using Clink to buy something on whatever, let’s say Amazon. Okay. If something goes wrong, am I supposed to go to Amazon? I’m or am I supposed to go to you? Or is Amazon Amazon gonna come chase you down? What what is the relationship here?
Patrick (32:51)
Yeah, yeah, good question. I think so we are being agent enabled there or the connector here. So that means we are not the merchant of the record. So you still own your order on Amazon, right? So that’s our relationship with the merchant as well. We are not going to replace a merchant, repla take over their customer relationship. That’s not what we’re going to do. So what we’re doing is here, like it’s being a connector. That means we pass along the information.
like the payment credentials and all that, then we also capture your intent instruction that the the user talk to the agent and then send that over to Visa. So after all you’re still seeing your Visa card paid Amazon. I’m just using Apple as example. You after all you still see that you have an order with Amazon. And then you still pay with your regular car, right? Just the the in the middle the two actors, one is the agent. So agent performs the excursion.
as you authorized. And the clin perform as another set of actors and take over the agent’s direct execution, but help you and your agent to place the order on Amazon because we have been certified by Visa. And we have certified by the payment industry because we are PCI compliant. PCI but by the way, PCI means that we can securely manage all the payment credentials, right? Being authorized by the certified agents that we can
persist and the process user’s kind numbers, right? And given that we have capability to pass that over to Amazon. But after all, still you place an order. We’re just solving the problem that agent may not be able to directly touch the credentials or you don’t trust the agent too, right? And I mean at this point. So that’s why the we as a connector in the middle, we persist your critical information and then pass along to the merchants. And also we
Help you validate the policy that you set for the agents and the the initial instruction. For example, you authorize the agent to buy a shoe, right, and under $100. And somehow the agent got mad and bought you a t-shirt for $200. Then clearly it does not match the initial instruction. And we will block you before we send that over to the particular merchant that the agent wanted. Right. So that provides additional layer of protection. And also because like all that is being also certified by the visa.
Like because we are the first agent enabler and the agent side of partnership with Visa in APAC in in the Visa intelligent commerce program. So all that is being like even though it’s still pilot, we see the potential that we can add additional layer of protection and also additional layer of privacy.
Grace Shao (35:23)
Really interesting because I actually have two questions. That means number one, I mean, this is not unique to you guys, like, but you essentially have a database everyone’s credit card. There’s definitely a risk there, right? And then the second part of the question is that, like you said, when you don’t trust your agents to touch your credit cards now. So, say one day the education has been done, the market has been educated, agents have been normalized. Five years down the line, everyone is using agents to do shopping.
Does that just completely skip over Clink then if we’re all just gonna give our agents our credit card numbers, if we even still have credit card numbers, whatever payment look like back in the future? Would that just actually frankly o omit the the the role of Clink then?
Patrick (36:03)
that that’s a good great question, but I don’t think that will happen because for like even though the agent development all that then the the the concern to like you still need additional layers of protection because it’s like we are the neutral layer and give you the additional protection, right? Even though you trust the agent much but you still want the third party. It’s just like at the very beginning
of the e-commerce in China, you have Taobao, you have the consumer, like you have the buyer and the seller. But you still need to about just LP in the middle, right? To prov provide you that layer of security and you you want another layer. Yeah, so a a neutral standpoint to to to also right to to monitor, to audit the agent behavior. And also from the car scheme or the merchant side, they want some additional layer of protection as well. Right. No one can simply trust a single agent can perform all of that.
Grace Shao (36:38)
See what I mean?
Patrick (36:54)
Right. It’s just like additional yeah.
Grace Shao (36:55)
I see what mean. So there’s guardrails
built into you as like the guardian almost of all this.
Patrick (37:01)
Yeah. Yeah, exactly. the an an a layer of protection and I think Clink by the time it will build the trust across the consumer side and also build the trust across the merchant side. Because we see that the potential of having more a merchants being aging readiness and into the aging ready merchant network will bring them the advantage across the five years development that you just mentioned.
Grace Shao (37:24)
So I have a question. I I have two three questions to wrap this up. first one is what is your vision of the future of agentic commerce then? Like where do you see this as going? You can start with this, and then I want to throw you another one first just for you to think about in the back of your mind, What is the smartest question you’ve heard someone else ask you about the agentic economy that I have not not asked? So
Patrick (37:44)
Sure.
I think yeah, I again I guess we can get started with agent commerce, how that will go and aging payment. I see that has a long way to go. That as you said, like you have a child and like if you have a child that they grow up, they become adults, they become mature. And actually at this point we see that agents are growing fast, but are they close to the AGI?
Right, are they close to a particular point that that they’re being smart enough enough to do all that? So I’m not sure about that. And on the scare side, if they are really that smart and as another adult, do you really trust them to manage all your assets? Right. So those are like I I see those questions are tough to understand, are tough to answer, but at this point we see that agents are being super good or super useful as a tool, right, as a helper.
to help you do the excursion to do all the repeatable work or like all that you want to skip. And the sum of part of the commerce is being part of that. But I also see that shopping is a great journey, right? I think you as a lady understand what I’m saying. So I I mean as I mean I I typically don’t do enjoy that process, but rather just get what I want, right? But my wife and they do enjoy like the journey of exploring products. I
I think those are entertainments, not just commerce or shopping. So those will never be replaced by agent. But it could be like you I I know there’s some services in in China I’ve heard of that. like you you got a guy or like a girl that accompany you or walk through the the big malls and do shopping all together, like a company, right? It could be like in the future those accompanying role replaced by agent. It could be. But still provides a a great journey.
help you sourcing and find a great product that fits you. But I think that also would be a very pleasant experience. But I I think in general like for now we see that it’s a delegation for e-commerce for most of that valuable part. it’s not a new demand. But we also see for the digital service side and also for the like a
Mostly on digital and the model requirements or all the model f model side of features, those are new requirements. But it it it replacing a lot of the previous like data sourcing work or like investing work investigation work that you’ve done as a worker. But this may be replaced by agents as well, and along that path will generate a different requirement or different demand for genetic payment. So those are new, I think.
And in general along the past, hopefully we see that agents and humans in a word that can both like pay and get paid. Right. So that like today the the we’re we’re talking about AI applications, but in the future the the the product may be served by agent. But you never know. So it’s like a human to human, human to agent, agent to human. We’ve seen an interesting case that Waymo plants order online.
get a human to close the door. Right. If s a passenger got off the the the ride and forgot to close the door, and then you you place an order and get a human to do that. Right. So it’s like aging to human, human to aging. We never know. It’s that’s what you said, agen agent economy, right? Just not not commerce. So I I think like that word will eventually come and along the way that we have the base models to develop being smarter and also a lot of
Issues like around identity, around trust in general to solve. And then that’ll that’ll be the case. And then to a second question, like the smarter question, I think is like really come to that when when intelligence or labor or execution becomes abundant, right? what remains valuable? Right, or die, right? And do we still need transactions?
Because whatever you want, like the goodies or the services is being so abandoned and probably at no cost at all. So where is the transaction? Where is the value? Right? I think then that I think that’s a smarter question or tough question could always come to me saying, like really like if we as a society or the entire world develop who develop to that level of then what what’s what what are you p people paying for? We never know.
Right. And maybe that that’s a time that come back to okay, the the taste and the the people that creating the value that I I don’t know, I just don’t know. I think that’s a very broad question, open question. And we may not have an answer when really that comes. Yeah. What becomes valuable and what are people paying for?
Grace Shao (42:07)
First, I wanted to say it’s really scary to think that we potentially have machine overlords telling us what to do and working for them. second of all, I think the second half is something that’s really interesting. It’s it’s not just in the whole abundance of a gentic era. It’s just like you’re seeing it play out already. Like even people are saying, you know, luxury is a structural short because why who is still buying luxury? When you have more money, are we really spending money on putting
Patrick (42:15)
Well, that’s a bit scary.
Grace Shao (42:35)
you know, brands and logos on our bodies anymore or is society moving towards like paying for experiences, praying for health, paying for, you know, things that actually cannot be so easily replicated because, you know, logos can be replicated easily these days. It it’s kinda interesting. So it it will be interesting once intelligence is abundant, what will happen. But then I do think that question is also overgeneralizing humanity because we do forget that, you know, not everyone
Is working in a career that actually requires intelligence. I’m not saying people are not intelligent. I’m just saying intelligence as a currency to for, you know, salary or, you know, whatever or or capital gain is actually not for everyone. And if anything, then do we say the value is more in craftsmanship, laborist work, you know, you know.
Patrick (43:08)
Yeah.
Grace Shao (43:28)
experience and different things. It’s just very interesting, but for sure it’s already moving away from objects, like tangible objects, like where people really want to pay, right?
Patrick (43:36)
Yeah. Right. And you th you mentioned labor go go ahead, no, that’s a sign. No.
No, I was just saying like a labor as well. Like the the labor may even the work may not require intelligence, but that work and like the labor side of work may still be replaced by robotics. And with a small robotic, that that’s even more scary, right? Like as a move yeah. Yeah, yeah, yeah. So but but hopefully we we humans still aren’t in in control when that comes.
Grace Shao (43:54)
I s I see where you stand in all of this, Patrick. I see where you stand.
I don’t know, Patrick, the way you’ve pro you painted the futures, we have machine overlords telling us to close the door. Why are we even leaving the house then? okay, jokes aside, I have my last question for you, which is the question I ask every single guest that comes on. What is one differentiative you you hold, something that you think is maybe against consensus or people still you think get wrong or underestimate?
Patrick (44:12)
Yeah.
I I think the two two sides I want to answer that. One is that I think we’re we’re spending too much attention or like on the itching payment side require a new brand new rail like broad chain a crypto rail. So I’m not saying that’s wrong and we see the potential of the stablecoin or the capability of blockchain that for cross border settlement, machine too much micropayments, all that, but like a a reel that
Like the the real exist because we move value, right? And most of the time today the buyers and the merchant that they use VAT and we see we may not spend enough attention to solving that side of the rails issue. Right. So but I I see that on the long run, of course, the blockchain aspect stable coin will have a great value and being super useful. But at this point if we want the agent to be capable of handling payment.
instruction doing all that payments and we probably want to solve more issues around the the fiat rail and to to make that smooth. So that’s one thing. A second thing is I I think like even you’re not a payment in the payment industry you may hear of a lot of protocol names, right? A P two, UCP, ACP, SPT, and it’s just so many of them, right? And I’m not saying like
like every single protocol as the has their values and standards and so try to solve a particular issue or a like a vertical scenario, right? Or some of them being try trying to be generic and some of them well like we’ll want to solve the issue of product discovery and all that. And but I feel just too many of them. Like just way too many of them. And that creates a barrier or creates a a like confusion for the merchant.
And for other actors in the ecosystems, like which one should I adopt? Which one should I integrate? And the more and more are coming out, right? Every single player has their own standard. And as a merchant, I may feel like so confused, like which one I should adopt. I think really that the entire ecosystem will want to align or I mean to get some agreement.
Well, I mean not necessarily a single one or two protocols, but to have some fundamental rules in place that people understand okay, that’s how we want move and that’s the direction we go and people start investing on that. And eventually some generic or universal protocol that’s been agreed by everybody, like by the majority of the actor in the market, they’ll come out. And that’s a time where I see that
agents or the merchants will be more comfortable. Otherwise it’ll be too many of them for for the ecosystem to advance to next step to to next stage that agents and merchants are like feel so like worry free to onboard. Otherwise they feel like okay if I bought I onboard A C P last year, right? And they’re really like like not all agents are supporting them. Right. And then if we are on board U C P this year, we see other agents not support them. So like those kind of issues are being
like a making just like a too too fragmented. And as a generic like a pass forward, I I I really want that we we somehow get aligned in one particular vision, the entire ecosystem and try to push that new outcome.
Grace Shao (47:38)
There needs to be some standardization consistency to help agentic commerce go forward. All right, Patrick. Thank you so much for your time today. Really appreciate all your insights. I’ve learned a ton. if anyone’s interested, feel free to reach out to Patrick or I’ll put the link in the podcast show notes as well as on Substack notes. thanks again, Patrick. Speak soon again.
Patrick (47:56)
thank you. Thank you, Grace, for having me here.
AI Proem is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.



Get full access to AI Proem at aiproem.substack.com/subscribe

When James Peng founded Pony.ai in 2016, many in Silicon Valley believed autonomous driving was only three to five years away. But he expected it would take at least a decade, because the challenge was never just teaching a car to drive. Commercialization also required regulatory approval, public trust, reliable fleet operations, and a cost structure that could support large-scale deployment. Ten years later, his vision is becoming reality.
In this conversation, we start with his founding journey, the milestones and how Pony.ai became a leader in the autonomous driving space. We also discuss the gap between assisted driving, Level 4 autonomy, and the longer-term goal of Level 5, as well as how Pony.ai uses simulation, real-world driving data, and increasingly capable AI models to improve safety. James explains that the hardest problems are often not the obvious ones but interpreting unpredictable human behavior and handling rare edge cases consistently.
The conversation also explores China’s cost advantage in robotaxis. A mature automotive and electronics supply chain, close collaboration with automakers, and faster iteration can materially lower vehicle and system costs. But moving into new markets still requires Pony.ai to adapt to different road conditions, regulations and driving cultures, from trams and roundabouts to local pickup behavior.
James’s broader point is that the industry has focused too heavily on the initial technological breakthrough. Getting a car to drive itself is only the beginning. The next phase is about deployment density, utilization, maintenance, charging, remote support, and economics. At the end of the conversation, I asked what he believes is underrated. James, an experienced operator, replied - scaling. Pony.ai may have crossed the zero-to-one threshold, but the harder task is scaling from one to ten, and eventually from ten to one hundred. Check out this insightful conversation.
For more interesting conversations with people who are charting the way of the future of AI, check out the podcast tab or follow us on Spotify!
Chapters
02:36 Why James Peng founded Pony.ai06:36 The milestone that proved robotaxis could work09:27 How passengers learned to trust driverless cars11:41 Level 2, Level 4 and Level 5 autonomy18:59 How AI and simulation improve self-driving24:13 Teaching cars to understand human behavior29:11 China’s cost advantage and global competition35:03 Expanding robotaxis into international markets41:13 Why Pony.ai is also building autonomous trucks48:40 Adapting to new cities, roads and driving cultures53:45 Why scaling is often harder than reaching zero to one
Transcript
Grace Shao: Hi everyone, welcome back to another episode of AI Proem Differentiated Understanding. This is your host, Grace Shao. Look where I am, the back seat of a car. Doesn’t look that exciting, does it? Let me flip this around. Look at that. There is no driver. I’m in the back seat of a Pony.ai robotaxi. Today joining me is James Peng, co-founder and CEO of the leading robotaxi company. It’s expanded its footprint across the globe, in Asia, in Europe, in the Middle East. But obviously, today we’re in its leading home market, China and Shenzhen, where it has a fleet of a couple hundred vehicles deployed on the streets already. Hi James, thank you so much for sitting down with me. I’m really excited to be having this conversation with you. So you left Baidu in 2016 to found Pony.ai when many in Silicon Valley were saying self-driving cars are only three years away. But obviously that wasn’t the case.
Grace Shao: So, what did you believe then that this consensus was getting wrong? Tell us about your journey from 2016 until now.
James Peng: Yeah, sure. We were founded in 2016, about 10 years ago. But even at that time, I didn’t believe that autonomous driving can be solved in three to five years. Just from a technical point of view, because even back then, 10 years ago, even a demo for autonomous driving was already very hard. Later on, there’s complexity involved in the autonomous driving industry that involves regulation, user acceptance, the readiness of the ecosystem. So because of the sheer complexity, even then, my prediction was it’s going to take at least a decade for this to be a real application. It turned out to be that my prediction was about right. Now, 10 years down the road, we actually have fully driverless commercial applications in many cities. Of course, it’s just the beginning of the long journey for autonomous driving. But at least now we have real commercial applications.
James Peng: So I think people, like any new industry, people were super optimistic for the short term, but they were underestimating the potential for the long term. So I think autonomous driving is definitely one of those industries.
Grace Shao: What really drove you to want to actually work on this, work on this technology and the future mobility?
James Peng: I think the motivation was twofold. One is that the potential, both commercially and also societal benefits for the autonomous driving is so huge. Think about like everyone needs to have some sort of mobility. Autonomous driving is much safer than a human driver. So it has huge societal benefit of saving people’s lives. So essentially, it’s just such a great industry to work on. Although back then, 10 years ago, it was very unclear when this can be done. The other reason is, of course, because the sheer technical challenge of autonomous driving involves because I was actually in my previous jobs. I worked on different areas, software, hardware, large scale distributing systems, AI and whatnot. But none of the things I worked on is as complex as autonomous driving, which is a field that involves hardware, software, hardware and software integration and Many other things. There’s AI, there’s real time system, there’s also large scale AI training and all that.
James Peng: So just from a sheer technical point of view, it’s such an amazing and challenging thing to work on. So I think those two reasons propelled me to start the company.
Grace Shao: There’s definitely a lot to unpack there. I think later on we can definitely double click on the hardware, software integration, as well as the safety concern there. You say that autonomous driving is much safer than humans. For sure, it’s safer than me driving. I know that. But some may argue otherwise. So let’s talk about that later. But first, I want to ask you about something that was quite interesting. During 2020-23, there was a bit of a public reckoning, I think, within the industry. A lot of peers folded during that time. People decided to pull out of this sector. Some people worried that autonomous driving would really become a reality. But you guys charged ahead and you really believed in your vision. Tell us about that period and how maybe that changed your vision or your growth mentality.
James Peng: I think 2020-23 was a period of time where the autonomous driving industry has evolved for roughly 10 years. I think that was the time of reckoning. That’s the time where the haves and have-nots have really diverged. So I think that’s actually exactly the time. As a company, we have seen tremendous progress. At the end of 2022, beginning of 2023, that was the time we actually finally had the first fully driverless commercial applications operations on the road. So because we made such progress, both from a technical and also from a regulatory point of view, that, of course, we made the progress. We finally see the glimpse of hope. Then, of course, we charge ahead. I think a lot of the other companies who weren’t able to, either from a technical point of view, or from a pure capital-raising point of view, or from a regulatory approval point of view, that weren’t able to have fully driverless applications. Then they were faded away.
James Peng: So it’s sort of like, well, everyone is in school. There’s no big difference. But after graduation, then there’s haves and have-nots. So I think that was the time of division.
Grace Shao: Yeah. So speaking of milestones, I want to kind of go back into history a little bit. So in 2021, Pony.ai had the third highest number of miles driven behind Waymo, Cruise. In 2022, Pony.ai became the first autonomous driving company to get a taxi license in China. In 2023, Pony.ai was licensed to operate robotaxis in Guangzhou, etc. And expansion continued. So kind of following what you just said, there was good momentum behind you guys. Now, today marks Pony.ai’s 10th year officially. You kind of talked about how you guys have grown. But what was one or two of the biggest milestones that you’re really proud of looking back now and that you think have really set the tone for your company Now as you are really expanding globally?
James Peng: Yeah, I think in my view, the biggest milestone, actually, I have already mentioned, is the end of 2022, beginning of 2023, where we were granted the fully driverless commercial license in both Beijing and Guangzhou. We start to have the operation to the general public. Actually, it was in mid-January in 2023 that I was the first road in our commercial robotaxis operations in Beijing. Surprisingly, it was exactly on that day, it was snowing in Beijing, and I was in the vehicle by myself. That was the moment where I actually saw our vehicles were able to drive by itself. Anyone besides me in the vehicle. Because of the snowing, it was also a very challenging scenario. We were actually not being suspended for operation. We continued to operate, and I was in there. That was the moment. Finally, it felt like a dream come true, right? Finally, it’s not just because our technology is ready.
James Peng: Also, because we actually got the approval from the government to have the license to operate. So it’s like all the seven plus years of efforts finally pays off. To me, that was felt like, as Lyndon Johnson said, the small steps for a person, but a giant leap for the human race. Although I wouldn’t call it as big as the Apollo, but to me, it felt like it’s finally from zero to one. So I think that was a deciding moment or defining moment for Pony.ai.
Grace Shao: That’s a personal Apollo moment. I love how you visualize it because I could just imagine how chaotic the roads were in Beijing. Also to be quite romantic when Beijing is snowing because it’s such a beautiful city. Okay, but let’s talk about what is a robotaxi and how the public actually even felt about it when it first rolled out. Before we started recording, Ivy was even telling me, I was like, hey, look, I get a bit scared when I see Waymos on the roads or Pony.ai vehicles when there’s no one Driving behind the wheel. Now, that’s because I’m not used to it. You said, oh, yeah, it’s okay. People get used to it eventually, right? But let’s look back at 2022 when it first was deployed to the public. What’s the public’s reaction?
James Peng: I think because it was a gradual process in the operational domains, in the operational zone that we had. We used to have a safety operator behind the wheel, although the driver actually didn’t touch the wheel or push the pedal. But people gradually get used to it. Actually, at the very beginning, when we were just deployed in Guangzhou, in those days, if you look at the picture of our first and second generation of Autonomous driving vehicles, you still see those spinning LIDARs on the top, and they were very much visible. People were curious. But gradually, people are just getting, this is like business as usual. As a rider’s point of view, the experience of a robot taxi is exactly like a typical taxi. The only difference is there’s no driver inside the vehicle, right? So the way you get the vehicle, the way you get in and get out is exactly the same.
James Peng: Also the other traffic participants, like the pedestrians and cyclists, they get used to it. So I think it just takes time. It’s just like the first cell phone comes out, the first real smartphone comes out. People were very curious. Now it’s just, nobody cares about it. So I think it just takes time.
Grace Shao: It normalizes eventually, right? Absolutely. I think we’ve really had a few years of consumer education done by quite a few of the players, including yourselves. Okay, well, let’s talk about the technical side of things. For outsiders, people might not understand the nuances between L2 and L4. Increasingly, we’re getting closer to L5 supposedly, are we? So help us understand your thinking on there. How do you structure your own teams, your products, working on different technology? Who gets held accountable for the actions in an L2 vehicle versus an L4 vehicle? Then finally, are we getting a glimpse into the future of L5? Are we going to be able to complete the road anywhere we want with autonomous vehicles? It’s a big, broad question, but I’ll throw it to you.
James Peng: Yeah, sure. So the definition of the level of automation for vehicles was actually defined about 20 years ago. So, of course, the industry evolved quite a bit. I don’t think that the levels from L0 to L5 might be the right way of defining what the level of automation is. So in my opinion, actually, there are two different products. One is what we call the driver assist systems, ADAS. The other is fully driverless. So in a broad sense, I think there are two categories. There are definitely two different products. The biggest difference is not on the technical side, but rather, as you mentioned, probably on the regulatory side, is who is first in line for the responsibility if there is ever an accident. I think for any ADAS system, any driver assistance system, it’s always the driver behind the wheel that’s responsible. Regardless if he or she is looking at the road or has their hands on the wheel.
James Peng: Whereas for the fully driverless systems, it’s the system that’s first in line. Because of that requirement, right? Think about if there’s a driver behind the wheel, it sort of serves as a safety net. So the system does not need to be bulletproof. It’s probably, well, as long as it can handle 99%, the case is probably good enough. Whereas for fully driverless, it has to be dealing with all the edge cases, all the extreme cases, and have a fallback system. We can get into those details later. But essentially, in my opinion, there are two different products. Of course, for the driver assistance systems, there are different levels, right? You can be, say, only highway or there’s only keeping in lane. Or they will actually even be able to handle some of the automations in the urban environment. For the fully driverless, of course, as you mentioned, there’s L4, L5 in a traditional definition. L4 means in certain areas. It can be fully driverless. L5 is everywhere.
James Peng: But I think it’s never a clear division. Essentially, you can think of it as how we drive, right? We start with the area, then we gradually improve. Eventually, it will be everywhere. So I think that will be a gradual process instead of a clear division.
Grace Shao: So actually, I want to double click on what you just said. So then help me understand, what is the gap between L4 and L5? Right now, Pony.ai is at L4, right? They’re robotaxis. Is that correct? How am I understanding this?
James Peng: No, I wouldn’t call them a gap. I think it’s a different product definition. Because they serve different purposes. I think most people view this as a process of evolution, right? From L0 to L2, L3, L4. But it’s actually a wrong way of looking at it. As I already mentioned, because the clear difference is that who’s first in line with responsibility. That’s decided by regulatory, actually. By product definition. By regulatory as well. So because of that, it’s essentially two different products. As the product is getting more and more mature, getting more powerful, in my personal view, the division of two different products is getting wider instead of narrower.
Grace Shao: Okay, then I’ll push on this. Then what is the real bottleneck right now for companies like you to deploy at a faster scale? Or to go into more cities quicker?
James Peng: I think that’s the reason that I wouldn’t say it’s one single blocker or one single bottleneck that prevented us to grow faster. I think it’s because the sheer complexity of the autonomous driving and what entails to ensure safety. There’s regulatory, there’s technical things. We also, because this is such a brand new system, that we need a manufacturing capacity. We need deployment. We need to get all the operational things ready, like all the garage space and whatnot. Also user acceptance, user education, as we already mentioned. I think all those take time.
Grace Shao: I believe also we have different partnerships with different managers of your local fleets. That kind of know-how also takes time for them to understand, to transfer over, right? For them to manage robotaxi fleets versus human fleets.
James Peng: Absolutely, absolutely. All those takes time.
Grace Shao: So I want to bring it back to technical. You have said publicly that you use the least compute footprint to reach L4. I thought that was quite interesting. Help us understand how you achieve that. How the model on the vehicle versus the large model you train in the labs actually work together.
James Peng: I think all the AI systems more or less take the same approach, is that you have data on the backend, on the data center side. You train a large model where you essentially try to get all the cases to be learned. In our case, we use the word model, where you can think of it as a simulated city, where we train the virtual driver and let us drive on all different kinds of roads and learn the driving ability. So that’s what’s condensed as a model from all the learning that we deploy on the vehicle. In the traditional AI sense, that’s called edge computing. You put it on the edge, put it on the devices, and put it on the car, where it’s a much smaller model. In a human sense, it’s like we learn everything. Then when we go to a test, we don’t need everything. We just need to be able to have the ability to handle the test.
James Peng: So that’s typically the training, where the backend system needs a lot of computing, but on the actual usage side, you don’t need that much computing power. So when the car is running on its own, it’s actually only using the model on edge, essentially.
Grace Shao: Absolutely. I see. Okay, so let’s talk about AI systems, because AI systems for language, images, code have improved dramatically. There’s also obviously a lot of hype right now around world models, but what you’ve been describing actually has been something that’s not been coined world models for a decade, over a decade. What has generative AI done for you guys? How has it changed how you view your own AI system? Do you think, I guess, the word world models do your system justice in that sense?
James Peng: Yes, it’s a little bit different, and they’re also related. Again, use human as an analogy. It’s actually quite similar to how we think, right? Because think about the large language model, how we process image, how we process knowledge. It’s sort of related to our memory and our logical areas of the brain. Whereas when we drive, it’s not just the memory and our knowledge. It’s also how we react, how we action, and all that. So the example is, one is related to how we learn a new skill. That’s the language model. Whereas for driving, it’s like how we learn to ride a bike. It’s actually different types of brain, different types of skill sets. That’s why it’s different. It’s not the same AI, because that’s how humans deal with different skills for knowledge.
Grace Shao: So Gen-AI has not affected you, but how do you view the idea of now calling, I guess, the physical AI world, world models? Because you guys have been doing this for more than a decade. That’s kind of my question, I guess.
James Peng: Yeah, that’s why I’m trying to get to it. For language model, it’s related with knowledge, with language, with logic. That means you need to be a very large model. Because think about it, if you don’t know a historical event, there’s no way you know it. So you have to have all the knowledge of a human ever created in your model for it to be powerful. So that’s why a large language model requires a lot of computing power and memory and everything. Whereas for driving, that’s how we learn riding a bike. We don’t need to have a PhD degree to learn how to ride a bike. But rather, it requires a lot of practice and training. That’s what world model is related to, or is assembled like. It essentially is a model where the virtual driver can start learning by itself, to learn how to interact with other cars, cyclists, pedestrians, and then learn the driving skill out of that.
James Peng: So it’s a bit related with the large language model, but it’s quite different. Because it’s related with action, related with manipulation, related with collision avoidance. So that’s the key for the world model.
Grace Shao: So tell us about how you simulate these systems. How do you leverage simulation systems for these edge cases?
James Peng: So essentially, that’s how we learn how to drive. There are several key factors for the world model. One is it needs to be very real. So we call the fidelity. It needs to have high fidelity, means it resembles the real world. Second is everything that moves in the world model, meaning cars, pedestrians, needs to be smart. Meaning that because the thing that’s driving is the interactive process. Our action, because we constantly make decisions in there, our action will affect everybody else around us. So they need to react accordingly. So it’s interactive. It’s more like interactive gaming, where we react with everything else. So that’s the second challenge is all the interaction needs to be smart, needs to be intelligent. The third challenge is how do we evaluate what is a good driving? You can interact and everything. You avoid collision. Is that a great driving? No. Not enough, right? Because there’s a passenger inside. Comfort is important. Efficiency is important.
James Peng: From A to B, we want to use the minimum amount of time. So essentially, it’s a multi-metric evaluation system in there to see what is a good driving. So there are three key challenges for the world model. We certainly, all our effort developing the world model related with that three areas.
Grace Shao: But human drivers can be so emotional, right? Either you can be scared or you can be rage driving. Or we can be communicating sometimes without obvious signs, right? We’re looking at each other. We communicate with eye contact, hand gestures. How do you train your fleets to understand human behavior right now? Because obviously, human drivers are still the majority of drivers on the road today. In your case, I actually think you’re right. At one point, maybe removing all the human drivers will make it even safer, right? Especially removing drivers like myself, if I say it again. But yeah, how do you actually help these cars understand all these non-obvious signals? Not someone quite directly clashing onto you. Someone forgetting to turn on the turn sign. Someone stop sign looking at you, waving to go.
James Peng: Like all the nuances. Absolutely. See, that’s why the first thing is how we become a better driver. Essentially, it’s a continuous learning process. The first thing, that’s how we learn how to drive, right? The first thing is you avoid collision. You were a cautious driver. Then gradually, you start learning a bit of everything else. All the signs, all the nonverbal cues, and the hand gestures. So that’s exactly the case for us as well. The earlier model of our system is just driving and try to avoid collision. Then gradually, we put a lot of new things, new recognitions, new perception models into our system where we start recognizing, for example, the hand gestures, especially all the policemen, all the typical police gestures, stop, Go, and all those things. Then we start recognizing, for example, potholes on the road, small obstacles on the road. So it’s sort of how we learn. We start getting all the big pictures first.
James Peng: Then we start learning all the nitty-gritty details down the road and put them to enhance our system. Regarding the second point, you’ll see, when all the cars are autonomous driving by themselves, it will be easier to drive. Yes or no? Because the thing that the big, especially in China, in the roads in China, the biggest challenge is not the other vehicles. In a lot of cases, it’s cyclists and pedestrians. While we can’t make them to be autonomous driving, so I think having the ability to recognize pedestrians, recognize the intention, their sign, and give You a specific example on a crosswalk, the way pedestrians look at and how they pay attention. For example, if they want to directly cross, they typically look straight. But if they were looking back, that means they will more likely not to cross the crosswalk. So we actually take those cues to decide whether we let them cross or proceed straight ahead.
James Peng: So a lot of those details need to be put into the system to make it safer and at the same time efficient.
Grace Shao: Is the judgment made on the spot using the cameras?
James Peng: Yes, using all the sensors. Cameras and LiDAR provide the sensor input.
Grace Shao: We take it all and then we make the comprehensive decision based on the input. Definitely China has more complex and less predictable driving conditions, especially given the number of pedestrians, cyclists, motorcycles we just talked about. So if you can drive safely there, I bet you can drive safely anywhere. But jokes aside, it’s really interesting because, we talk about as your fleet grows, you accumulate more and more world data, real-world data. Is that kind of data eventually becoming an advantage and a serious edge for incumbent fleets and a structural barrier that makes it very hard for new entrants to compete then?
James Peng: Data is important, but data is not everything. So how we understand that is this. Probably give you an example. Think about how we learn. Let’s say we learn math, right? You can think of the data is like the practice sets that we have. Of course, you need to do enough of practice to be a good knowledge about the subject. But doesn’t mean that you have the problem sets of the whole world that you become math experts. So that’s exactly the same case. We need enough of data sets in order to know what the real world driving condition looks like. But we don’t need everything because once we know enough, we can always generate enough knowledge about the driving. So in a way, I think the driving data is important, but it’s not everything. So that’s exactly how you view this.
Grace Shao: So as we speak of this, how do you view the whole landscape right now? Who would you say are your biggest competitors globally? How do you view the different markets playing out?
James Peng: Yeah, that’s a very complex problem. I think a question to answer because I think that I think the first and foremost, I think maybe I got some premises on this. First, the whole mobility industry, especially related with autonomous driving, is very large. They certainly have enough room for several players. Second, it’s still at a fairly early stage for the fully driverless. I don’t think the landscape is already divided in the set. So giving that two promises, I think currently, when I look at the players, I have to judge their current deployment. Although everybody can say, oh, they will have, they will, they will have thousands, hundreds of thousands of vehicles on the road. Actually, giving the current situation, I use the metric as having fully driverless commercial operation as a baseline. Giving that as a factor, I think in the US, Waymo is definitely leading the way. Because Waymo already have 4,000 or 5,000 vehicles on the road, 4,000 plus.
James Peng: Then, of course, there are some other players trying to play a catch up. Zoox, Cruise, maybe Tesla as well. So there’s, of course, some. So I would say in the US, Waymo is leading the way. There’s three to five players trying to play a catch up. In a global sense, I think from a technical point of view, China’s player is certainly on par with the US players. But from the total cost or the economical sense of a vehicle, for example, our vehicle is four or five times cheaper than Waymo’s vehicle. So in the global markets, such as Europe, such as Middle East, I think we will play a huge edge compared with the US players. Certainly, the whole landscape is still evolving.
Grace Shao: But especially in the global markets, you’ll see we will definitely not play a catch up, but taking a leading position. You’ve been an advocate for hardware optimization, software optimization, battery solution optimization. Is that the strategy behind being able to have a vehicle that’s four to five times cheaper than Waymo’s? Or where’s the edge? Or how are you building these comparable vehicles at a relatively cheaper cost?
James Peng: Yeah, I think as you mentioned, you definitely mentioned the most important factor to have the much cheaper price on the vehicles, which is optimization on Hardware, software, and everything else. I think another reason, of course, is because the whole ecosystem related with autonomous driving in China is relatively mature, and the scale is larger. So that price is cheaper. For example, the vehicle itself, the sensors, they’re relatively cheaper in China than anywhere else. Because of the ecosystem, because of the scale. So that plays an important role as well. That touches on something. A lot of physical AI, a lot of robotics companies are also now leaning into the Chinese supply chain. A lot of your peers, actually, autonomous driving, or even the EV players are now looking to expand into physical AI, whether that’s robots, humanoids, Quadrupeds, whatnot.
Grace Shao: So you’ve stayed really focused. You’ve not launched any robots out there or anything. What’s your thinking behind this?
James Peng: Yeah, absolutely true. I think autonomous driving definitely is probably the first large application of physical AI. All the others, humanoids, robots, and everything else, probably will have real applications down the road. For us, we view the autonomous driving as our brand and partner. Of course, as I mentioned, this is still early stage. I think we still have a lot of mileage to go. For all the other physical AI applications, we don’t have any specific plans yet. But I think they’re definitely interrelated. We may enter them down the road, depending on whether we need it or not. Because my judgment is that for the physical AI, it probably will follow the similar trend as autonomous driving. It might take another decade for it to mature. I think for us, it’s more like whether we have real applications for it. Give you a specific example.
James Peng: Even for our fleet, once we go to hundreds of thousands, millions of vehicles, how we maintain those vehicles, how we do the charging, cleaning, servicing, They may use robotic applications. So I guess my view is that we will not probably do robotic actions just for the sake of doing it. But we may do the related applications when we see the real applications.
Grace Shao: So it’s fair to say you’re cautiously optimistic that there is a potential use case further down. But it’s nowhere close to where it’s been hyped in the three to five years kind of use case.
James Peng: Yeah, I think it’s the same thing as autonomous driving 10 years ago.
Grace Shao: All right, well, let’s talk about your international footprint. You mentioned earlier, you have a global strategy. You’re in Europe and Luxembourg was your first launch, right? You’re in Southeast Asia, parts of East Asia, you’re in the Middle East growing really fast over there. Tell us about how you think of your next steps in your global expansion.
James Peng: Yeah, I think the mobility demand everywhere is the same, right? There’s a strong demand across the globe. But we have to focus on the most important markets first. I think eventually we’ll go everywhere, because that’s our motto is we have autonomous mobility everywhere. That’s our ambition when we started 10 years ago. But our first launch, we have several criteria. One is related with regulatory, right? It needs to have relatively accommodating regulatory environment. Second is it needs to be a relatively mature mobility market. In a more obvious sense is that the local taxi fares needs to be relatively high, because I think that our pricing anchor point is always a human driving Taxi. So that price needs to be relatively okay. The third criteria are that we need a good local player to partner with, because a lot of other things like regulatory, like the back end services needs to be Handled by the local partners.
James Peng: So judging from that three categories, I think Europe, Middle East, Southeast Asia, Japan, South Korea, Australia maybe. Those will be probably the potential markets for the initial launch. Of course, those are already big enough of number of countries. So we’ll pick and choose some to start with.
Grace Shao: How do I understand your partnership models? Because I believe you’ve quite a few different kind of models depending on the location, the regulatory environment, potential partnerships, know-how, etc. Tell us about that.
James Peng: Maybe I’ll take one step back first. Think about what is a robotaxi industry. The type of players, I’ll divide them into four categories. One category is for the user acquisition. Those are ride-hailing applications. Those are the Ubers and the Lyfts and the DDs alike. The second is a vehicle. You need a car, how you manufacture a car. The third is a driver. The fourth is all the back end services, cleaning, charging, servicing, insurance, and everything else. For us, our main job is creating a virtual driver, is making a really safe, efficient driver. So that’s definitely what we do. All the three other categories, we might have partners, we might do ourselves. So that, depending on the market, depending on what’s the strong local players, we might pick and choose players who’s handling one or two or three of the Other things. For example, we work with the ride-hailing platforms for the user acquisition.
James Peng: We work with some of the back end services who’s providing the parking space, who’s cleaning, charging our cars. We also have OEM partners that work on the cars. So that’s how we view the partnerships landscape.
Grace Shao: So after you deploy, say you send these out to Australia, what happens walk us through that. Because once these cars actually get off the boat and ships and they land in Australia, are they your responsibility or your partner’s responsibility? Do you send an engineer? Do you send your own management? Or do you transfer that know-how and maintenance know-how to the local partners to handle?
James Peng: Great question. That really depends on the different partnerships and different regulatory environment. In some markets, it’s the local player who’s first in line with managing the fleet. That means in those cases, we manufacture the cars with OEM. But then once we ship the vehicles to the local country, Australia, giving you an example, or Singapore, let’s say, then we actually, in those cases, we sell The vehicle to the local partner. It’s like selling hardware. It’s like selling hardware. But we will, of course, have engineers handling the driving because we are in charge of the driving. So all the driving related work will be done by us. But then the user acquisition, the cleaning, servicing, charging will be done by the local partner. So those are one case. But in some markets, we actually ship the vehicle and we apply licenses by ourselves. The vehicle is still on our own book. But those are rare cases.
James Peng: We actually, our preferred model is to have the local partner that handles most of the logistics and we will be the tech providers. We’ll essentially have the virtual drivers handling the driving and everything else is done by the local partner.
Grace Shao: I see. Would you ever view OEMs as competitors in any way? Because right now you’re partnering with them. You’re giving them the software enablement, right? Would they produce their own robotaxis?
James Peng: I think in most cases, in my view, that they probably will be partners instead of competitors. It’s very different because they were mostly on the hardware business. Very few of them will be in the robotaxis business because they’re quite different.
Grace Shao: I see. I see. Something a bit niche is, I know you run robotaxis as well as trucks. Walk us through how you think about that. Why do you guys also have a truck business? What kind of scenarios are they already being deployed in? I believe they’re the heavy trucks and then the light trucks. How do I understand this?
James Peng: Yes. As I already mentioned, think about our business is that all our technology is that we are creating a safe virtual driver. Virtual driver is our core. As a driver, you should be able to drive all different types of vehicles. The two biggest applications for driver is one is for the transportation of human beings and the other is for goods. That’s related with robotaxis and then for all the trucks. Within the logistic industry, there are actually three categories. One is for the long haul, which is typically done by the heavy trucks, the 18 wheelers and whatnot. Then there’s also in-city network, which is the light duty truck. Then there were also the last mile, typically is handled by much smaller vehicles. Our main focus, of course, is on the long haul and the intra-city transportation. On the last mile, we are the providers of the domain controllers, but that’s not the areas we’re working on. So think about we’re creating driver.
James Peng: Driver should be drive different types of vehicles. That’s how we view the trucks versus the robotaxis.
Grace Shao: Usually, I would assume these are like ports, airports, maybe?
James Peng: They will eventually be everywhere as well. We start with ports. We start with some dedicated routes, for example, like a minefield to the local distribution center, those 30-50 mile routes. The reason we start with those applications is because typically it’s mostly because of regulatory reasons. Because the ports and the dedicated routes and those are typically a semi -private road. It’s much easier to get the regulatory approval. Of course, we are working on long haul trucks as well. We already actually have a fleet of heavy-duty trucks doing the real goods transfers on highways, but still with a safety driver, of course. Eventually, we’ll be fully driverless as well.
Grace Shao: You’ve said you have a target of running fleets commercially across more than 20 cities by the end of this year. What do you know today that you could not have learned without actually operating at scale already on the streets? What makes you have the confidence to do that now, I think, compared to maybe a few years ago?
James Peng: Again, I think for robotaxis commercial business to be a reality, there are three important factors. One is technology. Second is regulatory approval. The third is user acceptance. I think within the last three to four years, we have gained a lot of experience on all three categories. The reason we were confident to deploy in 20 cities is because clear vision on the regulatory approval. There’s a lot of cities globally, both in China and in some global cities, they actually start coming out with regulations for supporting fully driverless commercial applications. Also we have planners. Planners want them. So I think all the important factors are falling into place. That gives us confidence.
Grace Shao: I’m going to play devil’s advocate a little bit here. With the rise of AI right now, there’s a bit of a fear of replacement of people’s jobs. The rise of autonomous driving obviously lead to job loss in people who are currently drivers. How do you view that? Because just now we talked about robotaxi drivers. We talked about people driving heavy-duty trucks that could potentially be replaced. Frankly, I’m in a camp that people could be maybe freed up to do more things that they can do otherwise. People will find alternative careers. But are regulators becoming more cautious. How do you feel about the current public pushback a little bit on AI, autonomous driving, autonomous everything at the moment?
James Peng: Yeah. Actually, driving is a hard job. Driving is a lot of cases in a stop vehicle for 10, 12 hours a day. It’s a really tough job. The thing that because autonomous driving itself is a highly regulated industry, the pace of our roll up is determined by the number of licenses. The thing about also a lot of the drivers were not young. The young generation, younger generations actually don’t want to be drivers. So I think, especially a lot of the global markets, we actually come in to fill the gap for the labor shortage for the driver. We’ll not change the human driving vehicles overnight. It will be a gradual process. So that’s sort of the development of the cities and the human society. It takes time. It becomes gradually a norm. Then, as you just mentioned, then the drivers can find other jobs.
James Peng: Even we actually absorb a lot of jobs, for example, for the remote assistance, maintenance, which are much safer and much less strenuous job conditions. So I think society as a whole has always a way to absorb jobs. To adopt, adapt, and then evolve.
Grace Shao: The current pay for a lot of times for these heavy truckload drivers are like 200 to 300k USD. They’re considered very high-earning jobs. But at the same time, people forget they’re extremely dangerous. There’s life lost constantly on the roads. So I can see that could be very valuable if people can actually replace those routes with robo-drivers.
James Peng: It’s not just replacing. Look at the truckers. Their average age is 45 plus. In North America right now? In North America. In China, they’re 40 plus as well. So a lot of younger generations, they don’t want that type of jobs. We’re coming not only to replace, but actually to fill the void for that job shortage.
Grace Shao: All right. So I think I want to wrap up our conversation soon about this. Is there anything I’m really missing, you think, about robotaxis and your business at this point?
James Peng: I think we’ve probably covered a lot of topics.
Grace Shao: Oh, I had one question. Another one about your business before we go into your personal thing. You mentioned Croatia just now when we were talking offline. I thought that was so fascinating. In my mind, I thought these robotaxis were being deployed mostly in futuristic cities like Silicon Valley and SF, out here in Shenzhen where we’re here today. But Croatia, help us understand the need for robotaxis in these countries where a lot of the roads are aged, are not really made for cars to start with, Are not easy to drive in, actually, even for humans. Then how does that make sense even for your economics, actually?
James Peng: Of course, there were some challenges. From a technical point of view, two challenges initially. One is there’s a lot of roundabouts. Actually, there were not many roundabouts in China. So although a lot of other very complex situations like heavy storms and whatnot, we were able to handle them really well. But roundabouts, we had some, but we haven’t trained that much. So we actually have to retrain a bit on the roundabouts. The second is the trams. There were just a lot of trams in the Zagreb. Their behavior of the trams is different from cars. So we need a little bit more training to get used to it. But it’s like how we drive. When we go to a new city, we might not drive as a perfect driver initially. But then we learn and adapt. Once we have a good learning system set up, then we can quickly learn. That’s exactly our experience in Zagreb, Croatia. Two things that we actually have to learn in Croatia.
James Peng: One is the roundabouts. The other is trams. Because those are not something that typically you will see on the roads in China. So for those new situations, it’s like how we learn. How we learn driving. When we go to a new city, we probably know 95%, 98% of the situation. Some of the scenarios probably we didn’t encounter previously. Then we learn. We adapt. So that’s exactly the case for us in Croatia. After three to four months of learning and training and retraining, we actually were able to handle those cases like roundabouts and trams really well. Because there’s a lot of roundabouts in other cities in Europe. They actually have different rules for roundabouts. Some of the roundabouts, I think the cars outside roundabouts have right-of -way. Some of the vehicles inside the roundabouts have right-of-way. But we can adapt once we have the system set up.
James Peng: So as I mentioned, the most important characteristic of our system is not how powerful it is, it’s how adaptive and how easy to learn on our system so that We were able to adapt.
Grace Shao: Brilliant. So a lot of localization as well for your vehicles. I have two last questions. One is, what is something you think people still get wrong often about your sector, in this case, autonomous vehicles, autonomous mobility? The second question is a bit of a curveball. I’ll throw it to you first, you can think about it. What is one differentiated view you hold? Something that’s a bit against consensus, maybe.
James Peng: Autonomous driving industry, I think people put too much focus on technology and probably underestimated the complexity of robotaxi as a business. Essentially, of course, technical is the most important. If you can’t drive safely, you’ll not have a business. But once you even have the most safest driving, you still have to, as a business, there’s a lot of other things involved. For example, how you deploy a fleet, how you make the pickup and drop off easy for the user, how you handle all the edge cases of the complaints of the Passengers, how you make the charging, servicing, cleaning efficient. For example, especially give you a specific example, the electricity fares during the day fluctuates. If you have the charging at the low fare, you can save a lot of cost. Then how you manage your fleet? Although you have the low fare for electricity, but the demand of the passengers is really high. How do you make a decision?
James Peng: So essentially, it’s a lot more optimization involved than just the driving itself. I think a lot of people underestimate the complexity with the management of a fleet of autonomous driving vehicles. We actually, as a company, have put a lot of emphasis and take a lot of efforts in optimizing everything. So that’s why I think those will be a very strong competitive edge down the road.
Grace Shao: Once you guys scale further, especially.
James Peng: Exactly, absolutely. Very interesting.
Grace Shao: The second one, I’ll put you on the spot again. What is one differentiative you hold?
James Peng: I think I’ll take the one related to the answer of my first question. Is that, again, people always put too much emphasis or give too much credit on zero to one and think about less for one to ten. Give a lot of examples, right? People always think an invention is so hard, but putting an invention to be a scaled application is equally hard or a lot harder. Because the scale involves cost optimization, involves user education, involves a regulatory approval, it involves making the things a lot easier to use. So many examples like this, right?
Grace Shao: Definitely. Say a rocket is put in the sky. Oh, it’s so hard. But having the rockets to always be able to safely take off and recycle, that’s extremely hard.
James Peng: So I think related with autonomous driving is we certainly crossed zero to one. I think we crossed one to five, maybe. But from five to ten, ten to a hundred, I think there will be still a lot of challenges ahead.
Grace Shao: That’s very insightful. I agree with you. When we look at the internet era and a lot of players that still stand today versus who are the actual ones that created a lot of the internet use cases we know of today. Thank you so much, James. It was a pleasure and an honor to learn more about your business, yourself, the man behind the company that is changing the future of autonomous mobility. Thank you again.
James Peng: Thank you for having me.
AI Proem is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.



Get full access to AI Proem at aiproem.substack.com/subscribe

In this episode, I spoke with Gary Ngan, CFO of Meitu, about how the company is evolving from its roots in consumer photo editing into a broader AI-native visual creation platform across photo, video, design, and agents. For many investors, Meitu is still associated with beauty editing and selfie apps, but Gary frames the company today as an AI application company serving both leisure use cases and productivity workflows.
We spent a lot of time on Meitu’s business edge: why visual AI is not just a foundation-model race, and why aesthetic judgment, controllability, and vertical context matter. Gary argues that visual creation is highly subjective. The same prompt can mean very different things across countries, cultures, product categories, and commercial goals. That is why Meitu is building verticalized products such as Picchi, DesignKit, Kaipai, Vmake, and RoboNeo, instead of relying only on one general-purpose AI model.
We also discussed the business model. Consumer subscriptions have become Meitu’s main revenue engine, while advertising is no longer the strategic growth driver it once was. Gary explained the shift in Chinese consumer willingness to pay for apps, the higher ARPU potential in overseas markets, and how new AI-native products like Picchi could introduce additional monetization through personalized models and AI credits. He also addressed AI compute cost, why more than 90% of Meitu’s AI outputs come from its own models, and why the company sees AI as a TAM-expanding opportunity rather than simply a margin risk.
Finally, we covered competition and globalization. Gary explained how Meitu thinks about competing with ByteDance, Kuaishou, Canva, Adobe, Shopify, Alibaba, and other AI-native visual tools, and why Meitu’s approach is more vertical-driven than general design-platform driven. Lastly, we touched localization, from different beauty preferences across markets to why true globalization requires understanding culture at a much deeper level than translation or marketing campaigns.
CHECK OUT THIS CONVERSATION. Gary’s so cool.
To find the previous episodes of Differentiated Understanding, see here.
Every episode, I bring in a guest with a unique point of view on a critical matter, phenomenon, or business trend—someone who can help us see things differently.
Season two will host a series of guests from analysts, VC investors, builders, researchers, founders, and product managers. For more information on the podcast series, see here.
Chapters:
00:00 What is Meitu today? Mapping Meitu’s product portfolio04:14 Why vertical focus still matters in the age of AI agents06:36 Aesthetic standards, subjective prompts, and visual AI nuance11:31 How AI changes art and creative expression15:02 MeituHub and MiracleVision as visual AI infrastructure17:01 Why Meitu needs its own models20:55 How Meitu chooses models and the role of designers25:09 Meitu’s AI legacy and generative AI strategy28:27 AI compute cost, ROI, and gross margin38:22 Subscription growth and advertising dependence40:47 Partnerships with consumer chatbots and platforms43:27 Competition with ByteDance, Kuaishou, Canva, Adobe, and others49:49 Deepfakes, misuse, and AI safety safeguards52:42 Globalization, localization, and cultural differences59:16 The biggest investor misconception about Meitu
Transcript (AI-generated, for reference only)
Grace Shao:Gary, thank you so much for joining us today.
Gary Ngan:Hi Grace. Good to be here.
Grace Shao:I’m really excited to have this conversation. To start, tell us what Meitu is up to these days. For a lot of investors and users, when they think of Meitu, they still think of the selfie and beauty-editing app. How would you define Meitu today? Is it still simply a consumer AI company, or is it much more than that now?
Gary Ngan:Meitu is no longer just a selfie or beauty-editing company. I would define Meitu today as an AI application company specializing in photo, video, and design.
We focus on very high-value verticals where we can leverage AI to deliver high-quality results to users. We often refer to these users as prosumers: people who have strong design needs, but no prior formal design training.
So that is how I would define Meitu today.
Grace Shao:That makes sense. Tell us about the products, because you have quite an array of them. Some are more consumer-facing, some are more prosumer-facing, and some may even be a bit more enterprise-facing. There is Meitu, BeautyCam, Wink, Picchi, DesignKit, Kaipai, Vmake, RoboNeo. Help us map out the ecosystem.
Gary Ngan:We think about Meitu’s product portfolio in two main categories: applications for leisure and applications for productivity.
Applications for leisure include the Meitu app, BeautyCam, Wink, and Picchi. They serve use cases such as photo-taking, photo editing, and video editing, usually for sharing on social media.
The Meitu app and BeautyCam are our core consumer applications. Wink extends our capability from photo to video editing. Picchi is our latest portrait-retouching agent, focused on personalizing editing styles.
The second bucket is applications for productivity, which includes DesignKit, Kaipai, Vmake, and RoboNeo. These products serve professional and commercial content creation needs.
DesignKit focuses on e-commerce product-listing design. It helps merchants and creators produce product images, model images, and marketing materials much more efficiently.
Kaipai and Vmake focus on talking-video and marketing-video creation. Kaipai is more focused on the domestic Chinese market in verticals such as insurance and real estate, while Vmake is seeing strong traction in the U.S. fitness and wellness market.
To give you a sense, as of May this year, Kaipai had about three million monthly active creators, and cumulative content creation exceeded 400 million pieces. For Vmake, ARR in the first quarter of 2026 was about US$4 million.
Then there is RoboNeo, our AI-native agent product launched in July 2025. It is currently targeting the AI short-drama vertical. Its agent workflows can support scriptwriting, characters, storyboards, visual generation, and asset management.
So that gives you a rough idea of the different vertical products. But the ecosystem logic is very important, because many new products come from user insights we observe in existing products.
For example, DesignKit came from the poster-design function within the Meitu app. Kaipai came from the AI teleprompter feature in BeautyCam. Picchi came from new user behaviors we observed in the Meitu app.
So our portfolio is not a random collection of apps. It is a structured expansion from consumer imaging into AI-native workflows across photo, video, and design.
Grace Shao:That makes a lot of sense. But in the age of AI agents, would it make sense for Meitu to consolidate a lot of these apps? Or do you still think it is better to keep them separate for different types of users and workflows?
Gary Ngan:In the age of agents, we still believe we should focus on high-value verticals, because different verticals have many differences.
First of all, aesthetic standards are very different. I’ll give you an example. The phrase “handsome guy” would be interpreted very differently in an application serving the U.S. market versus an Asian market.
Even within the Asian market, if you are addressing e-commerce merchants selling gym products versus formal apparel, the word “handsome” will also be interpreted very differently across those verticals.
So being able to separate these different verticals gives you a very good head start in focusing on the aesthetic standards that each vertical needs.
Also, users in different verticals have very different behaviors and workflows. It is very important to build those workflows and that know-how into each vertical in order to create the right products.
With agents, you can cover a slightly bigger boundary. But I still think you want to focus on different verticals to maximize the output for the user, and also make it more efficient and easier to market within each vertical.
Grace Shao:That is really interesting. You touched on something that a lot of people discuss when they think about visual AI, which is how to ensure consistency and accuracy when translating language into visuals, especially when text can be in different languages and words can be subjective.
As you said, if you say “handsome” and I say “handsome,” that could mean very different things in our heads. How do you ensure that identity, description, and nuance are not lost? You mentioned vertical focus, but what is the technical side of that?
Gary Ngan:Instead of calling it fragmentation, I would say vertical focus is very important. That sets the tone.
Behind that, we also have a large team of designers who control different points in the model fine-tuning process. They help set the right direction for the aesthetic standards within each vertical. That is something differentiated in our product offerings.
Then, if you move one step forward, the data flywheel is also very important. Users within a vertical give us data through their behavior: which photos they use, which photos they edit, which ones they throw away. That is very important for us to improve image creation.
We are also in the camp that believes controllability in visual applications should not just come from AI. You still need manual touch-ups at the end for users to make last-mile improvements, because aesthetic judgment is very subjective.
Even if an AI model works with you every day, you will always have subjective comments and small edits you want to make.
One other interesting point is that when we talk about aesthetic standards, in the case of leisure products, the face is usually yours. So you have a strong say and a strong sense of what is good for you.
The way we learn that is by studying the trend in your geographic location, giving recommendations, letting you try them, and then as you use the application, you tell us what is most suitable for you.
Picchi is a newly launched app where you can upload three to five sets of original photos and edited photos that you have done yourself. We are then able to learn that pattern and create a specialized model for you. The next time you want to edit a photo, you can call up the model that you trained yourself and apply your own aesthetic standard to your photos.
On the productivity side, however, aesthetic is not the ultimate holy grail of an image. It is very important, but whether that photo or video is effective in driving conversion, likes, or comments is also very important.
When we deliver images and videos to users, we take into account key data from that particular vertical and the metrics that matter for results. It is not just whether someone is subjectively handsome. It is whether this person, image, or video can help sell your product.
So the two camps are quite different.
Grace Shao:That is really interesting. From a pure consumer point of view, you pointed out an important nuance. If I upload my own face to a Meitu product, it might give me very smooth, pale skin and a more angular jawline or chin. But if I use an American fine-tuned product, it might give me more contouring. It is a very different aesthetic.
But to your point, if you are a prosumer, a content creator, influencer, or e-commerce seller, then it is not only about whether the image looks good. It is about whether it drives sales.
I want to ask something slightly more philosophical before getting into the businesses. Technology often changes art. Photography changed painting. Software like Final Cut Pro and Photoshop changed photography and video. How is AI now changing how visual artists approach their vision and craft?
I have also spoken to companies like Kuaishou, where they have Kling and are partnering with AI-native film studios. These people are creatives, but they do not view AI as disrupting their work. They use AI as a tool to create their work. What do you think about this at a high level?
Gary Ngan:If you look at our core value proposition, our mission statement is uniting art and technology.
One step down from that, we are trying to democratize design, art, and creative expression.
What AI changes is that it enables people who have creative ideas, but not the actual training or skills, to express those ideas.
A lot of the time, we have creative ideas that we want to express, but our motor skills are not refined, or we do not know how to put colors together. AI can help us deliver those ideas.
That is fundamentally changing the artistic landscape to a certain extent.
Other companies may say that existing professional filmmakers and designers can use AI to make things more efficient or create things in a different way. That is great. But I think the bigger impact on the world is enabling many people who previously could not create anything. They had ideas, but could not express them. Now they are able to express them.
That is what is fundamentally changing the industry.
Grace Shao:We have talked about how you have many different products, and you explained that they are targeted at different verticals. How should we understand MeituHub and MiracleVision? Are they the operating system underneath everything?
Gary Ngan:MiracleVision and MeituHub are the visual, image, and video infrastructure that we have. Our applications are built on top of these things, so they go hand in hand.
We are still fundamentally an AI application company, but we also need visual infrastructure.
To give you an example, over 90% of our AI outputs come from our own models. There are many situations where we think the models we fine-tune ourselves perform better than third-party models. There are things that other people do not necessarily focus on, so we have to invest in R&D and create that infrastructure ourselves.
MeituHub is also a way for us to export that technology. People can use our APIs and skills to build their own applications or integrate them into their own systems. That also reinforces our vision of democratizing design.
Grace Shao:That is a perfect segue to my next question. I understand your team fine-tunes your own models, but you also build on various open-source models. Why does Meitu need its own model?
Traditional application companies often did not need to own the foundation layer. So why does Meitu need that? And more broadly, why are so many Chinese consumer internet companies pushing out models? You see even companies in food delivery, ride-hailing, and other consumer internet sectors releasing models. Is this a cultural push, or something else?
First, how does Meitu think about it at the company level? And if you can comment, how do you view this competition across the China ecosystem?
Gary Ngan:It is harder to comment on the overall market, because what we do, visual image and video models, is quite different from language models. So I will focus on why we do our own models.
Our belief is that one general model will have difficulty performing well across all verticals, because context is so important.
If we do not have our own models, then aesthetic standards will be set by third parties. When a third party creates a model, they have their own idea of what aesthetic standards should be. They have their own idea of what should be generally good given a certain prompt word.
But that may not be applicable to the verticals we are working on. That is why we need our own models to serve those purposes.
At the same time, we integrate third-party models because even within a vertical, there are corner cases or edge cases that our core model may not be optimized for. In those situations, we call on third-party APIs to serve users.
As an AI application company, the only point of optimization is user satisfaction. We use a combination of our own models and third-party models to serve that purpose.
Sometimes we see more and more users calling third-party APIs for similar prompts or similar creation scenarios. Then we will augment our models to cover those scenarios as well.
Our model is continuously growing, but we make it very vertical-driven. We have told the market that we are not in the business of creating a general-purpose model. We are creating vertical models. But that does not mean we are giving up model training altogether.
Grace Shao:So there is a lot of industry know-how in each vertical that you have.
When it comes to which foundation model you choose for each task, how do you make that decision? I spoke to one of your colleagues at SuperAI, Rocky, your VP of R&D. We discussed the fact that you use a series of open-source models and also work with different model providers. What is the main factor in choosing which model to build on for which vertical? How do you delegate tasks across models?
Gary Ngan:At a high level, there are two main forces behind that.
One is user behavior. If a user uses Model A to create a certain task, and many users do not press save or do not continue working on it, then we probably need to serve that task with a different model. It is a data flywheel type of operation.
The other factor is our large team of designers, who are very involved in training these models. Designers help set the standard for what the right model should be for a particular task.
This is a very important differentiation for our company versus most technology companies.
I am not sure if you are aware, but our founder and CEO was an art student by training. In his day, he was the top student in the Tsinghua Arts Academy entrance exam for oil painting.
In his mind, aesthetic standards are always very important. Because of that, designers in our company have a very strong say in every product and every feature we launch.
Over more than a decade of designer training, the rest of the company has also developed stronger aesthetic standards. Product managers and R&D engineers also have quite high aesthetic standards now.
Our company is organized toward delivering the best aesthetic standards for users. That is very differentiated from most tech companies.
Most model companies may think: We solved this problem, the photo is done, the video is generated, the main character is stable throughout three minutes, so the mission is accomplished.
But for our designers, apart from the stability of the main character, they also look at whether the lighting is realistic, whether the color fits that vertical, and whether anything feels wrong from an artistic point of view.
Those are the things we really focus on when fine-tuning. That is something we are very proud of, and I think it is a major differentiator.
Grace Shao:Even as a consumer user, I can say your products have that extra last-mile touch-up tool that others often do not offer. It is meticulous and accurate. You can zoom into pores or details in the background. It is interesting to hear about your founder’s background and that artistic legacy, because that culture really shines through the products.
Speaking of legacy, I want to understand Meitu’s AI legacy and strategy. You have been working in image and video for over a decade, so you obviously have a vast database and deep know-how in visuals. How does that industry expertise translate in the age of generative AI and in the future agentic world?
Gary Ngan:Generative AI has changed the speed, scope, and value of what we can deliver.
First, speed. New AI capabilities can now be translated into user-facing features much faster, helping us launch popular effects globally and drive overseas growth.
Second, user experience. Generative AI enables effects that traditional computer vision technology could not fully achieve.
For example, facial and body retouching is no longer just manual adjustment. AI can reconstruct details, lighting, and texture in a much more natural way.
Third, target addressable market expansion. AI helps us broaden into productivity workflows like DesignKit and Kaipai, which were things we traditionally could not do.
Overall, AI is very empowering in helping us get to where we want to go.
Before AI, all we could deliver was better tools. But in order to use those tools, you still needed pretty good aesthetic standards or some understanding of the basics. Otherwise, giving you those tools did not really help much.
With AI, you still need maybe 10% or 20% of that understanding, but the requirement is reduced massively. AI can give you many choices to choose from, and then you can start building from there.
That helps us move from leisure applications to productivity applications. That is really what the strategy is about today.
Grace Shao:AI can act like a guide or mentor if you are new to a certain craft or sector.
Let me ask the spicy question. AI compute cost is obviously extremely high. Image and visual generation are expensive. How does the economics work right now? Does AI compute affect your gross margins, or are you seeing ROI already?
Gary Ngan:As I said, currently over 90% of our generative outputs come from our own models. As long as we are using our own models, the cost is very manageable. Our gross margin is still over 70%.
Also, when you are editing your own face or editing a product photo, these things are not purely AI-generated. You may want AI to edit a little bit, remove someone from the background, or create a new background for a product, but the entire photo is not purely AI-generated.
It is true that AI inference has a cost, but it is not as if every photo now incurs a lot of cost. We need to make that distinction first.
As we move into new verticals, like music videos and AI short dramas with RoboNeo, those are more experimental. We are using more third-party models, so margins on those new applications will be much lower than something like Meitu Xiuxiu.
But as we continue to progress, we will develop our own models to replace some of the third-party costs. Over the longer term, we also believe API costs will come down.
So we do not see this as a threat. In fact, the integration of AI has expanded the addressable market so much that it is a much bigger opportunity than threat.
Grace Shao:I appreciate that nuance. You are explaining that the first type of usage does not use as much AI or token cost as people might expect from the headlines. The second part may be more expensive, but we are still in very early stages.
Let’s take a step back. For some of our American or Western audience, they may not be as familiar with Meitu. How do you fundamentally make money?
In your public disclosures, consumer paid subscribers grew more than 30% year-on-year. What is driving that growth? Is it that the AI features are much better now? Is it global expansion? Help us understand the business model and what is driving growth.
Gary Ngan:Our main revenue source is subscriptions, mostly on the leisure side.
That is our second growth curve. The first one was advertising, but that business has matured.
The second growth curve, which is still growing quickly, is subscription on the leisure side. The main driver has several parts.
The first is China user behavior. Paying for apps really started after COVID. Before COVID, virtually all applications were free. They competed through free usage, advertising, or redirecting traffic to other applications to generate money.
After COVID, many user-facing applications realized advertising was under pressure, and they wanted new revenue sources. Without colluding, many of them started charging users. That kickstarted the user subscription process.
What is less understood is that users then began realizing that applications have to be paid for. As time goes by, the behavior of paying for applications grew on them.
Now there is much less of the issue of, “This app has to be paid, so I am not using it.” That was a real mindset before. Now it is more like, “This app costs 15 RMB a month. Is it worth it?”
That is what I would describe as the beta factor, meaning the overall market. Users are becoming more and more used to paying for mobile products.
That is one reason we are confident that paying subscribers and the paying subscription rate of our leisure applications can continue to grow.
To give you a sense, we have done surveys. The global paying percentage for photo and video applications is about 20%. If you benchmark music and video apps globally versus Chinese equivalents, the Chinese equivalent is usually around half. For example, if Spotify is around 40%, the Chinese equivalent might be around 20%.
So if global photo and video applications are at about 20%, China should at least achieve about 10%. Right now, we are around 5% to 6%. So there is still another 80% to 100% growth headroom there.
The second growth potential is international expansion. In high-ARPU areas like the U.S., Europe, and East Asia, including Japan and Korea, the base ARPU is already much higher than China, anywhere from 100% to 200% higher. The paying percentage can also be much higher.
To give you a sense, one of our applications called AirBrush has over 50% paying percentage in the U.S.
As we launch stronger operations in these high-ARPU countries, we expect our blended paying percentage to grow further.
One final point about monetization is that we are integrating more generative AI capabilities into these applications. For example, Picchi is an application for leisure, but it uses an agent for editing, and that has a completely new business model.
On top of regular subscription, if you want to create your own model to apply your own editing skills, you have to pay for that model separately. That is another monetization test we are currently working on.
Grace Shao:When I was reading your earnings reports, I was a bit surprised that your highest revenue generator is consumer subscription, because the default mindset is that people have very little willingness to pay.
But as you said, whether it is the change in behavior in China, or people having more appetite for premium add-ons or AI-plus features, willingness to pay is changing.
There is also the fact that advertising can be annoying to sit through. You do have a lot of advertising, I have to say. Spotify does too, and I think that drives people to pay to get rid of advertising.
On that note, do you think you will gradually reduce your dependency on advertising? It is still your second-largest revenue model.
Gary Ngan:We have not relied on advertising since 2022. At the corporate level, we made the point that we are no longer strategically trying to drive advertising.
You have seen our advertising business grow at low single digits over the past few years. Advertising is not what we are fundamentally trying to drive.
However, we are experimenting with advertisers on fun and engaging AI-infused campaigns.
It is hard to describe with words, but you can imagine users generating viral photos with a brand advertiser’s branding that fits the brand image. That gives the uploader a lot of likes and gives the advertiser a lot of exposure.
So we continue to experiment with those things. But in any case, we are not relying on advertising for business growth.
Grace Shao:On partnerships, I had this idea and I do not know if you are doing anything like this. Would you partner with some consumer-facing chatbots in China to help them with video and visual capabilities?
For example, could someone go into a consumer chatbot and call up Meitu’s capabilities? There may also be competition there. How do you view your relationship with these players?
Gary Ngan:We are open. In fact, we are already an official partner with WeChat, not on the Xiaochang side, but in another area. I do not remember the exact English name, but basically when you are using the chatbot, you can call up Meitu.
Right now, it is still a lighter relationship, almost like traffic redirection. But our goal is to democratize design. Being able to work with more people and enable more people to access that power to express themselves is something we are open to.
Grace Shao:That makes sense. It feels like they may not want to put as many resources into this specific use case, and you have the know-how in doing the best video and image editing.
Gary Ngan:I would not say they do not have the edge. I think they may just not want to focus on that.
Creating these applications requires a lot of focus. It requires the right organizational structure and a laser-sharp focus on trial and error, and on creating the best aesthetic output for users.
These may not be the things that larger companies want to invest in. It is important relative to our size, but to them it may be something they do not want to focus on. If they wanted to do it, I think they could.
Grace Shao:Let me challenge you a little bit on big tech. In China, ByteDance and Kuaishou clearly have a lot of edge and moat from massive pools of image, visual, and video data. In the West, we have Canva, Adobe, and other global applications. Even Shopify and Alibaba are creating e-commerce staging and design tools.
In this big world of competition, or peers if we put it more nicely, how do you see Meitu’s strength? Who are the most relevant competitors that are similar to what you do? And who may look similar on the surface but are not actually doing the same thing?
Gary Ngan:We have to separate it into two categories.
On the leisure editing side, with the exception of one business unit within ByteDance, there are not many large companies doing that globally. I do not think there is any real large company doing that in the U.S.
There are smaller companies, but they are much smaller compared to us.
On that side, our edge is really continuing to follow and set the trend for the latest aesthetic standards and what helps users stand out on social media. These are the things we have been doing for more than a decade, and we will continue to excel in them.
On the productivity side, there are many companies doing similar things, but taking a much more general approach.
For example, Canva and Adobe use one product to satisfy different verticals. Adobe is organized around media: photos, vector diagrams, video, effects, and so on. Canva is one editor trying to fit many situations. Figma is also one app serving many applications.
They are design-driven. We, on the other hand, are much more vertical-driven.
We are not restricting ourselves to a specific media type. We are saying, within e-commerce, what do you need?
You need product photos. You need very short product videos. You need the ability to generate batches and batches of photos. You need to know the latest trend on the e-commerce platform you are selling on. For that particular product, you need to know the selling points. You also need to know the rules of Amazon or Temu and what you need to abide by when selling those products with pictures.
All these things are baked into DesignKit.
If you are using Canva, I highly doubt it will have a red flag saying, “You should not be using minors in this product photo.” Canva may not even know that you are creating a product photo in the first place.
So these are the different focuses we have.
In terms of competitors, it is hard to say who is a direct competitor, because at the end of the day, you can use Photoshop, Canva, or our products to create an e-commerce photo. They are all peers, but we take different approaches.
If we take a step back, generative AI is still very early. It is 2026 now, but generative AI really only started in earnest late last year for visual use cases. Before then, a lot of generative AI photos still looked AI-generated.
Grace Shao:They were quite bad. There might be six toes, or the face was disproportionate.
Gary Ngan:Even if there was nothing obviously wrong, you would look at the photo and know it was AI-generated. It did not feel real.
Now we are just starting to see things that are harder to distinguish between human-made and AI-made. This is how we can empower the industry and increase efficiency.
We are still very early in this market. That is why we are very optimistic and see a lot of opportunities.
Grace Shao:A little side note: in 2019, when I was still with CNBC, I covered deepfakes. At the time, there were a lot of deepfake videos of Obama or Zuckerberg. A startup even made a deepfake of me. It was literally just plugging someone else’s face onto my head and body, and nothing really worked.
But now, fake images and videos are getting very hard to distinguish with human eyes. How do you view the ethical side? How do you stop misuse of the technology?
Gary Ngan:First, we have put in safeguards.
For example, on Picchi, if you generate a model of yourself using your own photos, that model cannot be applied to anything other than your face. If we detect that it is not your face, we will not allow you to apply that model to another face.
In some of our generation applications, we have also put in safeguards around certain words, such as violence or pornographic images. You cannot generate those using our applications.
So there are safeguards that we put in place. Obviously, we can only do so much.
One thing that makes it slightly easier for us is that we organize our applications into different verticals. Users come into our applications with a very strong intent. They know they are creating e-commerce photos, for example.
Instead of giving them a general chatbot where any random person can come up with a random idea like putting their face onto the President of the United States, it is harder to imagine someone using DesignKit to run a prompt like that.
Organizing into different verticals also helps us mitigate the risk a little bit.
Grace Shao:The last area I want to talk about is globalization and your global strategy. Meitu is globally available. It is interesting because, as you said, you focus on each vertical, and you have not done a big splashy general marketing push. It also feels like that is true geographically. You are in Southeast Asia, Japan, Korea, Europe, the U.S., and so on.
Help us understand global scaling. What have been the challenges? How have you done it successfully? And how do international users from different regions behave differently from users in China?
Gary Ngan:I will answer the second part first. Users in different regions all behave very differently.
Grace Shao:Give me all the stereotypes.
Gary Ngan:Not stereotypes, but I will give you one example.
We were doing a user focus group in the UK and spoke to a male influencer. He said, “Your app can edit my jawline? That is incredible. I would totally pay for it. But I do not think it is a good idea to smooth out my skin.”
Grace Shao:That is interesting. So it is not okay to pretend you have better skin, but it is totally okay to have a chiseled jawline?
Gary Ngan:He did not mention whether it was ethical or not. That was just his feedback, word for word.
The challenge, or the interesting thing, is that we have to really listen to what users want in those markets. Different geographic locations need the right mix of features and marketing campaigns.
I will give you another simple example. A few years ago, we were looking at Lunar New Year. Koreans also celebrate Lunar New Year, and in China Lunar New Year is a festival where we get a lot of usage.
We had launched features in China that were very popular that year, but in Korea there was no uptick. Later, when we had local Korean colleagues helping us run marketing campaigns there, they told us that Koreans generally celebrate Lunar New Year with white clothing and a white theme, while Chinese people celebrate with red.
Our Chinese marketing team was surprised, because in China, white is usually associated with funerals. It did not register.
That example tells us there are many things we need to immerse ourselves in culturally to understand how people behave, what they care about, and what the standards are.
We cannot stereotype anything. Every place and every person behaves very differently.
That is the biggest challenge, but also the biggest opportunity.
Now we are setting up offices in different parts of the world. We are sending product managers overseas regularly to do more focus groups and, more importantly, to experience the lives of the users they are trying to serve.
In the past, we relied too much on consultants, reports, or reading online. That is not enough anymore. We are fixing that, and I think we are making progress in some countries.
Fingers crossed, we will continue to grow bigger in Western markets.
One other tailwind that has helped us is TikTok and K-pop. Back in the day, editing a photo seemed socially unacceptable to a certain extent. But with TikTok, people are more relaxed about filters being applied and playing around with your face. It is no longer as taboo in many Western countries.
The rise of K-pop is also influencing cosmetic styles, and that becomes a segue for us to try different things in Western markets.
There are very interesting things happening. But the most important thing is for us to really understand, live, and breathe those cultures so we can create things users want.
Grace Shao:That is meaningful, and it feels important for a new generation of Chinese companies going global. Localization cannot just be reading headlines or high-level reports. You have to understand the culture, because culture influences the business.
To wrap up, I really appreciate your time. My last two questions: first, what is the biggest misconception investors currently have about Meitu’s business?
Gary Ngan:One of the biggest misconceptions is that general models are going to destroy everything, and that there is no place for AI applications.
We think that is quite unlikely on the visual side. I am not sure about the language side, but on the visual side, aesthetic standards are very subjective and personalized, and a lot of controllability is needed.
Different verticals have different interpretations of the same words. So the way models are trained and organized, even with agents, makes it unlikely that a one-size-fits-all general model can satisfy all verticals.
Every vertical has its own workflow and standards. AI application companies are very important in making those adjustments and optimizing workflows for users.
That is the biggest misconception.
Grace Shao:I agree with that. We are seeing more of that realization in the market now. You have strong vertical use-case AI-native companies coming through, like Harvey. I have also met companies where former investors are building equity analyst research tools.
You can say generic GPT can be used for research very easily. But to your point, these teams know the niche use case. They know the process, the standard, and the workflow better than anyone else. Even if the TAM is small, it can be big enough for their business.
The last question I ask everyone on the podcast is: what is one differentiated view you hold? Something that is a bit against consensus.
Gary Ngan:Is it related to the company or the industry?
Grace Shao:It could be anything. Usually people answer about their topic, but it can be anything.
Gary Ngan:I think life expectancy will be a lot longer than we think today for our generation.
Grace Shao:So we are going to live to 150, thanks to Bryan Johnson’s experiments?
Gary Ngan:Possibly. Then there will be more time.
There is a lot of advancement in AI. It speeds up many pharmaceutical processes. You can run different trials much faster and understand the underlying issues more efficiently than before.
And with more time, there is more time for us to create more art.
Grace Shao:And live a healthier life. Although right now, anyone working in AI knows AI never sleeps, and I think we are all working more than ever.
But thank you so much. That is definitely a differentiated view. I really appreciate your insights and your sharing today. Thanks again, Gary.
Gary Ngan:Thank you so much, Grace, for this opportunity. Really nice talking to you.
AI Proem is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber.



Get full access to AI Proem at aiproem.substack.com/subscribe