"We Have to Be Expensive First!" — Only a $200/Month Price Tag Can Validate True AI Market Demand | A Conversation with Zhenyu Guo: Founder/CEO of Sandwich Lab
Seeing the world and understanding it are two different things.
Seeing the world and understanding it are two different things.
👦🏻 Podcast interview: Koji, Ronghui
🥷 Edited by: Starry
🧑🎨 Layout: NCon
This week's Crossing guest is Zhenyu Guo, founder and CEO of Sandwich Lab. Their product Lexi[1] is an AI Agent designed specifically for SMBs to run ad campaigns on Meta platforms.
In the interview, Zhenyu Guo explains in detail how Lexi helps clients deploy Meta ads through "shadow mode," and breaks down their thinking behind the $200/month pricing strategy as a way to validate genuine demand. He also shares the "dopamine moments" and "epiphany moments" from his entrepreneurial journey.
Guo emphasizes that Lexi is only a small piece of Sandwich Lab's larger vision. He shares the company's ambitious goal: using AI to help businesses "grow revenue." Meta ad deployment is just one path toward that objective — they plan to roll out Email Marketing Agents, as well as agents for finance, tax, legal, supply chain, and HR. Sandwich Lab's core philosophy: any agent that can help businesses grow revenue belongs in their roadmap.
Guest bio: Zhenyu Guo, founder and CEO of Sandwich Lab. He earned his bachelor's degree from Zhejiang University and his PhD from the University of British Columbia (UBC). During his doctoral studies, he co-founded a company that became a Tier-1 supplier to Tesla. As AI Director at Postmates, he led the development of autonomous delivery robots. He later joined DAMO Academy to lead the Xiaomanlv autonomous driving project.
About Sandwich Lab: Provides automated revenue growth solutions for global SMBs using AI. The company has raised over $10 million from Jinqiu Fund, Mobvista, 5Y Capital, and Gobi-Alibaba Entrepreneurs Fund.
Listen on WeChat:
Listen on Xiaoyuzhou:
Rapid Fire
👦🏻 Koji
Hello everyone. This week's Crossing guest is Zhenyu Guo, founder and CEO of Sandwich Lab. Their latest product Lexi[2] is an AI agent that helps SMBs run ads on Meta. This episode of Crossing is also our first video experiment — it'll be simultaneously uploaded to Xiaohongshu, Bilibili, and WeChat Channels, so please follow us there.
As always, rapid fire first. Zhenyu, how old are you?
👦🏻 Zhenyu Guo
I'm 38.
👦🏻 Koji
Where did you go to school?
👦🏻 Zhenyu Guo
I got my bachelor's from Zhejiang University and my PhD from UBC.
👦🏻 Koji
Your MBTI and zodiac sign?
👦🏻 Zhenyu Guo
ENTP and Sagittarius.
Let's Talk About Lexi
👦🏻 Koji
Can you describe your company and product in one sentence?
👦🏻 Zhenyu Guo
Our company is Sandwich Lab. We help global SMBs achieve automated, sustainable revenue growth by creating AI agents in shadow mode. Our first product, Lexi, fully automates customer acquisition through ad channels on Meta.
👦🏻 Koji
Can you share your current funding situation?
👦🏻 Zhenyu Guo
We completed two rounds in the past four months, totaling over $10 million. The first round was led by Jinqiu Fund and Mobvista; the second by 5Y Capital and Gobi-Alibaba Entrepreneurs Fund.
👦🏻 Koji
What about revenue and profit?
👦🏻 Zhenyu Guo
We'll disclose more later, but growth is very fast — month-over-month revenue growth is above 150%, and we're gross margin positive after accounting for customer acquisition and compute costs.
👦🏻 Koji
How big is the team?
👦🏻 Zhenyu Guo
Around 20-plus people.
👦🏻 Koji
What were you doing before founding Sandwich Lab?
👦🏻 Zhenyu Guo
During my PhD, I co-founded a hardware IoT startup called Neural with classmates. After becoming a supplier to Tesla Solar City, we sold the company. Then I founded the sidewalk autonomous delivery robot project internally at Postmates, Silicon Valley's largest food delivery platform. That project is now listed on NASDAQ as Serve Robotics, valued around $800 million to $1.2 billion. After returning to China, I joined DAMO Academy to lead the commercialization and scaling of Xiaomanlv, Alibaba's first robot and an autonomous last-mile delivery project.
👦🏻 Koji
I've actually seen Xiaomanlv on the street.
👦🏻 Zhenyu Guo
Yeah, they're pretty common now.
👦🏻 Koji
Lexi has been out for a while now. What results have you seen?
👦🏻 Zhenyu Guo
Lexi is a B2B AI service with no free trial. Subscription fees are $200 per month (depending on plan), and users' ad budgets paid directly to Meta are separate. Three months after launch, we have paying users from 94 countries, with month-over-month revenue growth above 150%.
What excites me most is that we've delivered ad performance and significant revenue growth beyond expectations for SMB owners worldwide. Our users include a hand drum studio doing team-building events in the UAE, a local hardware store in Southeast Asia, a US clinic focused on ADHD, an Australian accessibility renovation service for people with disabilities, a UK real estate agent, and a Cameroon-based vendor selling dreadlock wigs. These are small business owners from all over the world, different skin colors, different cultures — typically with 3-5 employees — who previously couldn't complete ad deployment and customer acquisition growth. Lexi enabled them to try it for the first time. That's brought us a lot of joy.
👩🏻 Ronghui
Did these users lack methods, channels, or something else before Lexi?
👦🏻 Zhenyu Guo
Running ads on Meta is actually a relatively complex job requiring experience and technical skill. Even in developed markets, the number of agencies that can provide such services for local businesses is very limited, and extremely expensive. In the US and Canada, for example, local agencies mostly offer account setup and advisory — they can't guarantee results — and typically charge $6,000 to $20,000 per month on retainer contracts.
In smaller or less developed countries, there may be no such services at all, and even if there are, they're unaffordable. By comparison, our service costs around $200 per month. Many of the users I just mentioned are advertising on Meta for the first time; they simply couldn't have done it without us.
👩🏻 Ronghui
So you're targeting SMBs that need to advertise on Meta but have never done it or lack experience?
👦🏻 Zhenyu Guo
You could say that about 30% of our users have some prior ad experience and are looking for better tools and methods; 70% are the type you just described — they have strong advertising needs but have never completed a campaign. That was actually one of our original motivations for building an ad system. Our team had very strong advertising needs at the time but had never completed a Meta campaign, so we built this product for ourselves. The original motivation came from that.
People might think "AI for marketing," especially "AI for advertising," is a hot赛道, something natural to pursue. But we're quite different from most startups or application companies. Others focus more on the "front half" of advertising — the analysis and creative fundamentals: gathering advertiser and service information, analyzing product value, competitive landscape, and market demand, then deriving which countries, markets, and audience segments have paying capacity and real demand. After this analysis, they typically hand off to an ad creation agent that uses advertising expertise to produce content with selling points, tags, audience info, and creative assets — images, copy, and video.
Most companies, after creating the ad, deliver it to the user, who then deploys it on the platform themselves. However, on computational advertising platforms like Meta, advertising isn't a one-time prediction problem of finding "the single optimal solution." Ad creatives and ad sets have very short lifecycles — usually needing refresh every two or three days to avoid fatigue and other issues — so ad deployment isn't something that ends with one-time delivery.
So we don't treat ad delivery as a one-time handoff. We also handle the subsequent "technical trading." After launch, we dynamically monitor all objective feedback in real time, adjusting budgets and parameters based on ROI and odds estimates, merging or killing ad sets, rebuilding ads, and looping back to fundamentals for iteration. In slower cases, this cycle happens several times a week; in faster ones, some regions and ads need multiple adjustments per day. That's the second part we've built out.
If you look at these two stages separately, you'll notice that most companies, after completing the first step and delivering to users, end up serving people who have run ads before and understand them — because these users know how to maximize ad value on the platform. We, however, want to serve users with strong advertising needs who have never run ads and don't understand them. Only our full-process service can truly meet their needs.
If you hand ad creatives to someone who doesn't understand advertising, they still won't know what to do with them. That's why we built fully automated ad delivery. How big is this market? To put it dramatically, there are over 130 million SMBs with a presence on Meta globally. Even when you narrow down by country, region, and revenue scale, there are at least over 50 million potential users — advertisers who rely purely on social platforms to drive growth.
The reason we focus on them isn't simply because they can generate higher revenue. In fact, this kind of growth is much slower than working with KAs (key accounts). What we truly care about goes back to our founding mission: we believe AI progress represents a leap in productivity, and my personal long-term focus on automation and robotics stems from my deepest concern with "distribution." I believe fairness and rationality in distribution lead to a better society, yet wealth distribution itself is often inefficient and not something I can directly participate in. So I hope to use productivity gains from technological progress to allocate to those who truly need it — that's the essential motivation behind our desire to serve these additional 50 million SMBs.
👦🏻 Koji
This is really interesting because it actually influences many of your decisions. For example, your current decision not to serve KAs — it's because KAs don't really seem to need you right now.
👦🏻 Guo Zhenyu
Yes, Koji is absolutely right. We decided early on not to serve KAs in our first phase, because large enterprises have the ability to hire excellent talent — there's no doubt about that. They may want to improve efficiency and reduce costs, they may want leaner talent structures — that's another consideration. But clearly, fully automated AI is not a hard need for large enterprises that can afford to hire people.
👦🏻 Koji
But won't this directly affect Lexi's or the company's commercial outcomes? It sounds like you believe this will lead to greater commercial results in the long run, or is commercial outcome not the top priority for you in this venture?
👦🏻 Guo Zhenyu
I think it will lead to greater commercial outcomes. I worked at Alibaba for four years, and many colleagues were drawn by Teacher Ma's vision — the way it was articulated and genuinely practiced was hard not to be moved by.
👩🏻 Ronghui
"Make it easy to do business anywhere."
👦🏻 Guo Zhenyu
Yes, very simple and direct. In the B2B space, to put it somewhat bluntly, there may be no new demands — the demands are all ancient, just solved in new ways. The name Alibaba itself has meaning: hundreds or thousands of years ago, Arab caravans solved the problems of commercial circulation and matchmaking.
Along this path, achieving massive commercial outcomes often comes from starting with altruistic value. Building a hundred-billion-dollar company isn't something someone purely self-interested can accomplish. That's the first point.
From a specific technical and socioeconomic structure perspective, in an era of rapid technological progress, lowering the barrier to using advanced technology usually brings the largest and fastest commercial results. By serving SMBs, we're essentially continuously lowering the barrier to using advanced technology in key resources and methods.
👦🏻 Koji
I'm thinking that when KA clients hear about a service like Lexi, they'd probably be curious too, want to try it out — I bet quite a few have reached out. So are you firmly turning them down right now? Not even willing to chat?
👦🏻 Guo Zhenyu
We do have a world-renowned FMCG group as a client, and we're discussing a pilot experiment. The other party initially approached Lexi from the surface value of running ads on Facebook, but I'm more looking forward to guiding them to understand Lexi's greater value — not just ad delivery, but channels for reaching potential users.
Many enterprises have numerous ways to reach existing users: SMS, private domains, in-product A/B testing. But relatively few ways to reach potential users and validate demand. Lexi can provide greater value at this level, which is especially helpful for large enterprises. Small businesses typically don't need full-chain production capabilities across every endpoint, so we would consider this type of collaboration with large enterprises.
👦🏻 Koji
Does Lexi use Lexi itself to advertise on Meta?
👦🏻 Guo Zhenyu
Yes, and it's a fun inside joke. When fundraising, we often talk about how Lexi grows Lexi, using Lexi to advertise Lexi. This is actually our process for testing the product and iterating algorithms. Our growth team is very small — we mainly rely on Lexi for advertising. The data results so far are quite interesting, bringing relatively very low CAC (customer acquisition cost), and even after accounting for acquisition and compute costs, there's still gross margin per order.
👩🏻 Ronghui
I want to add a follow-up: beyond growth data and temporarily giving up on KAs, were there any more granular selection criteria for your initial customer base that led to your later results?
👦🏻 Guo Zhenyu
We had two observations that may sound a bit crazy — not necessarily absolutely rational, but I'll share them:
The first observation is that we very subjectively lean toward serving local businesses, what in China is called "daojia daodian" — home services and local stores. It's a form that covers all categories, and what it mainly distinguishes is e-commerce and online business. Although Meta ads work more directly for e-commerce because e-commerce is already online and understands advertising, returning to our product's original intention, what we want to do is lower the barrier, enabling more people without business capabilities to gain them. So we're consciously targeting the local economy, which is also my personal hope.
The second observation comes from comparing Chinese and US markets. For many years, I've spent half my time in China and half in North America and other regions annually, and this comparative feeling is very intense. Chinese local businesses have extremely strong operational capabilities, with online and offline operations refined to the extreme. Milk tea shops, barbershops — they're all using private domain traffic for precision operations. Even my regular barber in Hangzhou told me that managing private domain traffic is more important than cutting hair.
This refined commercial capability isn't accidental. It stems from several factors:
- Talent spillover from major platforms like Alibaba and Meituan has influenced vast numbers of people living within these ecosystems;
- Platforms like Meituan continuously distribute usable business strategies through SaaS tools, SOPs, and operational campaigns, enabling stores from first-tier to third-tier cities to constantly learn and operate tiny businesses with data-driven thinking;
- There exist millions of ecosystem operators helping complete marketing activities like opening new stores, cold starts, user acquisition, coupon distribution, and package deals — these people often work remotely, living in inland cities like Xi'an and Lanzhou, at appropriate costs, forming a complete ecological chain.
The accumulated result of this model is an undisputed miracle of human commerce, enabling even barbershop owners to understand the essence of "growth hacking" — extremely rare on a global scale.
And none of this has happened in the US. Not because Americans can't write Meituan's systems or SaaS tools, but because this ecosystem requires millions of strategically-minded talents working seamlessly together to effectively land. In developed countries like the US, UK, Australia, and Canada, or emerging markets in Africa and South America, this talent ecosystem and scale simply doesn't exist. This made me realize: what AI-powered full automation should probably do most is these very things. For local life and local economies outside China, this will to some extent bring massive efficiency gains.
👦🏻 Koji
When I tried Lexi, I found the operation very simple — just input product information and budget, and it directly gives a delivery plan and starts executing. But I noticed I couldn't see the ad creatives Lexi generated for me; I looked everywhere and couldn't find them. I later realized this might be by design — you don't let users see the generated creatives before direct deployment. How did you think about this design? Because right now many AI agents show users every detail to build trust — what's your view on this? Are users accepting it so far?
👦🏻 Guo Zhenyu
Excellent question. Actually, we do have a very deep entry point where users can see copy, tags, and creatives, but it's buried very deep — intentionally so that users don't see them. Even if users can see them, they cannot modify copy or tags. This was a very clear decision from the start.
Returning to Koji's question: we think what Devin and Manus have done is excellent — showing the process helped users quickly accept the feasibility of fully automated agents, which was necessary at that time. But that was more about building trust in the concept of "fully automated agents" as a category, not trust in any specific product. At this stage, showing the process for an application like ours is no longer that critical for building trust. When we started the project, our fundamental choice was to build AI agents in Shadowing Mode — "shadow" means completing the user's work, not showing the process.
Second, users have different expectations in different scenarios. For example, when Devin helps users automatically write code, its essence is directly delivering software to the user. There's an important term here: hands off. This is the endpoint of participating in automation tools or AI-assisted work — the product gives you software, puts it in your hands, and you use it. In this case, the AI agent system has no way to participate further, no opportunity for subsequent improvement. Even if there are ways to improve the product later, new insights, you rarely get another chance to touch your work. So accumulating trust through the process, showing that you're working hard, and helping users understand what you did — that may be very important.
What we chose instead is the "infinite game" scenario. No one is in business for just a day, and no one runs just one ad — advertising is a continuous process of operation. In this scenario, users don't care about specific deliverables; they care about sustained operational results. So we introduced the concept of the "Revenue Generating Machine," where users don't need to worry about the process, only the outcome. In this model, the quality of any single ad isn't critical — whether users stick around depends on long-term returns. Also, our R&D resources are limited, and showing the creative generation process isn't a current priority. We might explore it in the future if needed, but it's not essential right now.
The third and most fundamental reason — and this sounds somewhat counterintuitive — is that ad copy and creative assets themselves don't actually have that much impact on that month's ROI. Especially on a social advertising platform like Meta, and since we primarily serve SMBs, our clients aren't like big brands that need strict brand consistency and PR safety. Users only care about ROI. It's a non-zero-sum, continuous game of strategy. As I mentioned earlier, ad creatives on Meta typically survive no more than two days. If we made users approve every generated ad, it would create an enormous burden. For our long-term users, when they've spent $100,000 on ad budgets, the creatives generated number in the tens of thousands — they couldn't possibly review them all. So having users review creatives during initial onboarding helps build trust, but in the long run, it doesn't affect core value.
So how do we solve the trust problem? This was actually an interesting part of our startup journey. We'll be launching two new products in August, and both address needs where users have no alternative solutions. When there's no alternative in the market, formal trust isn't the core issue — as long as you can deliver ROI, you're good.
On the Company's Vision: Starting with Lexi, But Not Stopping at Lexi
👩🏻 Ronghui
I want to add two questions: First, regarding the "needs" you just mentioned — how did you identify these two "needs"? Second, you mentioned some early judgments that were later validated, as well as some counterintuitive observations. What else did users later demonstrate that aligned with or diverged from your initial expectations?
👦🏻 Zhenyu Guo
Let me start with how this "observation" came about. Before last year, I had already started thinking about building a "fully automated AI agent," but I didn't yet know exactly what it would do. For the past 15 years I've been in the robotics and autonomous driving industries — I'm very familiar with supply chains, perception, decision-making, localization — but business operations required completely relearning. So for this startup, you'll notice most of our team actually doesn't come from advertising backgrounds. We're "methodology-driven" entrepreneurs — we have hypotheses, we validate them.
When I was traveling around the world, I discovered that social marketing had become the most important thing. Wherever Xiaohongshu exists, it's always the top platform; where it doesn't, people look to Instagram. When people travel abroad, they don't actively search for rankings because the value is limited — they just scroll social media. So all local businesses want exposure. In Tokyo, on both US coasts, in Canada, Australia, Southeast Asia — everywhere I'd ask: local coffee shops, restaurants, boutiques — they all had beautifully curated Instagram accounts. I'd ask if they ran ads? And 99.99% of them only did organic traffic, no paid advertising.
This was strange. Intuitively, if you're already doing well with organic traffic, adding ads should be an amplifier — I was certain about this intuition. But why weren't they doing it? I started putting myself in a local merchant's shoes: if I opened a milk tea shop or coffee shop, how would I advertise? So I actually created an account and tried running ads myself. I found it was not only difficult but incredibly tedious. I'd say I'm a reasonably fast learner, but after a full day I still hadn't figured it out — which was pretty embarrassing. I could probably pick up SEO or Google Ads in a day, but social advertising is genuinely complex.
Then I looked for service providers. In the cities I visited, I'd search Google Maps in the local language for someone who could help me run ads. What I found were traditional digital marketing agencies that had been around for over a decade. These companies would first help you with graphics, then build you a website, and only then say they could help with Meta placement — not the kind of automation and efficiency I had imagined. Through this "dumb method" of small-sample observation, I confirmed this market was indeed a "rigid need" that no one had truly solved. The logic was simple: if nobody's doing it, then whoever helps them do it is automatically better than the rest. Though this inference wasn't rigorous, that's how it started.
Now for the second unexpected discovery — where user behavior diverged from our initial assumptions: when our product first launched, users needed to have their own ad accounts, pay us service fees, and we'd help them run ads through APIs and such. But we didn't expect that a large portion of users didn't even have Meta ad accounts, and couldn't even configure them. This surprised me, but also gave me a lot of insight.
I discussed this with Koji before: the core value of SaaS isn't really the subscription fee — it's solving complex localization and configuration for customers. In the pre-SaaS era, buying software meant installing from CDs, and every company's IT environment, operating system, network setup, and database was different. Often it wouldn't install, and you'd need to send engineers on-site. When we bought games on CD back then, installing for an hour was normal, and you'd breathe a sigh of relief when it finally worked. Young people today don't have this concept — click on Steam and play. But even after two or three decades of SaaS development, many "configuration" problems still persist. Ordinary users can't solve them themselves, and the barrier is much higher than people imagine. This was something we completely didn't expect at first, and had no idea this need existed. So what we're doing next is minimizing this legacy "usage barrier" as much as possible, helping users with account setup, registration, and these kinds of things.
👦🏻 Koji
Is this one of the new products you mentioned launching in August?
👦🏻 Zhenyu Guo
No, this is a major feature upgrade within Lexi. Our new product is an Email Marketing Agent — it sounds like a very old industry, but we're very excited about it.
👦🏻 Koji
That's indeed quite surprising. Many people might assume after Meta placement, your next step would be Google or TikTok, but you're going to Email instead. How did you arrive at this choice?
👦🏻 Zhenyu Guo
Almost everyone asks me: "When are you launching on TikTok? When are you doing Google?" This is the most common question. I think this is exactly where "methodology entrepreneurs" differ from "AI practitioners."
If I came from a traditional marketing background, I might naturally want to do everything. But we need to first understand: what is the fundamental principle of Google Ads? What is the core principle of TikTok advertising? What about Facebook? Though they're all called advertising, CCTV ads, Hunan TV sponsorships, and Facebook ads are fundamentally completely different. Google is "active search" — users initiate demand first, and Google knows user interests intimately, almost 100% accurately — whatever users search for, it knows. So on Google, automation is essentially a centralized allocation problem, without much middle-layer value. And Google only covers active demand, not passive recommendation.
👦🏻 Koji
So you believe Meta comparatively doesn't understand users as well?
👦🏻 Zhenyu Guo
Yes. Meta is fundamentally a social network. Its core is your friend relationship graph, not search-based interest profiles. Of course it works hard to understand users better, but it's inherently not an infinite content feed, nor is it as aggressively interest-based as TikTok — it needs to preserve the essence of social relationships. So its understanding of user interests isn't as deep. There's also less e-commerce on it, marketplaces aren't particularly strong, and with privacy policy restrictions in many regions, its data is increasingly limited, so its understanding capability can't compare to TikTok or Google.
So Meta's ad system transfers the crucial problem of "how to understand potential users, how to find them" onto advertisers and media buyers — the system is just an execution tool. We need to guide Meta on where to look and how to find people. As for TikTok, it's inherently a recommendation feed, and ByteDance's core strength is recommendation, so it understands users deeply. But the biggest problem with TikTok ad automation isn't placement strategy — it's creative. What kind of short video content attracts people? This is a creative problem, something humans excel at. Human life is the source of good content. So I don't think AI can do this well in the short term, and frankly it's not my personal interest either.
So to summarize:
- Google: Doesn't need a complex AI strategy layer, because it doesn't require much structured, transactional strategy to place ads.
- TikTok: Needs good creative and rapid testing, not placement strategy.
- Facebook: Has limited user understanding and needs third parties to do "black-box strategizing" for it — this happens to be our core capability.
So we're focused on Facebook, and not doing Google or TikTok for now — not because we look down on them, but because they fundamentally don't need our technology and product capabilities.
👦🏻 Koji
This explanation is quite interesting, and it's very generous of you to share so much detail on the podcast. But I believe some of our listeners may have reservations, and some might say it sounds very commercially calculated, or that we might be underestimating things.
👦🏻 Zhenyu Guo
Of course, because my own views are very cautious. I'm somewhat of an outsider here, possibly showing off in front of experts — I hope our audience can correct me and help our product continuously improve.
👦🏻 Koji
Are there clear metrics that prove Lexi's placement performs better than human media buyers? Or is this not important to you?
👦🏻 Zhenyu Guo
It's very important to us. We have three key categories of metrics:
- Traffic — how many people visit your page.
- Leads — how many people leave contact forms or phone numbers.
- Sales — how much revenue ultimately converts.
We review these metrics daily, continuously optimizing point by point. We're still in rapid growth phase, and most of our users are new, so data is continuously accumulating. From settled clients so far, our average performance reaches at least the level of human agencies.
The core metric for Sales is ROAS (Return on Ad Spend). Generally, a ROAS above 2 is solid, and for SMBs, above 1.5 might be acceptable. Many of our clients hit 8 or even 10 — spending one dollar to earn ten back. Traffic performance has been steadily improving too; automation naturally has an edge on traffic optimization since there's so much clicking involved. Leads are harder to call — we're still refining our tracking. Some local businesses rely heavily on phone calls, so we're building automated tools to trace those conversion paths and turn phone interactions into quantifiable data. All three metrics are currently within expectations.
👦🏻 Koji
Your Discord seems pretty active — users are actively discussing how to use it. But I've also seen some complaints, like slow customer response times...
👦🏻 郭振宇
Yes, we didn't have resources for customer service before. We recently launched Intercom and now have support in place.
👦🏻 Koji
Are you using Intercom's AI Agent?
👦🏻 郭振宇
Still considering it. The team's feedback is that it's quite good.
👦🏻 Koji
We should chat about this. Another guest previously mentioned they use Intercom too, but found Intercom's built-in AI Agent not great. Since Intercom is a third-party platform, they integrated an external AI Agent instead, and the results were pretty good.
But back to the main topic — I saw user feedback on Discord saying Lexi doesn't seem to do much optimization for them, that it just opens accounts and starts running ads directly, and that AB testing didn't show obvious results. What's your take on this kind of feedback? Was this expected? Or is it a product selection issue?
👦🏻 郭振宇
We review all this feedback, but we've observed that many users turn off their ads on day one or day two. But with Facebook ads, you need to run at least 3 to 7 days, and traditional agencies will tell you it takes two weeks to see results. Because advertising operates at scale — it's a mass, group behavior. Who clicks, who doesn't — that takes time to settle.
Many users don't understand how Facebook ads actually work. They expect instant results the moment they go live, and when they don't see anything on day one or two, they shut it off. That's the main reason we've identified. Recently we've added more product guidance and operational incentives like rebates to encourage users to keep campaigns running longer, at least 3-7 days, so they're more likely to get meaningful results.
👦🏻 Koji
Have you considered deliberately doing something on day one or two to make users feel like Lexi is "working for me"?
👦🏻 郭振宇
Actually, that wouldn't be the right approach — it would undermine ad performance, which is common knowledge. Going forward, we're choosing to be more upfront, helping users understand the basic cycle that advertising requires. This may increase conversion costs, but in the long run, actually improving outcomes for users matters more.
👦🏻 Koji
Lexi is currently on a fixed subscription model, right? So a big client spending $100K and a small client spending a few hundred dollars — they pay you the same amount?
👦🏻 郭振宇
Currently yes. But once we roll out our next model, there will be a new pricing structure, more like traditional agencies: clients send us their full budget, we allocate a portion to Meta, and Lexi's revenue structure becomes more flexible — with supply chain financing, platform management fees, service fees, and other diversified income streams.
👩🏻 Ronghui
What's the thinking behind this pricing model?
👦🏻 郭振宇
To be honest, the pricing was somewhat subjective. Before launch, I felt it needed to be "expensive enough." At the time, the market standard was around $200, with a few exceptions like Devin at $499, but mostly around $200. Why expensive? It was a very subjective, almost crazy idea. I believed the fastest way to validate PMF is to charge a premium — if the product isn't perfect and people still pay, then you've truly found a market need. We didn't even offer a free trial at first — you had to pay $199 to use it. And people still paid. That made our team even more convinced this was a real need.
From a more effective operations perspective, we still haven't opened free trials, but around August or September, we'll experiment with more marketing tactics to attract more sign-ups and usage.
👦🏻 Koji
I think you've made a lot of bold and decisive choices — whether it's the $199 pricing, committing to SMBs, or explicitly not doing Google and TikTok. These are all very gutsy decisions. Beyond these, were there other decisive choices you made at the time that felt important?
👦🏻 郭振宇
Next we're doing Email Marketing. Many might find this strange — an AI startup in 2025 entering a 50-year-old industry, a seemingly dead-end space. But as I said earlier, business doesn't have "new needs" — only evolving methods. Business schools have taught the same things for centuries: financial management, supply chain, marketing. Nothing's changed. We want to serve all markets. Business schools never divide by industry — financial management and supply chain management work the same across sectors. So for us, Marketing equals growth. We're doing it.
The direct reason for Email Marketing is that we need it ourselves. Lexi Ads can acquire new users for Lexi, but to reach existing users, we need Email Marketing. Especially in less developed markets and parts of the United States, SMS is also commonly used — we'll do that too in the future, but starting with email.
When we looked into doing Email Marketing, I found there was no tool that truly solved my problem. Giants like Mailchimp and Kllaviyo have high market caps, but they solve "workflow problems" — importing email lists, compliance, scheduled sending, click-through rates. What to actually send, when to send it, how to improve landing page conversion — users still have to figure that out themselves. It turns out nobody in the world is solving this problem. Lots of people selling shovels, almost nobody actually helping you mine, yet the demand is huge. There are very expensive Email A/B testing tools where you pay dozens of dollars and they assist with A/B tests, but you still have to design them yourself. I was baffled that such a large market had no product addressing this need. So we're building this Email Agent to directly help users improve clicks and conversions, not just send emails. You just tell me: what do you want users to open, and which landing page do you want them to land on — I'll help you improve that. Similar to our advertising logic: not just helping users place ads, but helping them get results.
I think the value here is enormous — at least I see value that hasn't been fully overlooked. I believe in the elite ToB SaaS market, there are many rigid needs that have been ignored for 20 consecutive years, simply because technology couldn't address them before. Now it can, and we'll solve them one by one.
👦🏻 Koji
So you define Sandwich Lab's goal as not just Lexi's automated ad placement, not just Email Agent, but continuously launching products and services that truly help businesses grow revenue. Will the third, fourth products also follow this cross-domain Agents path? Can you briefly reveal the direction?
👦🏻 郭振宇
Actually our thinking was quite simple — we just followed the four classic directions from business school: finance/tax/legal, marketing, management and supply chain, and human resources.
For finance/tax/legal, we won't do compliance-related services since that's not our strength, and there are already very mature, professional solutions on the market. What we mainly do is insight analysis — deep analysis from a financial perspective to uncover potential revenue growth points for businesses, and convert these growth points into executable actions.
Notably, these executable actions aren't fundamentally in the financial domain itself — they return to customer acquisition, operations, and supply chain-driven growth. Simply put, finance/tax/legal is more about discovering growth opportunities and formulating sound growth strategies for us, which we then execute through other channels and means. That's why we do financial analysis while also working in other domains.
Though these areas may sound scattered, and none seem grand on their own, they're all essential yet neglected rigid needs for business growth. We want to stitch them together into an automated system that helps businesses sustain growth.
👦🏻 Koji
Here I want to circle back to Lexi itself and talk about competition. Many people ask — you're doing automated ad placement, and as mentioned in previous "Crossing" podcast episodes, Mark Zuckerberg has said that one of AI's biggest meanings for Meta is automating ad placement. He wants advertisers to just tell Meta what product they're selling and their budget, and Meta handles everything else automatically. That sounds almost exactly like what you're doing. If the platform itself is doing this, does your space still exist? How do you view this?
👦🏻 郭振宇
This is a frequently asked question. Even if Meta hadn't publicly stated this, many would assume they're working on something similar.
I look at this from three angles:
First, from a principles perspective, Meta doesn't fully know user demand and purchase intent the way TikTok or Google do, so to do this, they'd need to approach the black-box problem with methods and logic similar to ours. This isn't something a small branch in their system could just cover.
Second, from a business composition perspective, Meta's main revenue still comes from ad placement, and there are over 50 million people globally who don't know how to run ads. We're addressing an incremental market — helping Meta, not threatening them. If someday we grow large enough that Meta views us as competition, that means we've already succeeded tremendously.
Third, from a business choice perspective, if we help Meta achieve automated coverage of the long-tail market they couldn't reach before, that's beneficial to Meta, not harmful.
👩🏻 Ronghui
Earlier you mentioned observing customer needs and conducting research. Now that product entry points are shifting for users, why did you still choose to continue going deep on Meta, a large platform that's existed globally for so long?
👦🏻 郭振宇
We chose Meta by starting from the lives of potential customers and the world they inhabit. Whether it's large enterprises in the United States or users in emerging markets, people still spend enormous amounts of time in Meta's world every day — on Facebook or Instagram. Snapchat has risen, but Instagram and Facebook still dominate users' attention and time across most regions. That's a clear fact we've observed.
TikTok is growing fast, but in terms of ad budgets and market share, it remains far smaller than what countries spend on Meta. So what we chose is still the entry point that the world's customers need most — there's no doubt about this. This is especially obvious in the United States, not to mention Africa, the Middle East, South Asia, Southeast Asia, and other East Asian countries. If I'm not mistaken, the less developed a country's physical infrastructure, the more it relies on Meta's digital infrastructure.
People in these regions handle daily necessities — food, clothing, housing, transportation — all on Meta platforms. They transfer money via WhatsApp or Messenger, or complete transactions directly on Meta platforms. Although there are many automated website-building tools, most people still find Shopify or Squarespace too high a barrier. Only practitioners like us can barely use them, but my parents' generation would never use automated tools to build a website. The barrier to these tools is much higher than people imagine. And this holds true in the US and UK markets as well.
So Meta still dominates many businesses. Many small merchants don't build websites at all — they just create a page on Facebook or Instagram as their online presence. Will new entry points emerge in the future? More entry points? Possibly. But returning to our vision: we focus on revenue generation for SMBs, launching different products as the times change.
👦🏻 Koji
Coming back to "fighting on multiple fronts" — it's hard enough for a startup to do one thing well. Why choose to do so many things simultaneously? Where does your confidence come from?
👦🏻 郭振宇
Investors ask me this every time. When I initiated projects last year, I actually listed many needs but held off from acting rashly. So why start suddenly now? Two observations.
First, a recent realization: my biggest advantage may not be having the Lexi product or this automated iteration algorithm, but rather that my editorial cost for creating new products for users is extremely low compared to others. I need to fully leverage this. Because whatever application we build, we're facing the same global users. In this case, our go-to-market channels, operations, customer service, and DevOps are all fully reusable — no need to rebuild from scratch. From an R&D perspective, I estimate we can launch a new product with just ten people. Because my algorithm and core logic focus on objective feedback and continuous iteration rather than one-time delivery, the operating model stays consistent. So algorithms, backend, even frontend components are highly reusable.
Therefore, my R&D costs are much lower than independent companies or large corporations building new products in the same space. And to truly validate PMF, you can't just finish writing code — you need to seriously operate the product. As I mentioned earlier, our customer acquisition and operating costs are also reusable and validated, with the same user base and growth methods, so learning costs are low. Thus, even seemingly unrelated new products on my main line — from R&D to launch to initial validation and operations — my editorial costs are much lower than others'. This is something I must fully exploit.
After realizing this, we began executing without further hesitation. Initially I worried that fighting on multiple fronts would bring extra burden, but I found these are all reusable, which actually pushes the team to do standardization and automation even better. The same functional teams can serve different product lines simultaneously, with consistent audiences and growth methods, and consistent payment models.
The second question: should these products be integrated into a single entry point? I struggled with this for a long time — when we last talked, I hadn't figured it out. Only recently have I somewhat clarified. Previously I had an extreme thought: build one product, one feature, hiding different functions in Shadowing Mode where users don't even know what capabilities exist behind it. Users open the website, enter a URL, click one big button, and revenue starts growing the next day. That's the product we wanted to build.
But I later realized this kind of aggregation is what large companies like Microsoft and Baidu need — companies in stages of fixed traffic and slow growth. They distribute traffic to different products through aggregation. The ByteDance approach is different: individual apps fight alone and grow rapidly, because ByteDance is an extremely data-driven growth team, whether in product iteration or customer acquisition methods. If different functions are aggregated together, the data-driven growth logic becomes extremely complex: should we change? Change which one? Who does it affect? It creates many problems.
In fast-growing emerging fields, we should avoid aggregation and instead let products fight alone and grow continuously — this better serves data-driven strategy implementation. And AI applications clearly belong to such a market, where faster is better. So we choose to separate each product independently and let them grow continuously.
Entrepreneurship, Life Goals, and Self-Actualization
👦🏻 Koji
You sound optimistic and decisive when facing challenges. This reminds me — some serial entrepreneurs easily develop PTSD after experiencing big-company competition or unexpected setbacks. Especially when they've been crushed by a big company during entrepreneurship, or achieved decent success, then unexpected events cause everything to fall through at the last minute. This makes them hesitate when starting something new, subconsciously recalling past difficulties and pitfalls, making it harder to launch new things or challenge bigger ambitions. You've also experienced entrepreneurship — what keeps you charging forward this time?
👦🏻 郭振宇
I haven't truly been a CEO before. During my student startup, the CEO was a business school classmate. The second time with the robotics project, although it eventually spun off independently as a NASA-appointed company, it was actually still managed by the parent company's CEO and CTO — strictly speaking, not a typical solo entrepreneurship process.
During the first startup, we went without funding for a long time, surviving half a year on Kickstarter's $500,000. We didn't really have product-market fit either — we just somehow signed deals and sold. Looking back, it was pretty precarious, but at the time we were still studying and didn't think much of it. We approached it with a "just for fun" mentality. For things we hadn't done that were complex, curiosity was usually the biggest driver.
The thing that's brought me the most joy recently is finally getting our AB testing system live with the team — yesterday we just started running the first feature test. The previous three features were all purely manual AB. On launch day we discussed how to do AB, and I said we shouldn't just look at Fisher, but should use orthogonal AB to test subsequent correlation and causality hypotheses.
Although it sounds like motivational talk, I said: "Correlation lets you predict the future; causality lets you change the future." This actually expresses my most intuitive experience of entrepreneurship. Entrepreneurship is constantly proposing hypotheses and doing correlation or causality validation. Correlation is for specific tactical choices — which ad, which algorithm, who to sell to, how to sell. And when occasional causality judgments appear, whether validation succeeds or fails, both are important. This validation process is deeply rewarding.
So having gone through those two or three processes, for me, making things happen is not just the goal but perhaps also the means — the essence is understanding the causal structure of the world. If Lexi succeeds, it means we did things right. If the email marketing product hits a million MAU as a free product in three months, it also validates our direction. This validation is rewarding not just as a result itself, but also guides us toward bigger and farther-reaching things.
By contrast, large companies often have long paths and mixed feedback mechanisms, easily leading people to chase false feedback — promotions, raises, stock — but these aren't real feedback. The entrepreneurial environment helps people see what truly important feedback is, because in a startup, giving yourself a bigger bonus can't solve growth and anxiety problems. Growth is the only cure.
Another thing that deeply resonates with me: before several product launches, our team was in terrible shape. By Alibaba or ByteDance standards, every function on the team was weak. But we still launched products and achieved growth. This firmly convinced me: as long as you find real demand, you can launch products, because where no one competes and others look down on, you can push through. And Sandwich Lab's team today far exceeds what I experienced back then, so I'm very confident. This is also what makes "feasibility" feel real and important to me.
👦🏻 Koji
So what kind of company do you hope to build Sandwich Lab into? When would you feel satisfied, even reaching your ideal state of success?
👦🏻 郭振宇
This is actually quite hard to answer, because I haven't really thought about what "success" means. It may sound like talking without having experienced hardship, but I've been relatively fortunate since childhood — no major setbacks at any stage. School, employment, entrepreneurship all went smoothly. The startup teams I joined all had successful exits and made money, which is actually quite a low probability. So the word "success" rarely appears in my mind — not pretending, but genuinely not thinking about it, because it's inherently hard to define.
What is success? Is it enlightenment and self-awareness about life? Harmonious relationships with others? Impact on the world? Or concrete achievements? I think all are possible, but none is the only answer. Compared to "success," I more often think about "why do people live" — a question that may have no ultimate answer at different stages, only that we constantly satisfy ourselves in phases.
I don't particularly like competition, nor do I want to compete on speed and scores in the same exam. I'm neither good at nor interested in such things. I prefer to do things without competition but sufficiently interesting. And for "to what degree is good enough," I actually have no answer either. On one hand I may be hard to truly satisfy; on the other, I'm easily satisfied. Honestly, I'm someone who pursues dopamine, and I indulge myself in pursuing this immediate gratification.
And entrepreneurship is the activity I've found with the highest dopamine density and concentration — more satisfying than gaming, more than sports. So I treat entrepreneurship as my main activity for pursuing dopamine, which itself satisfies me greatly.
👩🏻 Ronghui
No wonder you read Borges.
👦🏻 郭振宇
Right, Borges isn't a writer I particularly love, but he's rational and exquisite, with wonderful structural circularity.
👩🏻 Ronghui
Do you share these thoughts with investors? What feedback do they usually give you?
👦🏻 郭振宇
I often bring up topics like "what makes a better society" with investors, and at first they’re usually caught off guard. But most of them, after thinking seriously, will discuss it with me: what kind of society is better? What investments should we make to improve society? Then I’ll smile and say: "Invest in me—I’m working on making society better." A lot of the time, that’s how I open my investor pitches.
As for other things, I’m not sure how directly they relate to running a company. There was a period when I suddenly realized that Kant barely ever left the small town he lived in his whole life—just walking around, writing—and yet he was one of the humans who understood the world most deeply. That made me realize: seeing the world and understanding the world are two different things. Many people travel the globe without necessarily understanding more than Kant did. That had a big impact on me. It made me value formal discussion and essential understanding more in my work.
For example, I often talk about computational advertising with candidates. Once, a candidate from Alibaba explained something to me that made me realize: the foundation of computational advertising is Pareto optimization. A lot of people haven’t thought about why computational advertising is Pareto. It has to do with free-market economies versus planned economies. If the goal is to maximize total system value, a planned economy actually struggles to achieve that. You need to design mechanisms where each participant, in pursuing their own self-interest, optimizes the whole system. That’s how Alimama and Tencent’s ad systems work too. Even though they have all the data, if they directly assigned ads, both advertisers and the platform would earn less. The reason auction-based ad systems are efficient is that when each bidder selfishly helps an advertiser maximize their own benefit, the whole system reaches optimality too. This logic is isomorphic to free-market economics, which I find fascinating.
Why doesn’t computational advertising use centralized, planned allocation, relying instead on each participant selfishly pursuing maximum gain? Because only this way can the entire system achieve optimality. The ad auction system is efficient because when each bidder selfishly helps an advertiser get the best outcome, the whole system reaches optimality too. This logic is isomorphic to free-market economics, which I find fascinating.
From here I went further: in our technology and business choices, we should maintain this structure too. My ideal is to hire colleagues who can inspire me, to discuss these underlying logics together, not just execute tasks. The core of Pareto optimality is: in the system, if any participant’s gains increase further, another participant’s gains must decrease. In other words, before reaching Pareto optimality, it’s all non-zero-sum, win-win. Wonderful. I later realized that what we’re doing is exactly non-zero-sum game theory. For example, doing ad placement for SMBs on Meta, or email marketing, or supply chain management—there’s no "opponent," no irreversible risk, allowing us to keep exploring and iterating.
In true zero-sum games, like quantitative trading or direct price wars with competitors, every wrong decision can mean total loss. There’s no room to explore iterative strategies. Choosing non-zero-sum scenarios to help all SMBs grow gives me great confidence, because even if a campaign goes wrong, we can start fresh the next day. It can genuinely contribute to global GDP growth, rather than the you-win-I-lose dynamic of zero-sum games.
👦🏻 Koji
I’ve noticed that what gets people out of bed every day really varies. Zhenyu once said entrepreneurship is the fastest way for him to get dopamine, letting him experience the world comprehensively and diversely.
That reminds me of a recent conversation with an investor around fifty, very accomplished. I asked him: "Why are you still so passionate, fighting so hard? You don’t have to choose battle every day." He paused for two seconds and said: "There’s not a single day I’m satisfied with myself." Maybe this perpetual dissatisfaction is exactly what drove him to achievements far beyond ordinary people. Another founder friend’s goal surprised me—"make a hundred million dollars first." He reverse-engineers every decision to hit that target.
👦🏻 郭振宇
It’s a good goal, especially when you’re young and need that sense of responsibility, though I don’t think it’s necessary. Whether you can enjoy your life comes down to whether you have the mindset to take the next step, whether your next move can be independent of anyone else’s judgment. But having stable income to support my family when I was young was important. Actually, taking the next step isn’t that hard. You don’t need to wait until you’ve made a hundred million—just step out. Of course, this friend’s goal is good; I hope he achieves it soon.
👩🏻 Ronghui
Really not that hard? Hahaha.
👦🏻 郭振宇
I don’t mean a hundred million is easy. What’s truly hard is being honest with yourself and taking that step. If you need inspiration, read Siddhartha or Hesse’s books—you’ll find some answers. I actually envy people with clear goals. Like that senior—he’s dissatisfied with himself, but at least he knows what satisfaction would look like.
But most people don’t really know their core needs. The true underlying needs are "to be understood" and "to be seen." Work, creating value—it’s fundamentally about hoping to be seen and understood by society. I think human meaning operates on two levels: material and social. One is your interaction with the world (work, hobbies, sports), the other is interaction with others (being seen, being understood). The latter may be more core, more essential.
👩🏻 Ronghui
In your past work experience, do you feel those needs to be understood and seen were ever met?
👦🏻 郭振宇
Impossible to satisfy. Entrepreneurs are never satisfied. We might seek understanding from family and friends, but being truly understood and seen in a work environment or industry is almost impossible. If what you’re doing could really lead to a hundred-billion-dollar outcome, then for a long time it shouldn’t be noticed by too many people—otherwise competition emerges and you never reach that scale. So if you’re doing the right thing, you’ll likely be in a state of "not understood, not seen" for a long time. If not, you’re probably doing something wrong. That may sound extreme, but it’s what my investor taught me. He said the reason Amazon started with only books was that the less visible you are, the farther you can go.
Often, doing the right thing means being temporarily invisible. Why did Kuaishou start with only GIFs, not video? Why did Facebook start only with school social networking, not professional networking? It wasn’t that they didn’t want to—they wouldn’t have been seen if they had. If "not being understood" is an indicator of doing things right, then you can calmly face every instance of not being understood, every single day.
👩🏻 Ronghui
You seem very emotionally stable, without too many ups and downs, not affected by short-term results.
👦🏻 郭振宇
Right, my expectations are very low. I’m clear about my goals, but I keep expectations very low for each specific action and outcome. If you have high expectations for every move, you won’t make contingency plans, you won’t have margin to adjust—you’re basically gambling. Entrepreneurship requires discipline, calculation, and reserved capacity.
👩🏻 Ronghui
I have another question. You serve clients from all over the world, many of whom you might never have been able to reach otherwise. After serving these clients, how has your understanding of the world changed?
👦🏻 郭振宇
I’m super optimistic. I think the world is much better than we imagine. I have users in Palestine and Pakistan, both living in war zones, but in serving them I feel that people of different ethnicities and backgrounds all radiate a kind of hope for life and vigorous vitality. By comparison, China is entering a cyclical downturn. Chinese friends haven’t experienced cycles in the past 30 years, so suddenly hitting an adjustment, people feel unadapted. But my clients from places like Palestine, Pakistan, and Israel still face life with longing—that moves me deeply.
Because of this, I did some simple research, and with AI’s help research is easier now. I found that globally over the past 30 years, wealth inequality has actually been shrinking. Though many Americans may feel the gap is widening, World Bank data shows the gap between people in the DRC, Liberia, Zimbabwe and those in the UK is narrowing. The gap between countries and regions worldwide is clearly converging. Though internally the US and China may be different cases, this discovery gave me great confidence.
People often say "there won’t be another China"—that may be true. But I still believe humans are remarkable. Different cultures and regions at different stages will have their own creations and civilizations. I’m full of confidence about many countries and regions, something I didn’t feel as viscerally before. Unlike the politicized macro narratives in news reports, real individuals haven’t given up on their lives. They’re living happily, working hard—that’s my biggest feeling.
👩🏻 Ronghui
What you said reminds me of the book Factfulness, which also discusses similar statistics. Many conditions we think are terrible have actually made huge progress over past decades. It’s different from what we imagine—like you said, many things are still continuously developing.
👦🏻 Koji
Bill Gates recommended The Better Angels of Our Nature, which makes a similar point. Though the news is full of crises every day, from a long-term, aggregate data perspective, human living standards and safety are continuously improving. Various crises and diseases are decreasing.
👦🏻 郭振宇
Speaking of entrepreneurial pressure and mindset, I have to say I might be too relaxed. But I remember we had a Palestinian e-commerce client, one of our very early users. When he first connected with us, there was still fighting in his area. His business was buying small goods from Yiwu and transporting them daily in six vans to sell in the Gaza area. Unable to build a website or use other channels, he posted photos on Facebook, used our service to run ads, local residents saw them and messaged him on Facebook circling the items they wanted, and he’d drive and deliver them.
At that moment I thought: we have nothing to be anxious about. Compared to his business, ours is safer, easier, and growing faster. And he was still going at it energetically in that environment, genuinely happy, just focused on doing the business well and serving everyone, in a very optimistic state.
👦🏻 Koji
Thank you so much Zhenyu for joining us at Crossing today. Your sharing included not just business and strategic insights, but also how to live better, how to gain energy, and what entrepreneurship means to you—very inspiring for me, and a really enjoyable conversation. Looking forward to having you back.
Finally, a reminder that Sandwich Lab is always hiring in Hangzhou. If you’re interested in what they’re doing, you can find contact info in the comments. Thanks Zhenyu!
👦🏻 郭振宇
Thank you all, bye bye.
👩🏻 Ronghui
Bye bye.
References
[1] Lexi: https://lexilexi.ai/