Midway through 2024, when Guo Zhenyu left Alibaba to build a company in the agentic artificial intelligence space, he faced a key decision: build a fully autonomous agent capable of completing tasks independently, or develop a copilot-style assistant that keeps humans in control?
At the time, the idea of fully autonomous AI agents was still fringe. Foundation models were evolving, but few believed they could handle end-to-end tasks without human oversight. Most investors favored the safer bet: human-led, AI-assisted systems. Guo, however, had a different view.
Before founding Sandwich Lab, Guo spent over a decade working on autonomous driving and robotics. While pursuing his PhD, he founded a hardware startup called Neurio. He later joined Postmates, one of Silicon Valley’s largest food delivery platforms, where he launched an internal project to build sidewalk delivery robots.
To Guo, autonomous vehicles, delivery robots, and AI agents all serve the same fundamental purpose: building fully automated systems that help users complete tasks. Autonomous vehicles and delivery robots transport people or food from point to point. An AI agent, similarly, should deliver results and not just assist.
Sandwich Lab was officially founded in mid-2024. By March this year, it had launched its first product: Lexi.
Lexi is an advertising-focused agentic platform offered on a subscription basis starting at USD 200 per month. It provides small business owners with end-to-end ad campaign execution within Meta’s ecosystem.
The process is designed to be simple. For example, a bakery owner only needs to input a monthly budget and define the target radius from the storefront.
“Everything after that—analysis, image creation, copywriting, ad buys, optimization—is handled automatically by Lexi. The baker just needs to check how many new customers came in and how much revenue was generated,” Guo said.
Lexi targets real-world business scenarios. Its typical users are overseas small business owners who have never run ads before, whether they operate online or offline.
Guo observed that many small businesses begin with a single skill or goal of earning a living. “They can work hard and generate high revenue, but to scale, they need more complex capabilities,” he said.
Three months after launch, Lexi said its paying users span 94 countries and monthly revenue growth has exceeded 150%.
In parallel, Sandwich Lab reportedly closed a new financing round, bringing its total funding to over USD 10 million. Backers include 5Y Capital, Gobi China, and the Alibaba Entrepreneurs Fund (AEF).
Users don’t need to know what an agent is
Lexi’s defining feature is “continuous delivery,” meaning it requires no human intervention.
The product follows a prescriptive design. Users aren’t shown which creatives Lexi runs, and there are no interface buttons to modify campaigns.
To launch an ad campaign, users input only basic information: company details, budget, and duration. The system handles market analysis, asset generation, strategy setting, and iterative optimization.
Sandwich Lab positions Lexi as a revenue generator, directly accountable for business results. Advertisers set a budget and campaign period, then monitor revenue.
“Users only need to know Lexi helps them make money. They don’t need to understand how it works or even what an agent is,” Guo said.
“Running a business isn’t about a single day. People care about sustained revenue growth,” he added. For Guo, the biggest opportunity in this AI paradigm lies in shifting from one-off outputs to continuous operations.
At launch, Lexi focused on Meta’s ecosystem rather than Google or TikTok, and it was a choice that raised some eyebrows.
Guo’s response challenges conventional thinking. He argues that the differences between Facebook and Google are even greater than those between Facebook and a traditional TV station.
His reasoning centers on the platforms’ commercial logic.
Google is intent-driven: users type search queries, often with clear purchase intent. Advertisers focus on keyword targeting and intent matching.
TikTok, as a recommendation platform, pushes content to users based on predicted interests. But these preferences may not always translate into buying behavior.
“The biggest challenge for automated advertising on TikTok isn’t bidding. It’s about pushing the right video that captures attention. Ultimately, it’s a creativity contest, and that’s something AI still struggles to master,” he said.
Meta’s monetization system is less mature than Google’s or TikTok’s. That’s exactly where Lexi can add value: platforms that require strategic judgment and targeting.
Facebook remains a critical customer acquisition tool for many overseas small businesses. Over 130 million of them have run ads on the platform. Yet most agencies offer only account setup and website development, which are services that don’t guarantee ad performance.
Agency fees are high, often ranging USD 6,000–20,000 per month. Lexi aims to fill that gap.
Starting at USD 200 per month, Lexi is priced low enough to allow small business owners to test ad performance from scratch and see real returns, even without prior experience.
Its operating model is akin to quantitative trading: Lexi runs and tests multiple strategies simultaneously. Some may lose money temporarily. That’s considered part of the process.
Rather than choosing the “best” platform, Lexi targets the one where AI adds the most value. By operating in an uncertain market and playing the long game, it seeks to outperform what individual users could do on their own.
Many unexpected customers
Guo is candid in conversation. “I may not have figured this out fully,” he said, noting that his startup is driven by personal conviction. “This is a very personal idea. It can’t be fully explained by startup logic. I just want to do it.”
A recurring theme in his thinking is “redistribution,” a term he uses often.
“Alibaba’s mission of making it easy to do business anywhere influenced me deeply. With Lexi, I want to use technology-driven productivity gains to redistribute opportunities to people who otherwise wouldn’t have access, and make society a bit better.”
Lexi’s intended users were clear from day one: small business owners with products or services but no online marketing capabilities.
While Lexi currently operates as an ad-buying agent, Sandwich Lab’s ambitions go further. Guo emphasizes that ad buying is simply the first proof of concept.
The team is preparing a new product for release by year-end. It will serve a different scenario but be built on the same foundational principles.
“What we’re really building is a system of automation and productivity that can be applied across industries and use cases,” Guo said. Wherever efficiency can be improved, the system can intervene and reshape workflows.
Guo follows Lexi’s own recommendations, using it to acquire users and guide market strategy. When the system detects higher willingness to pay in a specific region, the team tailors its product for that market.
Today, over 60% of new users come through Lexi’s own campaigns.
For instance, in North America, Lexi found that local service providers, such as gyms, beauty salons, were 40% more willing to pay than e-commerce clients. These businesses rely heavily on local foot traffic, which Lexi can precisely target through Meta.
In response, the team optimized features for local services, including geofencing, localized copy generation, and Google Maps integration.
Asked whether he trusts Lexi’s strategies, Guo didn’t give a straightforward yes or no. “It’s not about trust. It’s about understanding why the AI suggests a particular strategy,” he said.
In 2025, OpenAI released GPT-5. Hallucinations persist, largely because large models understand the world through language, not numbers. Like humans, they can let intuition override evidence.
In commercial settings, whether AI can objectively incorporate feedback, or whether it will mirror human biases, remains an open question.
There are many ways to build in the large model era. Sandwich Lab chose a different path: don’t let AI make one-off guesses. Instead, build continuous feedback loops and evaluate outcomes using revenue.
That approach has led to many unexpected users. Early Lexi adopters include a team-building drum studio in the UAE, an Australian firm serving people with disabilities, and a business in Africa selling dreadlock extensions.
These discoveries represent deeply meaningful moments for the team. “Before Lexi expanded into these markets, I wouldn’t have known those customer profiles existed,” Guo said. “Even with Palestine at war, we have a customer there selling Yiwu goods on Facebook. He drives them to the border daily. He emailed us saying Lexi helped him sustain his income during the hardest times.”
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Deng Yongyi for 36Kr.