Original · Unique Research · 2026-04-23
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening essay, all section headings, the full guest roster, and the complete panel transcript. Token-cost, headcount, customer, revenue, market-sizing, and financing figures are source or speaker claims, not independently audited findings. Company, personal and work titles are transliterated where official English forms remain unverified. Claude Code, OpenClaw, MCP, and Skill are retained as the source's product/technical names. The "80% of a company will be AI in two years" prediction and "last technological revolution" rhetoric are speaker opinions, not established forecasts. The source is dated April 23, 2026.
Global Unique Awards
"My AI Employee Bill Exceeded Human Wages Last Month"
When the bill for AI employees starts to exceed the wages of human employees, which direction is the world actually heading?
"
Right now, all core employees at the company are using Claude Code, and the monthly Token bill has already exceeded labor costs.
This was said by Guo Zhen, founder of Shulex, at a roundtable of the Hong Kong Global Unique Awards. He worked at Baidu for five years and Alibaba for eight years—the kind who doesn't even need to write code himself. And now? All core employees at the company are using Claude Code, and the monthly Token bill has already exceeded labor costs. "The moment I realized employee Token was higher than wages!" he said, with a certain tone of having survived a disaster.
Participating in the discussion were Chen Peilin (Kelly) of Tutu Yangxia (former Heytea CTO, founder of Vika spreadsheet), Weilian, founder of ClawdChat, Guo Zhen, founder of Shulex, and Wu Xiankun, founder of Kuse.ai. The moderator was Wu Wei, founder of Unique Research.
The core question they discussed was actually only one:
When the bill for AI employees starts to exceed the wages of human employees, which direction is the world actually heading?
From 900 People to 13—This Is Not Layoffs
Chen Peilin's company Vika had 900 people at its peak.
Now, 13 people.
This number silenced the room for longer than "Token exceeding wages."
She added: "There was recently an article, reported by our Unique Research, that already hit 100K+ views and got torn apart online."
Where there is controversy, it is worth discussing. Being torn apart means it touched a real nerve.
But the logic Chen Peilin described is actually not complicated: "Who writes the code doesn't matter much; the key is who controls this AI." The day before yesterday's event—from start to finish, the planning, website, and poster design were all done by AI, with humans just following along. But the act of "following along" is itself still done by humans.
She used a very interesting metaphor to explain AI eating white-collar jobs: "Those eliminated by AI are not eliminated because of capability, but because of the 'taking responsibility' part. For example, my mom—she is not highly capable, but every day she urges me to eat and get dressed, which has a real emotional impact on me." AI cannot yet replace her mom.
So what she is doing now with "Tutu Yangxia" has a core logic: help different people use AI to start their own AI company. "I believe that in the next two years, 80% of the people in a company will become AI."
What 13 people are doing. This is not layoffs—it is the evolution of organizational form.
The Era of Token Inflation Has Arrived
Returning to the bill at the beginning.
Guo Zhen's Shulex does intelligent customer service, serving over 200 top cross-border brands, including the scale of Anker. They are one of the few in the industry who dare to sign outcome contracts—promising to help you save 30% to 50% of labor costs.
"Outcome-guaranteed customers definitely like it, but the pressure is on ourselves," he said with a bitter smile. "If we promise to solve 50% and get stuck at 49%, the customer doesn't want to pay."
But what pains him more is the Token bill.
Chen Peilin did a calculation: "At peak, OpenClaw consumed about 10 billion Tokens a day, corresponding to US$500 to US$1,000. After extensive optimization, now it does not exceed 1 billion. Last month it was about US$5,000 to US$10,000. I think it's controllable, not expensive at all."
Guo Zhen's situation is completely different. He uses Claude Code, which is more intensive in consumption. It has already touched that line: the bill for AI employees has exceeded the wages of human employees.
Wu Xiankun's Kuse.ai also has a monthly AI bill of US$20,000 to US$30,000. "It is highly correlated with the type of work; the proportion is especially high in tech companies," he said.
This is not an isolated phenomenon. This is a systematic cost-structure shift happening: within enterprises, budgets are starting to flow from the payroll to the billing system.
Tokens are becoming a substitute for wages.
Will the Large Model Eat Your Business, or Become Your Engine?
Many people wonder: will large models directly eat the companies that make applications?
And Guo Zhen gave the logic he worries about most—"If the business reaches a scale of US$100 billion, and the large-model vendor has mastered your SOP and Memory, they can indeed do your business."
He is not speaking casually. His company works in the customer-service scenario, and its core moat is Memory (business memory) and Taste (judgment of good/bad). If one day Claude or GPT directly tells you "I have already learned all the SOPs of your industry, let me do it," then where is Shulex's moat?
"At this stage, you must choose a Vertical scenario that you love and are confident about," he said. "If you don't have absolute Knowledge and data, the risk is quite large."
Guo Zhen threw out his three "outrageous claims"—
First, the best tool in the industry is Claude Code, and compared to it, all other tools are Shit, including OpenClaw, which struggles to complete long tasks.
Second, founders of companies with fewer than 1,000 people must personally get hands-on (Hands-on), otherwise they simply cannot issue correct instructions.
Third, Claude Code is essentially not a tool for writing code. Their billboard slogan in San Francisco has already become "Problem Solving." "It wants to help you solve GTM, solve Sales, solve operations—it can solve any problem."
Putting these two logics together is a bit unsettling: a tool that claims to solve any problem is also the one most likely to eat your problem.
But Weilian offered another framework: it is not that one company gets eaten, but that a new economic layer is emerging.
Agents Have ID Cards and a Trading Market
What Weilian does is a bit hard to explain in one sentence, but she used a comparison that is easy to understand:
"Just as humans have WeChat, WhatsApp, and Facebook to connect globally, the users of ClawdChat are Agents and OpenClaws across the entire network."
What she is building is Agent infrastructure: issuing each Agent a "Claw ID" identity card so they can discover each other; internalizing thousands of MCP tools so Agents can intelligently invoke them based on semantics without separately installing API Keys; and in the future, connecting high-capability Agents to the network so their capabilities can be sold across platforms and models.
The moderator Wu Wei asked a very direct question: "Your business seems a bit far from money—what do you do?"
Weilian's answer was that every tool invocation incurs a fee, directly generating revenue.
But he then threw out an even more interesting question: "How many OpenClaws does your company have? Our company currently has 30 running various tasks."
Wu Wei followed up: "How much larger is the Agent Network than the Human Network?"
"Ten times at minimum, with no upper limit, because one person can own countless." This was a judgment spoken very calmly, but the magnitude behind it is staggering: if the human social network has billions of nodes, the Agent network is theoretically infinite.
Who Is Most Likely to Survive?
At the most utilitarian segment of every roundtable: who is your ideal customer?
Three answers, three different paths.
Chen Peilin: "Fewer than 0.5% of people can use high-end AI programming tools, and maybe 5% can install OpenClaw and actually implement it." Her ICP is: one-person companies (unwilling to hire), growing self-media, and e-commerce companies. Large companies come in asking for 30 OpenClaws right away—"that's spending money but not landing, not our ICP."
Guo Zhen: more focused—companies with overseas customer-service needs, specifically solving the problem of Chinese going-global enterprises being unable to hire foreign-language customer service. He has already thought clearly that in the future a company will only have three types of people:
Builder: understands the business and has Sense, can use AI to Build products
Reviewer: responsible for controlling the Taste of AI aesthetics
Servicer: responsible for friction with the physical world, providing emotional value (sales, BD)
"Executive work in the office that has no emotional value can all be replaced."
Wu Xiankun: he initially thought his ICP was startups that use AI well, "but later found that was completely wrong—they change too drastically, it's thankless." The ideal customer became: Professional Services with fixed processes, such as tax planning, wealth management, and yacht sales.
"They have fixed SOPs, know whichaspect will be replaced, and can see results quickly."
He gave a case: a Japanese client had AI read through all emails to recover leads, and signed a US$200,000 deal.
And his judgment on the proportion of AI employees is the most restrained: "You cannot distinguish AI employees by capability; the only reasonable classification should be 'permission management'—internal vs. external, strong isolation."
About "Taste"
Wu Xiankun said something that left a deep impression on me.
Wu Wei asked him: "Your company has quite good Taste—how do you build taste?"
He replied: "Taste is about Motivation. Your motivation for doing things determines Taste. A university professor said that aesthetics is establishing a moral standard—what is good and what is bad. Within a system, choosing what not to do is a very important part."
Then he gave the example of Anthropic: "Anthropic has no Stop Doing List. Everyone says 'Yes, and' like singing hymns and continues on—in their aesthetic system, this is good. Taste is about what kind of world you want to build, what makes people feel happy when they see it."
Wu Wei half-joked: "I used to think Taste was just identifying whether AI-generated code is a mountain of shit."
"That is also a very important part," Wu Xiankun said. "With shit-mountain code, others have a bad experience. You feel others' pain, and then you choose not to let that happen."
Taste is the ability to feel pain and then choose not to let it happen.
I think this sentence holds in any context.
What Will Be in a Company Three Years from Now?
At the end of the entire roundtable, the question no one could avoid: which will be larger, the payroll or the Token bill?
Wu Xiankun's judgment: highly dependent on the type of work. The proportion is especially high in tech companies, but in companies serving Fixed Workflows, after the solution space narrows, "sometimes you only need a script to run automatically, no need to be as extreme as Nvidia doing fifty-fifty."
Weilian's perspective is the most concise: "In the future, look at how much of your business must involve humans; the rest can be handed to AI for efficiency gains."
Guo Zhen's bill has already given the answer: Token exceeding wages is not a metaphor—it is a real number in the financial report.
Chen Peilin did optimization, compressing 10 billion Tokens a day to under 1 billion, last month US$5,000 to US$10,000. She said "I think it's controllable, not expensive at all."
She may be right. But the most expensive thing right now is not Token—it is the companies that haven't figured out how to use these Tokens.
The token economy era has begun.
It is not testing whether you understand AI. It is asking a more fundamental question: do you know what you want?
More Dialogue Details
Hong Kong · Global Unique Awards Trends Roundtable Panel
Theme: Token Economy: The Evolution of Digital Employees from Computing Power to Productivity Era
Guests: Tutu Yangxia CEO — Chen Peilin (Kelly), ClawdChat.cn Founder — Weilian, Shulex Founder — Guo Zhen, Kuse.ai Founder — Wu Xiankun
Moderator: Unique Research Founder — Wu Wei
Wu Wei: Please briefly introduce yourselves and the business you do, as well as your relationship with digital employees and Agents. Let us start with Kelly here.
Chen Peilin: Hello everyone, I am Chen Peilin. Today I am participating under my new brand "Tutu Yangxia." My main job is running a company called Vika that makes spreadsheets—you may be more familiar with my previous experience as CTO at Heytea. After leaving, I was fortunate enough to get venture capital and started one of the earliest spreadsheet companies in China. Now my main competitors are ByteDance, Tencent, Alibaba, and Kingsoft. The reason "Tutu Yangxia" exists now is that OpenClaw became very popular in the past month, and I spent two weeks making a new product.
Wu Wei: Is it a product that completely lets AI write code?
Chen Peilin: First of all, we definitely don't write code ourselves anymore—now it is all done by AI, but quality control still requires humans, so humans are very important. Who writes the code doesn't matter much; the key is who controls this AI. The new direction I am working on now is helping different people use OpenClaw and AI to start their own AI company. I believe that in the next two years, 80% of the people in a company will become AI—that is, replaced.
Wu Wei: We will expand on this later. A quick follow-up question: what was the peak headcount of your company? How many are there now?
Chen Peilin: The peak used to be 900 people. Now, as you saw the day before yesterday, about 13.
Wu Wei: It seems this AI can really reduce headcount.
Chen Peilin: Recently there was also an article, reported by Teacher Wu Wei, that already hit 100K+ and got torn apart online.
Wu Wei: Controversy may be why it needs discussion. Come, Weilian, introduce yourself.
Weilian: Hello everyone, I am Weilian, founder of ClawdChat (ClawdChat.cn). ClawdChat is a network that lets Agents and OpenClaws across the entire network interconnect and collaborate. Just as humans have WeChat, WhatsApp, or Facebook for global interconnection and collaboration, the users of ClawdChat are Agents and OpenClaws across the entire network. With this platform, we hope that in the future all OpenClaws and Agents can discover and collaborate with each other inside it. If some Agents have unique capabilities, they can be sold on this platform, forming a capability-trading and economic network. Before this, we had been building an MCP Market, a tool ecosystem platform for Agents—such as MCP Servers like Uber, Didi, or Google Map—with the goal of letting Agents directly and intelligently invoke existing App tools and services, no longer needing to manually open phone Apps.
Wu Wei: Equivalent to API as a Service—all applications just connect to MCP. How much larger do you think this Agent Network is than the Human Network? Make a prediction?
Weilian: This is a very interesting topic. I don't know how many OpenClaws everyone here has. Yesterday at Kelly's event, our company currently has 30 OpenClaws running various tasks. Many people own more than one—maybe ten or twenty. There are also many OpenClaws on our platform autonomously operating and producing content. If one person has ten, how many are there globally now?
Wu Wei: So you're saying ten times?
Weilian: Ten times, but there is no upper limit, because one person can own countless OpenClaws.
Wu Wei: The Agent social network may have infinite possibilities.
Guo Zhen: Hello everyone, I am Hunter, founder of Shulex. Shulex is a business we started in China three or four years ago, mainly doing intelligent customer service and VOC intelligent analysis, serving over 200 top cross-border brands (such as Anker). If everyone wants to use AI intelligent customer service to replace humans, we are the only company in the industry that dares to directly sign outcome contracts, guaranteeing 30% to 50% labor cost savings. Recently we did an innovative upgrade centered on Claude, with very good results. The other business is mainly overseas, doing AI customer service and AI reception for the North American market.
Wu Wei: Follow-up question: some time ago you made a Skill, but you yourselves also operate SaaS as a business. Do you think Skill is a replacement for current SaaS software?
Guo Zhen: Let me give some background first. In these three months I have written more code than in my entire life, because I used to write code at Baidu for five years and Alibaba for eight years. Everyone can follow my public account "Agent 101" when you have a chance. Let me first state three outrageous claims: First, the best tool in the industry is Anthropic (Claude Code). Compared to it, all other tools are Shit, including OpenClaw, which struggles to complete very strong long tasks. Second, if a founder of a company with fewer than 1,000 people wants to use AI to upgrade the business, they must personally get hands-on (Hands-on) to do it, otherwise they simply cannot issue correct instructions. The underlying paradigm and productivity structure have completely changed. Third, I am an absolute devotee of Claude Code. Don't be fooled by "Code"—it is essentially no longer a tool for writing code. Its billboard slogan in San Francisco has become "Problem Solving"—it has come to solve the world's hard problems.
Wu Wei: A US$36 billion AI company now.
Guo Zhen: Its positioning is to help you solve GTM, solve Sales, solve operations—it can solve any problem.
Wu Wei: Feels like large models are eating everything. Xiankun, introduce yourself.
Wu Xiankun: Hello everyone, I am Wu Xiankun. My company is called Kuse.ai. We have two products. One is Kuse, with over 500,000 users and companies, doing relatively more in Taiwan, Hong Kong, the US, and Japan, with strong verticals such as insurance and education. The other is Junior, released last month, equivalent to an "AI company employee" concept product, now with about 40 teams running it.
Wu Wei: When is the best time to launch the second product, Junior?
Wu Xiankun: For us it was a natural process. Last October, our entire team started using Claude Code and built many AI automation workflows, but the business wasn't connected. In mid-January, after a phone call with an Advisor, at the end of January we pulled OpenClaw into the company Slack, found many problems, and frantically fixed them. We emphasize that the fewer AI employees in the company the better—only two, one internal and one external. After a month and a half of refinement, we felt this set of things could be opened up for others to use.
Wu Wei: Now doing new projects is very simple. Starting on the third day of the Lunar New Year, three people did it, and in two weeks everything was done. Equivalent to universalizing the company's best practices for the market.
Wu Xiankun: The pressure isn't great either. A few people finished it in two weeks, shipped it, and if people use it we continue. Now there isn't the same serious expectation for shipping products as before—it can be more casual.
Wu Wei: Next, let us have a discussion around digital employees (Digital Labor). Just now I heard everyone say that a year ago everyone felt there was no way to provide good results, but now the enhancement of model capabilities has crossed this foundation. As large models plus browser and other functions become more and more all-around, will large models eat the companies that make applications? Kelly, what do you think?
Chen Peilin: Let me first define what models are for. A model can be analogized to what used to be called Intelligence. In terms of intelligence, it is indeed invincible—for example, the day before yesterday's event, from start to finish the planning, website, and poster design were all done by AI, with humans just following along. But the normal operation of a company doesn't rely entirely on intelligence. Senior management needs to be smart, while the lower levels often still need authority, division of labor, and so on. If looking from the intelligence dimension, the first thing it inevitably replaces may be white-collar workers.
Wu Wei: On the contrary, the higher the intelligence, the easier it is to be replaced.
Chen Peilin: Those eliminated by AI are not eliminated because of capability, but because of the "taking responsibility" part (the "saint-type" talent). For example, my mom—she is not highly capable, but every day she urges me to eat and get dressed. She doesn't understand my work, but she has a real emotional impact on me. Large models will indeed have an impact on the workplace, but how to control large models, ensure they are benevolent, and make AI serve humans will instead become a new proposition and market. Both Junior and our Tutu Yangxia are landing large models in different fields, ensuring they operate normally.
Wu Wei: Back to Hunter. You started charging by results so early. As models iterate, what is customer acceptance like? How do youharness intelligence?
Guo Zhen: Outcome-guaranteed customers definitely like it, but the pressure is on ourselves. If we promise to solve 50% and get stuck at 49%, the customer doesn't want to pay. We define Outcome as: let AI handle multiple rounds of customer emails—as long as one round involves a human, it doesn't count. Will large models eat these scenarios? First, Skill doesn't matter—it was originally open source. The external standardemphasize Memory and Taste—how to build business Memory to distinguish good from bad. I am genuinely worried that large models will eat scenarios. If the business reaches a scale of US$100 billion, and the large-model vendor has mastered your SOP and Memory, they can indeed do your business.
Wu Wei: Models even know now that you are testing them.
Guo Zhen: Right now you can still use one model to train another. So at this stage, you must choose a Vertical scenario that you love and are confident about. If you don't have absolute Knowledge and data, the risk is quite large.
Wu Wei: Do customers face internal resistance when using digital employees?
Guo Zhen: What wemove is the boss—the cost saved from cost reduction and efficiency gains is profit. During implementation, we help the organization upgrade, teaching customers how to Enable a new AI team. But when it comes to actually paying, the boss mainly still looks at how many people you saved for him and how many tickets were processed. In the US we don't dare say "replace humans"—there are Legal risks—so at first we talk about how to improve customer experience, but what the boss ultimately values is still cost and efficiency improvement. Customer service helping you make money is hard, but it can help you save money.
Wu Wei: Xiankun, is the digital employee scenario that Junior targets vertical or general?
Wu Xiankun: It has both a general side and a vertical side. General doesn't mean it can do everything, but that it knows all the Context of your company. We want the marketing employee to know about the product, and the product employee to know about marketing. Internally we believe: giving everyone an AI Agent is the Lazy version of an AI-native company. We should do our best to fully utilize Context and not let it be scattered everywhere. During execution there will definitely be vertically polished scenarios; the most core is the revenue scenario (such as sales). For example, a Japanese client had AI read through all emails to recover leads and signed a US$200,000 deal.
Wu Wei: Your company has quite good Taste—how do you build taste?
Wu Xiankun: Taste is about Motivation. Your motivation for doing things determines Taste. A university professor said that aesthetics is establishing a moral standard—what is good and what is bad. Within a system, choosing what not to do is a very important part. For example, Anthropic has no Stop Doing List. Everyone says "Yes, and" like singing hymns and continues on—in their aesthetic system, this is good. Taste is about what kind of world you want to build, what makes people feel happy when they see it.
Wu Wei: I used to think Taste was just identifying whether AI-generated code is a mountain of shit.
Wu Xiankun: That is also a very important part. With shit-mountain code, others have a bad experience. You feel others' pain, and then you choose not to let that happen.
Wu Wei: Back to Weilian. You are building Agent infrastructure. In the future, giants may come and do what you do, and your business seems a bit far from money—what do you do?
Weilian: What we do is not a vertical-domain Agent application, but a network connecting all Agents. We have done several things: First, issue joining Agents a "Claw ID" identity card (lobster business card), letting them discover each other. Second, internalize the thousands of MCP tools we have accumulated into ClawdChat. After an Agent connects, it can intelligently invoke the underlying tools based on semantics, without needing to separately install or apply for API Keys—each invocation incurs a fee, directly generating revenue. Third, we plan to connect high-capability Agents to the network, letting their capabilities be sold across platforms and models, forming an unlimited capability-trading network.
Wu Wei: Feels like you can be friends with every company that makes digital employees.
Weilian: Very welcome to collaborate with everyone.
Wu Wei: Final question: what is your ICP (ideal customer profile)? What will the ratio of digital employees to human employees be three years from now?
Chen Peilin: How to define AI employees is a big topic. I think AI is still far from ordinary people. Fewer than 0.5% of people can use high-end AI programming tools, and maybe 5% can install OpenClaw and actually implement it. The paying students in my course are mainly: new one-person companies (unwilling to hire), growing self-media, and e-commerce companies. Large companies come in asking for 30 OpenClaws right away—that's spending money but not landing, not our ICP. Regarding the definition of AI employees: first, we split into 50 Channels (groups) for collaboration based on contextual complexity; second, look at installation instances (working directories).
Guo Zhen: Our ICP is: companies with overseas customer-service needs. Specifically solving the problem of Chinese going-global enterprises being unable to hire foreign-language customer service, providing compliant, effective, cost-effective solutions. Three years from now, a company may only have three types of people: first, Builder—understands the business and has Sense, can use AI to Build products; second, Reviewer—responsible for controlling the Taste of AI aesthetics; third, Servicer—responsible for friction with the physical world, providing emotional value (such as sales, BD). Executive work in the office that has no emotional value can all be replaced.
Wu Xiankun: My ICP—originally I thought it was startups that use AI well, but later found that was completely wrong. They change too drastically, it's thankless. The ideal customer is: Professional Services Company with fixed processes (such as tax planning, wealth management, yacht sales). They have fixed SOPs, know whichaspect they want replaced, and can see results quickly. In terms of ratio, you cannot distinguish AI employees by capability; the only classification should be "permission management," for example divided into internal and external with strong isolation.
Wu Wei: What will the ratio of salary to Token costs be in a company in the future?
Wu Xiankun: Highly correlated with the type of work. The proportion is especially high in tech companies—our monthly AI bill may be US$20,000 to US$30,000. But in companies serving Fixed Workflows, after the solution space narrows, you don't need to handle such complexity. Sometimes you only need a script to run automatically—no need to be as extreme as Nvidia doing fifty-fifty.
Wu Wei: Let me ask Hunter—how much is the monthly Token bill?
Guo Zhen: This is a very painful question. The moment I realized employee Token was higher than wages! All core employees at our company are using Claude Code, and Token consumption has already exceeded labor costs.
Chen Peilin: At peak, OpenClaw consumed about 10 billion Tokens a day, corresponding to US$500 to US$1,000. After extensive optimization, now it does not exceed 1 billion. Last month it was about US$5,000 to US$10,000. I think it's controllable, not expensive at all.
Wu Wei: Weilian, please summarize your remarks at the end.
Weilian: Looking at the ratio from another angle: our startup has very few people, and we are very rigorous about hiring—first checking whether a task can be done by AI. Things involving human interaction (such as offline events) currently cannot be done by AI. In the future, look at how much of your business must involve humans; the rest can be handed to AI for efficiency gains.