Original · Unique Research · 2026-06-30
Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening narrative, themed sections, and complete panel transcript. Pricing, accuracy, and capacity figures are speaker self-reports attributed to the named founders, not independently verified findings. Company and person names are preserved as source attributions.
AI Industry Observation
A roundtable on AI, business, and human nature — the answer was simpler than expected.
"The food-AI person says: eating. The legal-AI person says: travel. The Web3 person says: sweating through sports."
The answer was simpler than expected.
But if you'd sat through the whole discussion, you'd understand its weight. Because minutes earlier, they were discussing that famous Sequoia article, talking about how their own digital employees already replace the output of 15 people, talking about how AI runs business while you sleep, never takes leave, never pays housing fund, has no emotions.
Then they were asked: so what do you yourself want to keep?
Many lines in this roundtable are worth chewing on; I'll pick the most interesting.
From Selling Tools to Selling Outcomes — It's Already Happening
This March, a Sequoia partner wrote an article called "Services Is the New Software." The core logic is simple: for every dollar an enterprise spends on software, it's willing to spend six on services. Enterprises will pay for outcomes, not tools. So the next trillion-dollar company will be a services company disguised as software.
If you think this is just VC hype, CK will tell you they're already doing it.
CK is Singaporean, runs Cuber AI. He said flat out that what they deliver is outcomes; they don't call themselves Agents, they call them digital employees. They've run by this logic since last June; in July they'll fully launch full-time digital employees and temporary digital employees.
The pricing logic is simple: complete a task, and you pay him for it. A customer-service digital employee rents to you for 5,000 RMB a month. The benefit? 24/7 online, no housing fund, no leave, no complaining of fatigue, no emotions even when you scold him.
He doesn't even write code anymore. CK's take: going forward all development will be AI Native; no one will write code; people only tell these digital employees what to do on the business side.
Honestly, a boss can't fail the math on 5,000 a month for a customer service who never complains of fatigue.
But you'll surely want to ask: where do the replaced people go?
CK's answer is interesting. He said: don't keep thinking about eliminating people. Think back to when computers first arrived — didn't the hand-ledger accountants think they were doomed? But in the end, accountants are still here; the way of working changed.
"The digital employee handles 80% of operational work; what do you free your hands to do? Build relationships with clients, do group interviews, provide emotional value. These are things AI can't do."
Human-machine collaboration isn't a slogan; it's pulling your own time out of trivialities to do higher-revenue work.
Predicts Your Favorite Flavor With 86% Accuracy
Erik's company, Digitaste, is special. He doesn't make chatbots; he uses digital twins to help food and beverage companies test whether a new product will be popular.
For example, should this summer launch bayberry flavor or lychee flavor? Sweeter or tarter? His system tells the brand the direction.
The principle isn't complex. Most food decisions are subconscious; what you grab in the supermarket didn't go through your head, and relates to familiarity. He builds each person's digital twin through online surveys, implicit-association tests, and behavioral tests, then trains the model on 70% of 200-plus questions and validates on 30%, comparing against real human results. Current accuracy is 86%.
He said they're also integrating electronic noses, electronic tongues, even brainwave data, aiming to cross 90% soon.
But Ahaya asked a question.
She said there's a scenario AI can never have. For example, she especially loves osmanthus-flavored drinks, maybe less about taste, mainly because at 16, under an osmanthus tree, she exchanged a glance with the basketball-star in her class. That moment makes her want to buy osmanthus flavor whenever she sees it.
AI has no such memory. It can taste sweet or not, but it can't taste the flutter under a 16-year-old's osmanthus tree.
Erik's answer was candid. He said their system gives directional judgments — bayberry vs lychee, sweeter vs tarter — but for the exact formula, food engineers are still needed. Those parts belonging to carbon-based-life experience are indeed where AI can't reach.
I think this line hides an important fact: the stronger AI gets, the clearer the things only humans can do become.
Large Models Strike Downward; Does Vertical AI Still Have a Moat?
Max's ThinkSpace is a legal vertical application. Legal AI has a star company called Harvey, valued at $1.1 billion last year, with 1,300-plus clients globally, just landed in Singapore this June.
Ahaya asked him directly: Harvey walked in — is your advantage a moat or just a time difference?
Max's answer, I think, was more honest than most.
First he told a fact that stings peers. Last night a friend doing R&D at a model vendor messaged him: can we cooperate, let us train on your vertical data and lawyer-labeled data. He was at once angry and amused — this is my core know-how; how could I give it to you?
But look at it the other way: big vendors are fully entering vertical scenarios. When the underlying model is strong enough, does vertical AI just become a feature of the large model?
Max said there's still a gap — that famous Sequoia article also noted the huge gap between model capability and actual enterprise adoption. And enterprise customers are sticky; once they put all data, workflows, and collaboration on your platform, switching costs are high.
But he didn't dodge the more essential question: how much time does vertical AI actually have left?
Ahaya added an interesting point. In an enterprise's decision-making process, many things aren't direct outcomes but process things. How was this decision made? Which departments did it pass through? What's the boss's preference? What's each leader's decision basis? A general large model can gather some data in every vertical, but it can't get the human part of those decision bases.
"These are the real moat of an enterprise Agent. Technical barriers get caught up sooner or later; the barrier of understanding people doesn't."
If an Agent Can Be Shared in One Click
Xiong Wei wants to build the App Store of the Agent era.
He did Web3 before, with faith in open source and decentralization. His logic comes from a very concrete pain point: he uses Agents himself smoothly and can do all sorts of tricks, but once he wants to share these Agents with colleagues and friends, he's stuck. Because an Agent's skills and data are too tightly bound to his personal productivity environment; it can't be copied to others in one click.
So he wants to build an open Agent App Store. Here, you can train your own workflow into an Agent, strip out privacy data, export and list it in one click for others to buy. The platform handles underlying operations; you don't touch servers. His "open ecosystem" means everyone can participate, not one big company controlling everything: developers train Agents, users use them, sharers earn, the platform takes a small operation fee.
Someone once asked him: "Aren't you afraid of ending up as another 30%-commission App Store?" He said he believes in open source and doesn't want to build a closed empire.
But frankly, for this vision to truly land, too much is needed. Ecosystem building isn't like writing code; it's never a one-person job.
The Real Question Is What You Want to Keep
Back to the opening question.
Notice the common thread in these four answers: what they want to keep is experience, not output.
No one wants to keep writing code. No one wants to keep reviewing contracts. No one wants to keep managing a team.
It's eating. It's travel. It's sweating. It's feeling.
CK said on stage: the digital employee does 80% of operational work; 20% needs humans to give instructions. But that 20% may be what makes humans human. What you want to do, where you want to go, who you want to be with, who you think of when you smell osmanthus.
AI can't compute these.
"May every carbon-based life form have a wonderful life."
Saying this at an AI summit is a bit absurd and a bit sincere.
Talking about how many people AI replaces gets boring after a while. The interesting question is different: with the replacements done and your hands freed, what do you want to experience yourself?
More Conversation Details
Panelists: Erik Yang (Digitaste Founder); Max Ye (ThinkSpace Founder); CK Ng (Cuber AI Managing Director); Wei Xiong / 熊炜 (VertrAI CEO)
Host: Ahaya (DAJ Innovations Founder)
Why Start From Singapore and Go Global? What Matters Most?
Ahaya: Today's panel is "Starting From Singapore, Then Going Global." This came up across panels all day, so let's set aside geographic advantages other guests covered. In one sentence, why did you start from Singapore, and what do you value most here?
Erik: Hi, I'm Erik. For us, the most important thing about Singapore is the East-Meets-West local culture. For our taste research, sensory science has always been built on Western systems, but in Singapore we can well obtain samples to study East Asian tastes.
Max: I started a company in Singapore mainly because I've been here over 20 years, from school to work. In one sentence, Singapore is the intersection of China and the West. The legal industry is the same; it fuses rules from both sides — a good starting point.
CK: I'm Singaporean myself, so I won't discuss (objectively) why here. Mainly Singapore gives a great geographic position; second, it links the world, and we have quite favorable political support. So starting here and going to other countries is a strategic point. In terms of English and Chinese acceptance, Chinese companies coming to Singapore and re-going out is a conversion model. The traditions resemble China; we're all Chinese, mutual understanding is high. Taking a product transformation to Europe, the US, or Australia becomes much easier — a springboard.
Xiong Wei: I've been in Singapore four-plus years, previously Web3. Here we mainly focus on law, compliance, and politics. Due to geopolitics, everyone cares deeply about data security and sensitive info and won't make data public. Against this backdrop, Singapore provides a safe, stable environment backed by stricter, more complete laws. It also offers a multicultural environment well-suited to an AI company serving global users; global users' data ultimately lands in Singapore. I think Singapore is a great base for AI companies, especially global-facing ones.
How Do You See Sequoia's "Services Is the New Software"? Future: Deliver Tools or Outcomes?
Ahaya: I'll introduce Xiong's project later. Let me introduce myself — I also start from Singapore, now connecting Chinese entrepreneurs with ecological vision to global capital and markets. In plain terms, I provide Go-to-Market services for some Web3 and AI Agent projects. In future I hope to connect more ecologically minded entrepreneurs across tech, art, culture, and community, helping them truly reach global markets.
Second round: I don't know if you've seen it — this March a Sequoia partner wrote "Services Is the New Software." The core: the next trillion-dollar company is a software company disguised as a services company.
His logic: for every $1 of software budget, enterprises spend $6 on services. Enterprises pay more for "outcomes" and "services themselves" than for tools. So the most valuable future AI Agent company directly delivers services, not just a Copilot-like tool. I'm curious: are your AI Agents planning to move toward services, or have you already embedded some direct service delivery?
Erik: I've never really thought of us as a pure AI company. As a company modeling high-dimensional human taste and smell, AI is an accelerating tool and a technical revolution. What we care most is whether it's tasty, how consumers feel, whether what the client makes is tasty. For us, delivering high-accuracy results is always first.
I've studied Sequoia's view. Sequoia splits the offering into Intelligence and Judgement. Intelligence will accelerate human development, but Judgement — taste and experience — is something AI can't cover. So we won't leave many tedious hands-on tasks to ourselves; AI can replace them, and we focus on things truly meaningful, exploring the edge of consciousness.
So I think the 1 (tool) budget vs 6 (service) outcome premium will rise further, even 1 to 20. The 1-to-6 ratio is from the existing market. We use digital twins to help clients validate new products and accelerate food production. Before they could only test 10 of 100 products; now we test 60 or 70. Putting an incremental market in, I think there's bigger potential.
Max: I read that interview too. My first thought: we always say SaaS is "Software as a Service," itself a service. I feel Sequoia more meant: what's the difference between human-to-human service and software service?
In my view there's an L1-to-L5 process, from manual to eventually Agent-like automated task execution. The difference is how far executing this task differs from going to a real lawyer in our industry. My feeling — I chatted with Xiong yesterday — they now look at contracts by first having a GPT lawyer take a look. In this case, sooner or later most users' first approach will be asking software or AI.
That's the ToC scenario. But ToB differs: when you automate an Agent, if it errs you don't know where. So you need lots of upfront engineering. In document structure, with many files in, you focus on which passage or clause matters — human judgment needed, then continuous training. So one day AI may become our first service direction, but we're not fully there yet.
CK: My view is a bit different; we now deliver outcomes. We don't call ourselves Agents; we call them "digital employees." You can find me on LinkedIn, CK Ng.
My recent cases are all outcome-delivery. Our current Digital Employees purely deliver and look at results. Future pricing follows: like hiring someone to work, you look at results then pay. We've run this since last June. Sequoia also invested in us, so we're doing by this logic.
In July, our company will fully deliver FDE (Full-time Digital Employee) and TDE (Temporary Digital Employee) outcomes. Going forward all development will be fully AI Native; no one writes code. People are on the business side, including how to prompt these Agents. So we've moved to delivery and outcomes.
Future pricing is the same: complete a task, and I charge you. Recently in Southeast Asia we started similar AI Service, an outcome-delivery node, paying by results. Buy a digital employee doing customer service; maybe I rent it to you for 5,000 a month. The benefit: 7x24 online, no housing fund, no leave, no "I'm tired," no emotional output. No matter how you scold or push it, it finishes the work — that's outcome delivery.
Ahaya: CK, comparing a digital employee to a traditional employee, at what discount is the digital employee's wage?
CK: Before, at a Shenzhen roundtable, someone asked: "Is your digital employee expensive, or my person?" I summed it up: my digital employee will start very expensive. At first it needs to learn, like a new hire you must teach. But once taught and familiar, today's AI has developed to what we call RIL, Recursive Improved Learning. Anthropic just started; if you use OpenClaw or large models, especially top models, after a few instructions it writes "I am improving something" — it achieves iterative auto-learning.
So at first you'll find it expensive. But once you use it long and see results, it's not expensive, because it keeps delivering results. Initial investment is expensive. Like educating a child — university upfront is expensive; later when they find a high-paying job, your payback is relatively high.
Xiong Wei: I hadn't read that Sequoia article until Ahaya recommended it; very interesting. Before starting a company, I built OA and CRM systems inside banks. I developed systems for a business department (like accounting) to operate. So the article describes my past experience.
But now switching to AI, including vertical and platform, we've revolutionized that old way. I used to build a system for people and a department to use. Now AI inherently knows this bank's data architecture; I don't need to operate a system, just connect to raw data and it directly gets the result. So from "people use a system, then produce an internal enterprise-architecture process," it's now "directly give the enterprise the final result through an Agent." The change is huge.
Against this backdrop, we tried cooperating with CK on a recruiting scenario. Recruiting used to build a system for employees to enter data and HR to process. Now our recruiting Agent (a digital person) no longer does that old work; it's like an HR dealing directly with the boss. As boss, you just add its WhatsApp, tell it who you want, send your JD. This Agent searches LinkedIn and platforms for talent, chats with them about fit and willingness, then after analysis reports to you. The boss looks at a dashboard of high-fit candidates, and it schedules meetings and interviews.
The interview itself AI can't do yet, but as voice and image recognition improve, maybe AI does it too. Of course there's a paradox: if AI can do all this, why do we still need human employees? So I strongly agree with the Sequoia article. My background and current work are exactly this breakthrough point. Technically there's still much to break through, but I believe a huge revolution is coming soon.
How to Build the Agent-Era App Store? Open Source or Closed?
Ahaya: Following your answer, another question. You shared you want to build the Agent-era App Store. The App Store may be business history's most successful business — 30% commission, opaque rules, yet every founder depends on it. Your background is Web3, which builds decentralized ecosystems. So in future, what kind of Agent-era App Store do you want — a closed empire, or a truly open ecosystem?
Xiong Wei: Conclusion first: we definitely want an open ecosystem. I'm an engineer by background, with deep faith in software open source.
Why build an Agent App Store? Core reason: there's a pain point. I personally use Agents smoothly, but when I want to share this Agent stuff with colleagues, friends, or users who've never used Agents, it's hard to copy. Because the skills and data an Agent uses are tightly bound to my own productivity environment. We feel the biggest pain point is whether Agents can be replicated — can I take efficient things I do and one-click-copy my workflow, my Agent to others?
If Agents can be replicated, it should be an open, open-source ecosystem, not controlled by one big company that only it can serve. Everyone participates, training their work into Agents, stripping privacy data, and sharing their Agent.
So with this idea we're trying an Agent App Store. We want first to build a basic framework separating open data and private data. Finally, whether engineer, developer, or ordinary user, if you find your Agent very usable, you can export it in one click and list it on our Agent App Store for people to buy.
Our platform also provides underlying operational infrastructure. Developers or users don't manage servers; these Agents run on an open service platform. End users train, use, share; other users can buy and use on the platform.
How Does the Digital Human Predict 86% Success? Where Is Vertical AI's Moat Against Large Models?
Ahaya: Next, Erik. Introducing Erik: his product serves food and beverage with digital-human tools to test whether users will like a new product before launch. In your deck I saw a case of 81% accuracy predicting market acceptance of an osmanthus-flavored drink. First, please share the principle — why can a group of digital humans predict whether the market likes a product? Use simple principles.
Second, my question: AI can't replace carbon-based life because it has no taste or emotion. For example, we like a food (say osmanthus drink) not just for taste but for memory — like 16 years old, under an osmanthus tree, exchanging a glance with the basketball star. That moment makes me buy osmanthus drink whenever I see it. But an Agent can't have emotional memory. Can an Agent really predict such human preference? Can the intelligent part not replace human decision-making?
Erik: Yes, the osmanthus tree resonates; for me it wasn't a boy (laughs). Our recent research now reaches 86% accuracy. How?
First, back to the system's underlying logic: humans distinguish your biological senses and multimodal experience. Biological senses are obvious — taste, smell, touch; your ambient experience needs these three to detect. Multimodal is whether the venue is noisy, who you're eating with, what mood. Memory, emotion, surroundings, and inner state are multimodal associations.
Another layer is identity attributes. Liking or disliking something is sometimes hard to say — is it because you like it, or because others like it so you do? If you've been in Shanghai, natural wine is very popular; I've never thought it tasty, but many like it, and if I don't like it I seem unsociable. So we have social attributes.
How is this measured? Through online surveys, implicit-association tests, and behavioral tests. Most food decisions are subconscious; we grab something at the supermarket without thinking much, tied to familiarity.
Here, we recall testers after two weeks for a comparison. Among 200-plus questions in experimental design, maybe 70% train the person's digital twin and 30% validate, then compare against real human results.
As for actual eating and drinking, we're recently connecting with electronic nose and tongue companies. Directionally we can tell beverage companies: this quarter bayberry or lychee, sweeter or tarter. But the exact formula needs food engineers. We're also moving toward brainwave processing, because the prefrontal and temporal lobes most honestly reflect familiarity and liking. We're fusing this data into the system, aiming soon for 90%-plus prediction. Thank you.
Ahaya: Next, Max. Your ThinkSpace is a legal vertical application. Legal has Harvey, very famous, now $1.1 billion valuation, 1,300-plus global clients, just landed in Singapore this June. So is Harvey a direct competitor to ThinkSpace? When such a giant starts local Southeast Asian service and hits your market and customers, which of ThinkSpace's prior advantages are real moats, and which are just a time difference? Coverage of Southeast Asia, Chinese and minor-language advantages, playbook knowledge — how do you respond to Harvey? Which are your real moats? I'm very curious.
Max: This question is core. Every AI startup may need to answer it: when a much bigger, more successful company is in front, how do you differentiate?
I'd split it. First, Harvey appearing in our industry is lucky for us. Legal barely changed for decades, so it first educated the market that "legal AI is valuable and worth doing." It reduced much of our education cost, because legal is already close to service. On this level we thank Harvey.
Next, when it enters our market, do we treat it as an absolute competitor? In function and product many things are similar — they do playbooks, Agents, legal applications. What lawyers do is those things.
But I'd ask a counter-question: as an enterprise, is everyone willing to pay at least $50,000 a year for a full suite? Or does everyone need every feature? Like Adobe has a family bundle (Lightroom, Photoshop), a few hundred a year. But another product does only Lightroom at a tenth of Adobe's price.
For users, we slice user profiles finely. One type may not need many features; on our contract-review product we do only "contract review," extracting past data into structured form to review future contracts — that's enough. We make it fine and precise.
Conversely, Harvey and I face the same bigger question: when the underlying model is strong enough, is vertical AI still needed?
This is interesting. Before today's panel, last night a friend doing R&D at a model vendor messaged me: "Max, I know you do legal; can we cooperate on your vertical data and lawyer-labeled data? We want to train our legal vertical scenario." I felt both angry and amused — how could I give this to you? Isn't this my core know-how and core data?
But sideways, you see big vendors fully entering vertical scenarios, especially legal, maybe others in future. I feel this scenario is a time difference. When the underlying model iterates to a certain level — as that Sequoia piece noted — there's a big gap between the model's underlying capability and enterprises' actual adoption rate.
But when the model is strong enough, can vertical AI be directly replaced? My feeling: Harvey's moat is user stickiness. When I already use your product, putting all my know-how, data, collaboration, and habits on your platform, switching to another product is relatively hard. This stickiness is a big moat.
Second, we found a scenario with customers. When you tell a customer "AI can do 1, 2, 3, 4, 5," in actual use they need education or hand-holding. What this feature can specifically achieve isn't known at the start. So I see a value-add: we truly help users run the whole process and, as CK said, deliver a result. In our view, outcome delivery may not be fully AI-automated; it needs our manual assistance. That's my view.
Ahaya: Enterprise AI Agents are very different from ToC. As I advise some enterprise AI Agents on Go-to-Market, I see vertical-AI founders anxious: if a large model (like Claude) does this too, we're just a feature — how much time do we have?
I think general, low-data-moat vertical SaaS will shrink. Data accumulated in verticals (cleaned, labeled, classified) may soon be caught by big vendors with strong models and capital.
What's truly valuable? My view: in enterprise decision-making, many things aren't direct outcomes but process things. How was the decision made? Which departments? What's at stake? What's the boss's or each leader's decision basis and preference? If you truly plow deep in one vertical, a general model can't go that deep in every vertical. It can gather useful general industry data, but it can't get the truly "human part" of decision bases and not-purely-rational thinking processes, including workflows and collaboration modes.
As Max said, enterprise customer stickiness is higher than individual. An enterprise decision involves budget approval and cross-department agreement; the decision and replacement cycle is long. Building an enterprise Agent has a higher threshold, and compared to personal Agents the moat is a bit higher. If you truly go deep in a field, the bar for founders is far higher than in the last SaaS era.
Will Digital Employees Really Replace Humans?
Ahaya: Next question for CK. In your materials, two lines stuck with me: "Cuber AI doesn't eliminate people; we augment people," and "Cuber AI's digital employee can replace the output of 15 full-time employees." Are these contradictory? Where do the 15 employees whose output you "replaced" go? Are they eliminated after all?
CK: Strictly, if you talk to a boss, the first definition is: "after using your digital employee, how many people can I eliminate?" So we tell them that directly "eliminating people" is unlikely, but you save many work hours.
My output math: doing whole delivery and final results, a person sitting at work can only do 10 cases or transactions a day. But a digital employee runs 7x24 and can do 50 transactions. So while you sleep, it's working.
So I save some work hours, equivalent to raising a person's work ability. Simple reasoning: before computers, we didn't use real Excel; we used handwritten ledgers. When Microsoft launched Excel, did you think "oh no, my math genius is finished, no work"? No. When computers came, you still had work. Today, even with computers so powerful, do you think accountants are jobless? Still here; only the accountant's way of working changed, less drudgery.
Especially in logistics, everyone works super-long hours; on vacation the boss finds them, customers find them. Today, with a digital employee as your assistant, can't it hold the fort first?
So we mean: people and machines will fight side by side — certain. Because digital employees have no emotions. As the lady in the front row said, ToB business needs emotion. You give clients emotional value, bargain with them, so they feel "I'm talking to a living person." So that part still exists.
What do people with saved time do? Let them do higher-value things that bring higher value and better income, like giving clients emotional value. Pull their time out of trivialities. When you're busy, you have no time to socialize; you say "I haven't finished all the system workflows, how can I socialize, how can I chat?"
Yesterday I met our CEO; he told me: "CK, though you're busy, you must spare an hour to talk with people." At first I thought, is talking necessary? Then I realized it is, because you need to know what's happening outside. The digital employee only solves your operational part.
So we're human-machine collaboration: 80% of work by the robot, 20% — like I'm CEO — I just give instructions and tell them what to do. What do I do with the freed time? Build client relationships, do group interviews, or provide emotional value.
After Productivity Is Liberated, What Do Carbon-Based Life Forms Most Want to Keep?
Ahaya: So the final liberated productivity is so people can better be human.
This circles back to my earlier question. I won't ask CK; I'll ask the other three: if one day AI is omnipotent and can do anything for you, what do you still want to do yourself?
Erik: Obviously, eating. Just that. Because we always care about the creativity part and the final experience part — these belong to our carbon-based life, and we must maintain them. Leave tedious work to AI.
Max: I'd probably travel. I love going out. Let your digital employee work for you; you're the CEO, you go play.
Xiong Wei: Similar to Max. AI can't replace taste; you experiencing nature, your heart beating, sweating through sports — only you can deeply experience these.
Ahaya: Thank you to the four guests here today, and thank everyone. Bowing to you. Because of you, this final panel ran smoothly. Again: no matter how AI develops, may every carbon-based life form have a wonderful life!