Original · Unique Research · 2026-06-25
Editor's note: The first-person report and its judgments belong to the original Chinese author (hosted by Huangchao Chen; five VC panelists). This English rendition retains the narrative on the first question, data flywheel moat, Day-1 globalization, three survival traits, and closing, plus the full verbatim Q&A. All panelists, companies, and figures are preserved. Investor statements and cited market figures are source attributions, not independently verified findings.
Most AI Agent Companies Won't Live to 2030?
Everyone talks about Agents, every forum shouts "the next wave," and market researchers have already penciled the 2030 market size at 50 billion US dollars. But go through the companies claiming to "do Agents"; most are just a shell wrapped around ChatGPT, a few more prompts filled in, and a decent-looking interface, and they dare to raise money.
At a recent Singapore Agent Summit roundtable, five top VCs — Jeffrey Paine of Golden Gate Ventures, Yiliang Zhao of Openspace Capital, Jim Lim of 59st Ventures, Arun Pai of Monk's Hill Ventures, and Manasi Shah of Accel — sat down for over an hour. Their consensus was blunt and cruel: most AI Agent companies won't live to 2030.
So who will survive?
This article isn't another puff piece about how great Agents are. What we want to tell you is: when the investors who actually hold the money sit across from you, what exactly are they looking at, and what kind of company deserves real money out of their pockets.
01 The First Question: "If Your Agent Disappeared Tomorrow, Who Would Cry?"
"If your AI Agent disappeared tomorrow, who would be affected? Who would notice first?"
These were Jim Lim's exact words at the roundtable. He is Chief Strategy Officer at 59st Ventures and founding CEO of Grab Health, having seen too many beautiful Demos end up as zombie products no one uses.
He says most Agent companies only created "curiosity," not "dependency."
No matter how stunning the Demo or how pretty the PPT, if users just say "wow" and go back to their old tools, you're just a gimmick. If someone panics the next day, a business process breaks, a key metric crashes — then you're mission critical. Jim's criterion is clean: no one notices you're gone, you're just another tool; someone notices, and you're an irreplaceable process.
Jeffrey Paine's question list is harsher still. Golden Gate Ventures invests very early; he fires ten questions at the founder in a volley, peeling open their understanding of the client's pain point layer by layer. "Vertical applications are already very crowded," he says. "You must deeply understand the customer."
"If you send me a BP still using Gmail, I really want to talk to you."
Jeffrey said this with a smile. Not because he doesn't care about professionalism; quite the opposite — he feels someone who can't wait to register a domain and build a website, and reaches him with a personal email, "is full of energy and running hard." "If you've already sorted the domain and website, you may be three months behind." Earlier action is better; full of energy beats fully ready.
After talking to 50+ Agent companies, Yiliang Zhao's ranking is also clear: talk Impact first, then technology. Are you solving a real and big enough problem? If your Agent suddenly became unavailable, would those key business metrics drop sharply? Not vague "optimized efficiency," but tangible, measurable impact.
Accel's Manasi Shah threw the question back at the founders themselves: "How obsessed are you? How fast?"
Her logic is more direct — Claude could launch your competitor tomorrow; how do you plan to fight back? The times change too fast; model capabilities evolve by the day, and all an investor can hold is you as a person and your hunger to win. Talking about Chinese founders, she says, "Chinese teams' speed and explosiveness are excellent — but what you're running isn't just technology, it's that near-paranoid obsession with customers."
Arun Pai is looking for something more mystical: the "Aha moment."
Hundreds of people worldwide are doing something extremely similar to you; anyone can draw themselves in the top-right corner of a two-by-two matrix. The only way a founder can convince an investor is to bring a truly unique insight that gives the investor an "oh, you can think it that way" moment. Everything after that is just execution.
At bottom, the five VCs' questions differ in angle but point to the same core: they don't care how fancy your model is or how flashy your Demo; they care how much your customer hurts.
Finding the pain point is enough?
Far from it. The investor's next question is immediate: what exactly is your moat?
02 Pretty Demos Are Worthless; What's Worth Something Is Called a "Data Flywheel"
Yiliang Zhao has heard this question too many times. The Director of Data Science and AI at Openspace Capital, after 50+ Agent companies, almost every one puffed out its chest: "our moat is proprietary data competitors don't have."
Yiliang's response is direct: the real moat isn't whether you have data, but whether you can use it.
"You should build an iteration loop — feed data in to optimize the model, users using it produce feedback, feedback becomes new data, and the model updates further. That's far more persuasive than simply claiming 'I have proprietary data.'"
He invested in a particularly good example: Thailand's Abacus Digital, spun out of SCB Bank, doing credit scoring specifically for low-income people. These people have no credit-card records; traditional banks won't touch them. Abacus doesn't just have a good dataset; more importantly they have a good strategy to constantly evaluate and upgrade the model, data, and features, making the model more and more accurate. Data turns, the model evolves, and the moat digs deeper.
Data itself is a static photo; the data flywheel is a movie still playing.
Arun Pai was clearly hooked by this, pressing on the spot: "wait, are you sure the team here can build a better fine-tuning iteration loop than Anthropic or OpenAI? What level are their talent density and compute density?"
The room went quiet for a moment.
Yiliang's answer was firm: if you are an expert in the field, sitting inside that industry, and can access business data, you are absolutely stronger than outsiders force-fitting proprietary data.
"I find it very hard to believe an outsider can win just on so-called proprietary data."
The subtext is sharp. OpenAI's engineers, however smart, don't understand the risk logic of a Thai rural credit union; Anthropic's compute, however strong, isn't camped in your customer's office. Understanding the industry, having the data, and iterating — the three together are the real moat. Missing any one, it's made of paper.
Jeffrey Paine drove in from another angle: "The more customers use you, the harder it is to leave you."
He's plain. Make your product a "tool," and customers can swap you anytime; make your product part of their workflow, embedded into daily operations, and leaving you means rebuilding a whole process. The pain of rebuilding the process is the source of pricing power. Jeffrey's own words are blunter — embed your product into their workflow, make them unable to live without you, with no escape.
From "usable" to "can't live without," there's a river called "workflow" between.
Manasi Shah's closing lifted this round to another level.
This Accel investor said something that lingers: "When underlying technology and infrastructure are commoditized, Taste and deep understanding of the top 0.1% of the field become most important."
It means everyone can call the same models, plug the same APIs, infrastructure gets cheaper. At that point, you're not racing who's more technical, but whose "taste" for the thing is better — who better knows what to do and what not to do, who better understands the judgment standard of that top 0.1% of players.
The data flywheel deepens the moat; deep understanding is a wall others can't copy; and taste — taste decides whether you can find what others can't see. Guard all three, and giants can't break in either.
No matter how deep the moat, if it's built on sand it's for nothing. The next question is more fatal: where are you fighting?
03 Jeffrey Begs Founders: "If You Can, Be a Global Company from Day One"
Jeffrey Paine leaned forward as he said this; his tone wasn't advice, it was a plea.
"I'll actually beg you to do it."
Golden Gate Ventures has invested in hundreds of Southeast Asian companies; Jeffrey has seen too many founders get the very first step wrong. His logic is bare: the Singapore market is too small to support the valuation you want. If the end goal is the US, move there directly; don't "transition" in Singapore.
"Singapore is sometimes too comfortable."
The government's mantra is "launch in Singapore first, then go global" — Jeffrey says that's wrong. Go wherever the company should be; staying wherever it's comfortable puts personal preference ahead of the company. He also punctured something many people won't face: team diversity. First ten employees all Chinese, pitch deck handed to American investors — too homogeneous.
But Jeffrey also gave a vivid path: often you don't "charge into" overseas markets; you get "pulled" out. Build a good product in APAC; a multinational's APAC branch uses it, recommends it to Western headquarters, and clients come to you. Dare you drop everything and move? Not everyone has that guts.
Globalization from Day 1 isn't because you're ambitious, but because your competitors are already doing it.
Once geography is right, you must pick the right sector. Jeffrey's judgment is pragmatic: foreign giants will come, but they have their own backyard to defend. Pick manufacturing, semiconductors, supply-chain logistics — areas where giants' capital won't pour in first, giving you a time gap.
"If you pick an industry giants can win quickly, you'll die miserably — their capital is 100x yours."
Choosing the right battlefield is as important as fighting it well.
Jim Lim is most qualified on this topic — he tripped and fell himself.
In 2018, with SoftBank and Ping An money, he built Grab Health, intending to make a splash in Singapore. But Singapore didn't allow AI to directly bill for diagnosis; no matter how much money he raised, he had to go to Indonesia first.
He told another story: EyRIS, a star Singapore AI company, took a full nine years to expand from local to the Middle East. Co-founder Daniel Ting later reflected that starting from the Middle East might have taken only 2–3 years — higher regulatory acceptance, patients traveling farther to see specialists, a bigger pain point.
Jim's advice: use the PESTLE framework (Political, Economic, Social, Technological, Legal, Environmental) to pick the launch market. Not whichever market you're familiar with, but whichever soil your problem roots in most easily and monetizes fastest.
Arun Pai gave a bitter smile here. His story is too real.
When founding Kristal.AI, the team fought on three fronts at once — Singapore, Hong Kong, India — like headless flies. None focused, none best. Finally nearly died, had to return to zero, rethink what the root pain point was. Only years later did they get back on a growth track.
"Failure to focus is the number-one reason people fail. You're not Google; you can't let employees spend Fridays on some moonshot."
The common founder ailment is wanting everything, but the cruelest part of early startup is — with few resources, not choosing means choosing death.
Yiliang Zhao's answer is most counterintuitive.
The host asks: if you started a company from zero, what would you do?
He says: I wouldn't do an Agent company itself.
"I'd do a data company upstream of Agents — Physical AI training data."
AI Agents will penetrate the physical world next, but the bottleneck isn't algorithms; it's the lack of real-operation training data. Hundreds of Agent companies will fight to the death; the people selling shovels always make money.
Strung together, five people's answers draw a map: aim globally from Day 1, pick launch market with PESTLE, dive into a vertical giants despise, all-in on one pain point until it breaks through. If you have the nerve, don't even make the Agent itself — make the thing everyone needs but only you have.
Once you start, every step is a deletion problem. Delete until only one answer remains, and you may live.
Jeffrey says globalization, Yiliang says the data flywheel, Jim says create dependency, Arun says unique insight, Manasi says speed — everyone seems to say something different. But piece these fragments together and they point to one answer.
04 The Five VCs' Unified Answer: What the Surviving Company Looks Like
Five investors, five angles, all pointing to the same answer. The Agent company that survives isn't the best at Demos, but the one customers can least live without.
Specifically, three traits.
Trait one: irreplaceability.
Jim Lim's words are sharpest:
"Create dependency, not curiosity."
Most Agent companies only achieve "making people curious" — beautiful Demo, stunning PPT, customers say "interesting." And then? Then there is no then. Curiosity isn't worth money; dependency is.
Jim told a vet-clinic case. The client originally used traditional clinical-management software, 200 SGD a month. The AI Agent quoted 10,000 SGD a month. Why would they pay 50x? Because this Agent can truly replace a call-center employee. If it can't, why should the client pay?
"If it just enhances efficiency, they won't pay."
"Enhance efficiency" translates to "nice to have."
Trait two: clear ROI.
Manasi Shah's standard is blunt:
"Can 100% replace an employee task and immediately show ROI."
Not "optimized 30% of the workflow," not "improved team collaboration efficiency" — these soft words don't invest. Yiliang Zhao gave a more concrete test: if this Agent suddenly became unavailable, would your key business metrics drop sharply? Yes means you've become infrastructure; no means you're just a plugin.
Real-money replacement, real-money return. Don't fool yourself with soft "efficiency gains," let alone the customer.
Trait three: workflow embedded.
Jeffrey Paine's criterion is simplest: clear problem definition, fully embedded in the workflow.
"If they truly need you, they'll use you heavily, and they'll never get away."
Behind this is a cruel truth — many Agent companies die from "not being used deeply." The client tried it, thought it was OK, but quickly went back to the old way. Why? Because the Agent didn't become "part of the workflow," just an external add-on. Add-ons can be uninstalled anytime; built-ins cost more to swap.
The deeper the use, the harder to leave. That's SaaS's iron rule; Agents are no different.
Three traits stated. But between knowing and doing lie a hundred thousand miles.
At the end of the roundtable, the host asked each guest to leave only three words — for the founders reading this. Here are their "startup first-aid kits"; screenshot them:
Manasi Shah (Accel): Global from day one / Edge and speed / Resilience
Commentary: globalization, differentiation, resilience — Accel watches for long-distance runners.
Arun Pai (Monk's Hill): Unique insight / Dogged determination / Spikiness
Commentary: "Spikiness" is well chosen — don't be everything to everyone; have one sharp edge that pierces the industry.
Jim Lim (59st Ventures): Pilot to production / Dependency not curiosity / Business continuity
Commentary: pilot to production, curiosity to dependency, flash to business continuity — every step a qualitative leap.
Yiliang Zhao (Openspace): Act fast / Data flywheel
Commentary: just two words, no fluff. Act fast; the flywheel spins itself.
Jeffrey Paine (Golden Gate): Why are you doing this / Close the knowledge gap / Sleep under your customers' table
Commentary: the last one is sharpest — "sleep two weeks under your customer's desk" — meaning: how badly do you actually want to solve this pain point?
Five first-aid kits, five ways to survive. But the underlying logic is identical: don't be the coolest-looking; be the most needed.
OK, back to the opening question: what are AI Agent startups really fighting over?
Not whose large model is stronger. Claude can copy your feature tomorrow.
Not whose Demo is flashier. A Demo is the cheapest competitive barrier.
Not who raised more. When the money burns out, the story ends.
What they fight over is who understands the customer's pain best — so badly the customer pays 10x for your Agent, so badly the customer panics when it goes down, so badly they never go back after using it.
Technology ages, models go open-source, capital flows to the next hot spot. But a product truly needed — no one can take that away.
Jeffrey once told founders a rough but vivid line: sleep two weeks under your customer's desk. Not literally on the floor, but asking how badly you want to understand that pain point. Understand it, and the product lives; understand it, and the moat deepens; understand it, and your Agent goes from "nice" to "can't live without."
So the last question is for you:
your Agent, if it disappeared tomorrow, would anyone notice?
More Conversation Detail
Panelists: Jeffrey Paine (Golden Gate Ventures Co-Founder & Managing Partner); Yiliang Zhao / 赵奕靓 (Openspace Capital Director & Head of Data Science and AI); Jim Lim (59st Ventures Chief Strategy Officer & Managing Director); Arun Pai (Monk's Hill Ventures Partner); Manasi Shah (Accel Venture Principal)
Host: Huangchao Chen / 陈凰朝 (China-Singapore AI Association Co-Founder)
Huangchao Chen: I'm today's host. Let me briefly introduce myself. I'm co-founder of the China-Singapore AI Association, Singapore's most influential AI association, with 1500+ members connecting AI entrepreneurs, investors, and researchers across China, Singapore, and globally. I'm also a serial entrepreneur with over a decade of AI experience, CEO of Distri, a company building Agent-as-a-service infrastructure. Today we gather to discuss what investors truly want to see in AI Agent startups. You're all experts in this field; for time, please each introduce yourself in about 30 seconds and your current focus.
Jeffrey Paine: I'm Jeffrey, co-founder of Golden Gate Ventures. We're an early-stage VC based in Singapore, about 80% invested in the region and 20% globally; recently we also set up a new fund in the Middle East (Doha). If you're raising Series A or below, come talk to me.
Yiliang Zhao: I'm Yiliang from Openspace Capital, Director of Data Science and AI. I've been with the company over six years, mainly supporting portfolio companies on AI, ML, and data-science practice, and recently more on investment-side technical due diligence and deal sourcing. I have a technical background, a PhD in machine learning, previously Head of ML Practice at Google, and one of Shopee's earliest data scientists. I've also been adjunct faculty at Singapore Management University (SMU) for over six years. Looking forward to the discussion.
Jim Lim: I'm Jim. I'm Chief Strategy Officer and Managing Director at 59st Ventures. Besides helping startups raise, we also provide consulting and fast implementation deployment training. I was previously founding CEO of Good Doctor Technology (Grab Health in Indonesia), a JV of SoftBank, Ping An Good Doctor, and Grab. Before that I worked in the enterprise world driving tech landing, for example as Huawei APAC Regional CTO.
Arun Pai: I'm Arun, partner on Monk's Hill Ventures' investment team. Monk's Hill is an early-stage regional VC focused on Southeast Asia; we typically write $5M–$20M checks at growth stage (Series B and beyond).
Huangchao Chen: Thank you, distinguished guests. Let's dive into today's topic. Everyone talks AI Agents now, but the reality is many actually use just LLMs like ChatGPT, not true Agents. An AI Agent is truly different — it plans ahead, acts proactively, like your AI employee. Market analysis projects the global Agent market to exceed $50B by 2030. Today we focus on two angles: one, as investors, what are you truly looking for in AI Agent companies? Two, setting aside the investor role, if you founded an AI Agent company yourself, what would you do?
Let's start with part one. When a founder walks into your office saying "we're doing AI Agents," what's the single first question you ask? Jeffrey first.
Jeffrey Paine: For me, the focus is the problem statement. I'll ask about ten questions to dig into how thoroughly they understand the customer's pain. Vertical applications are very crowded now; they must deeply understand the customer. Also, for some reason I personally like one thing — if when you email or BP me you're still using Gmail, I actually want to talk to you; it shows you haven't even bought a domain or incorporated, just used a personal email to reach me.
Huangchao Chen: Sorry to interrupt; I'm curious why?
Jeffrey Paine: Because I do very early-stage investing; I like people full of energy and running hard before everything is ready. If you've already sorted the domain and website, you may be three months behind. If you can't wait to talk to someone like me or a customer, even on Gmail, I think it's cool.
Yiliang Zhao: I've recently talked to about 50+ companies doing AI Agent-related business. From my view, the most important is being clear what problem you solve, how big the market is, and how scalable the solution is. Then we go deeper into technical layers — what's unique about your Agent, what proprietary datasets, how you keep data updated. The system also needs to plan, remember conversation content, self-correct, and use various tools well. Self-correction is a very important trait of AI Agent systems.
Huangchao Chen: To summarize in one short sentence, what's your first question?
Yiliang Zhao: It actually depends on the materials they sent and the problem they solve. But what I ask most is probably Impact — what impact have you brought, and is it a real and big enough problem.
Jim Lim: My answer is similar to theirs; problem statement and impact are definitely the foundation. If I must summarize in one sentence: we should invest in "Dependent," not "Curiosity." We've seen many companies with impressive Demos and beautiful PPTs, but many stay at Demo and idea stage. So my first question is: if your AI Agent disappeared tomorrow, who would be affected? Who would notice first? If no one notices, you're just another tool on the market; if someone notices, you're mission critical, part of an irreplaceable process, creating dependency. That's what makes a successful AI Agent company.
Arun Pai: I look for whether the founder has a truly unique and differentiated insight into the product they're building and why. This space is too crowded; everyone likes drawing themselves in the top-right of a competitive two-by-two matrix. But honestly, hundreds of people worldwide may be doing something extremely similar to you. The only way a founder convinces investors is to propose something truly unique that gives investors an "Aha" moment, feeling the founder has thought extremely deeply. Everything after is just execution.
Huangchao Chen: We have a new guest, Manasi. Welcome Manasi; please introduce yourself and answer: if a founder walks into your office, what's your first question?
Manasi Shah: Hi, I'm Manasi from Accel Venture. Accel is a global fund with $10B+ under management. We're obsessed with AI, fortunate to have invested in Anthropic, Cursor, Perplexity, etc. We're very excited about Chinese founders going global, which is why I'm here. My first question is: how obsessed are you, how fast? The times change too fast; the only thing we can hold is your team and your hunger to win. How obsessed are you with customers? With the product? Did you think about this deeply for a long time, or are you just a speculator wanting to cash in? Claude may launch your competitor tomorrow; how do you plan to fight back? So we value obsession. Second is Speed; you must run faster than everyone, and in the right direction. I think Chinese founders are excellent in speed and explosiveness.
Huangchao Chen: Thanks for sharing. I notice you value founders' passion, focus, market, and team. Next question for Arun. I know Monk's Hill led Saleswhale's Series A back in 2019, a Singapore AI SaaS later acquired by US revenue-tech unicorn 6sense — definitely a very early AI bet at the time. But today's Agents are completely different from that concept. If today they came to you with today's product, would your thinking differ?
Arun Pai: Glad you asked; coincidentally Saleswhale's founder Gabriel recently founded a new company, Bluewhale Energy, which we just invested in. Back then he did an AI sales agent and successfully exited; now he's using AI to enter a foundational area like energy to optimize energy delivery. For Monk's Hill, our philosophy is "entrepreneurs back entrepreneurs." When we find a founder we think is very special, we're happy to back them multiple times, whether or not their previous company succeeded.
Huangchao Chen: So if Saleswhale today came to you with the same team, vision, and product, would you still invest?
Arun Pai: If it were exactly the same as back then, obviously no. It's about timing. If seven years later the founding team has no new unique insight into the problem they're solving, given how much underlying infrastructure tech has evolved, it won't work. You must be able to better use underlying tech to build more scalable solutions. I believe their team learned a lot and has already shifted to a different tech stack. If today they were still doing sales-domain Agents, I can only wish them luck competing with people like Brett Taylor (Sierra co-founder).
Huangchao Chen: Next question for Jeffrey. You've invested in hundreds of Southeast Asian companies, including Carousell, Ninja Van, etc. Today anyone can create an Agent with tools like web. If a Singapore company builds an Agent, but a US or other-country giant also enters Singapore doing the same thing, what advantage do you think Singapore-native companies have? How should they protect themselves?
Jeffrey Paine: The first thing I strongly recommend: if you can, build a global company from day one. If you can move to the US, move there directly. I know about 15% of you are trying this; it won't necessarily work, but you're doing your best. Most of the rest stay local. If you stay in Singapore, the market is too small to support your valuation. If you must stay, I suggest picking a specific industry — manufacturing, semiconductors, supply-chain logistics — areas harder for foreign companies to grab, then focus and go extreme. Also, don't limit your sights to Southeast Asia; look at the Middle East and beyond, where B2B products are rising. Will those foreign giants come? Yes. But they have their own backyard to defend and their own competition; they must win their home market first before coming, which gives you a time gap. If you pick an industry they can win quickly, you'll die miserably, because their capital may be 100x yours and talent better. So you must be very strategic, planning seven years out.
Huangchao Chen: Thanks for the advice. You mention being a global company from day one, but that's very hard, right?
Jeffrey Paine: Yes, I'll actually beg you to do it.
Huangchao Chen: Can you share examples of companies you invested in that were global from day one? What did they do right?
Jeffrey Paine: Some have deep regional backgrounds; even locally, they have experience dealing with global clients. Most of the time, don't insist on signing the first client in the US; launch the product as fast as possible. Somehow overseas clients may start finding you, or a multinational's APAC branch recommends you to their Western HQ, and you find yourself "pulled" out of home base. Then you must decide whether to drop everything and move. Not every founder has that adventurous boldness, but sometimes the market tells you must. So my advice: put the company before your personal preference; go wherever the company should be, not wherever it's comfortable. Singapore is sometimes too comfortable. If the government tells you "please launch in Singapore then go global," that's wrong. If your target is China, go to China; if it's the US, go directly; don't waste time here.
Huangchao Chen: Understood. Should the team also be international?
Jeffrey Paine: Not necessarily. If you're from mainland China, your first ten employees may all be Chinese, your advisors Chinese. But when you're about to go to the US, flip the company to Delaware, and take US money, you'll realize: "oh no, investors see my pitch deck, the team looks too homogeneous; I need diversity." Then you'll scramble; that's a challenge you need to be aware of early.
Huangchao Chen: Thanks. Next for Yiliang. You now work closely with Openspace portfolio companies (like Halodoc, involving millions of patients' medical data) on AI strategy. Many Asian companies now claim unique data. With your technical background, how do you assess whether a company truly has a data advantage?
Yiliang Zhao: Good question. Many companies now run around saying their moat is proprietary data competitors don't have. But from our view, the real moat isn't the data itself, but how you use it to further optimize the model. It should be an iterative process: build a closed loop, feed data in to optimize the model, users using the model produce feedback, feedback becomes part of the data, and further updates the model. If you can prove this use-and-improve iteration runs effectively, that's far more persuasive than simply claiming proprietary data.
Huangchao Chen: With your AI-scientist background, can you share a team example you've seen with technical advantages in models or training methods?
Yiliang Zhao: Honestly, after generative AI took off, we haven't seen particularly promising companies in this area yet. But in the traditional machine-learning era we invested in several, like Thailand's Abacus Digital (spun out of SCB Bank). They built a credit-scoring model giving low-income Thais unsecured loans. They have a very good dataset and good strategies to constantly evaluate and upgrade models, data, and features, making the model more accurate. That's a good example of proprietary data used effectively to improve the model.
Arun Pai: Can I ask a follow-up? I'm a bit surprised you say proprietary data isn't most important. Do you really think teams here can build a better fine-tuning iteration loop than those well-funded foundational models at Anthropic or OpenAI?
Yiliang Zhao: My point is, if you're an expert in the field, sitting inside the industry and able to access business data, you're absolutely more practical than an outsider's model built by force-fitting proprietary data. I find it hard to believe outsiders can win just on claimed proprietary data; I hope that makes sense.
Huangchao Chen: Thanks for sharing. Next for Manasi. I know you focus on early AI; I want to ask, on the customer side, what big opportunities do you see on the 2C side?
Manasi Shah: Your question is about the excitement of 2C vs 2B, right?
Huangchao Chen: Yes, I'd like to hear your sharing, especially 2C opportunities in the AI Agent space.
Manasi Shah: Understood. We're excited about both 2C and 2B. I think 2C suits Chinese founders very well, because the astonishing growth speed and fighting spirit I've seen at Pinduoduo, Meituan, JD, ByteDance can't be found anywhere else in the world, not even the US. So you definitely have the advantage on 2C. The challenge is how you understand global consumers' taste, totally different from Chinese users; you need a data-driven way to tailor overseas product UX and product journey. Second is distribution — it used to be WeChat or SEO; now it may be GEO (recommendation inside Claude or GPT). Getting product and distribution both right is key. On the consumer side, I also see some interesting sectors, like agentic trading, AI short and long-form video entertainment; no one suits these better than Chinese founders.
As for 2B, it's also an exciting opportunity. We value bottoms-up entry products more, like open-source developer tools, Agentic infrastructure, video-editing tools — areas where Chinese founders can win. But the challenge is, if it involves heavier enterprise sales, that needs completely different skills and is very hard in the US market. If your product can run bottoms-up and gain some organic pull, and you can hire one or two people in the US to do sales, it might work, but it's still a challenging area.
Huangchao Chen: Thanks for the 2B and 2C insights. Now the market has hundreds of AI Agent companies, but most won't survive. In one sentence, what traits do you think surviving companies have?
Manasi Shah: The number-one reason a company survives is that its AI Agent can truly 100% replace an employee's work task and immediately show ROI; that's critical. Also, after underlying tech and infrastructure commoditize, Taste and deep understanding of the field's top 0.1% become most important.
Huangchao Chen: Jim, combining your entrepreneur experience?
Jim Lim: As I said earlier, "creating dependency" is your survival guarantee. In B2B, if you can't replace someone or augment them toward higher-value work, you're just a backup. For example, we recently looked at a startup doing a vet-clinic call center; the client previously used traditional clinical-management software for 200 SGD a month, but introducing an AI Agent might cost 10,000 SGD. You must truly replace call-center employees before they'll pay. In B2C, the core is ecosystem readiness. When we took Grab Health to Indonesia it didn't go smoothly, because Indonesia had no natural insurance traffic like Ping An Good Doctor in China. You must first plug into a B2B2C ecosystem; otherwise acquiring C-end customers alone is too expensive.
Yiliang Zhao: I agree strongly. From the Impact metric, if this Agent became unavailable and it causes certain impact-related key metrics to drop sharply, that's a successful Agent example.
Jeffrey Paine: Same for me. The problem definition must be very clear, then fully embedded in their workflow. You also consider profit and numbers. In short, the more customers use you, the harder they leave. If they truly need you, they'll use you heavily, and they'll never get away.
Huangchao Chen: Thanks for sharing; I see your success standards for AI Agents are similar. Now part two. If you were the founder yourself, at Day 0, small team, limited funding, before writing the first line of code, what's the first decision you'd make?
Jim Lim: I'd still split it into B2B and B2C, because these two businesses are completely different. Recently we're also founding an AI Agent company in real estate. As investor and entrepreneur, getting problem statement and impact right is a given. But in B2B you must prove real ROI after replacing or integrating AI into the whole workflow, and you must distinguish yourself from traditional "automation." Also, what happens if your AI Agent goes down tomorrow? In the traditional world this is called disaster recovery or business continuity, something many startups never think about. You must consider all stakeholders' acceptance (like clinicians in healthcare). As for B2C, like my Indonesia experience and the health-tech company I advise, it's ultimately about ecosystem partnerships. Without partners feeding you natural traffic, B2C acquisition cost is very high.
Huangchao Chen: You mentioned founding companies in healthcare etc. Before starting a startup, would you ensure you already have the resources?
Jim Lim: Definitely; resources are prerequisite. It depends on the app type you're launching; sometimes you also need to talk to regulators. For example, in 2018 we launched Grab Health in Indonesia because Singapore didn't allow AI to directly bill for diagnosis; even with SoftBank and Ping An money starting in Singapore, we could do nothing, so we proposed to the board to go to Indonesia first.
Huangchao Chen: OK. Anyone else?
Yiliang Zhao: I might not start an ordinary AI Agent company now.
Huangchao Chen: Just imagine.
Yiliang Zhao: I think a more promising space is an extension of AI Agents, like Physical AI — AI with actuators doing more complex tasks in manufacturing or construction. But the bottleneck is the lack of real-operation training data. So I might found a data company to collect data and sell it to startups wanting to do physical AI Agents.
Huangchao Chen: You mean you'd do an upstream industry for AI Agent companies, right?
Yiliang Zhao: No, I mean I wouldn't do an AI Agent company.
Huangchao Chen: That's cool too.
Jim Lim: Actually it's a good idea. For example, all Singapore public hospitals use an EMR system called Epic, under which runs InterSystems' platform. Owning the data pipeline makes real money; they also let startups build Agents on their data platform.
Huangchao Chen: Very inspiring view. I'm also curious about Jeffrey and Arun: if you started in AI, which direction would you pick?
Jeffrey Paine: I'd be an AI scientist focused on healthcare, specifically curing cancer. With no resource constraint, that's what I'd do; I don't want to watch people die anymore.
Arun Pai: Interestingly, every Monk's Hill investment team member was an entrepreneur before. Combining my past 10 years in banking plus six years at two different fintech startups, I'd probably do something fintech-related. Maybe some AI-empowered service in wealth tech.
Huangchao Chen: Thanks. Another interesting question: everyone knows the success stories, but truly valuable are often the failure lessons. Can you share some failure cases you invested in or saw, and how you think to avoid them?
Arun Pai: I'll start. For a startup, unless you just raised a billion from Accel, you don't have the financial resources to stretch the front too long. You're not Google; you can't let employees spend Fridays on moonshots. You must be extremely focused on the pain point in your mind and execute as fast as possible. Failure to focus is the number-one reason people fail.
Huangchao Chen: Can you give a concrete example?
Arun Pai: Rather than picking on others, let me talk about my own failure. When joining Kristal.AI's founding team, at first we were very lost on ICP and expansion regions. We tried like headless flies to build one in Singapore, one in Hong Kong for high-net-worth, one retail scheme in India. We spread too thin without lots of capital. Finally we had to return to zero and use first principles to rethink the root pain point. Only years later, refocused, did we get back on a growth track.
Jim Lim: Let me share two examples; I won't call them pure "failures." The first is my own Grab Health; starting B2C failed, then shifted to B2B2C and succeeded, because we had no natural traffic in Indonesia. The business model may fail, but if you pivot fast enough, the company overall can still succeed. Another example is EyRIS, a very famous Singapore AI software often held up as a poster success. But co-founder Prof Daniel Ting later shared that they don't feel it was an absolute success either. Because starting from Singapore, it took 9 years to expand to Australia, the Middle East, and Europe. He said if they'd started from the Middle East (higher regulatory acceptance, longer distance to eye clinics, bigger pain), it might have taken only two or three years. So using the PESTLE framework (Political, Economic, Social, Technological, Legal, Environmental) to choose your launch market is very important.
Manasi Shah: I'd add one point. We've seen some excellent founders in Asia build stunning products, but still fail to succeed globally. As Jeffrey emphasized: if you want to win the global market, you must go global from Day 0; there's no other way. Building for local clients is more comfortable, but it makes you late for going global. The best product alone isn't enough to win; the best product plus the best distribution wins.
Huangchao Chen: Yiliang, want to share?
Yiliang Zhao: A brief one. We previously invested in a computer-vision AI company with a strong team, top-school pedigree, and a very solid product. But they chose Southeast Asia as the initial market. When they talked to big companies here, the ROI analysis they produced simply didn't work, because labor here is too cheap; replacing humans with AI was even more expensive than the original.
Jeffrey Paine: Beyond the people factor, I think the failure reason is iteration speed. You must quickly acquire knowledge to make decisions. If you keep building behind closed doors in Singapore, I bet you're already a year behind. You think you're excellent, your first hire is excellent; actually not. Same for us investors; we must keep running to Silicon Valley to see what world-class looks like, to close the knowledge gap. For example Accel has 40 years of history; they've seen what the top looks like. If most people don't know what the world's best product looks like, it's blind leading blind, and in three years everyone dies together. Those Chinese friends of yours working at Silicon Valley giants or joining YC are already a year ahead of you. If you still want to transition in Singapore before going to the US, you may be 18 months behind. Take strategy extremely seriously and close the knowledge gap as fast as possible.
Huangchao Chen: Thanks for the valuable sharing. Since time is short, I'll end with a very practical question to ensure the founders here leave with something directly usable. If a founder walks into your office today, could each guest give three keywords or three core points to help them keep going, raise, and succeed? Starting with Manasi.
Manasi Shah: I might sound like a broken record.
1. Global from day one;
2. Find and play your edge, and keep speed;
3. Resilience — entrepreneurship inherently includes lots of resilience; even if you fail the first time, you must adjust and persist.
Arun Pai: My three:
1. Unique insight;
2. Dogged determination to solve the problem;
3. Spikiness in some area — the trait that makes you top 0.1%.
Jim Lim: Also three:
1. How to go from pilot to production, don't stay in Demo;
2. Create dependency, not curiosity;
3. How to handle it if AI fails, i.e., business continuity.
Yiliang Zhao:
1. Act fast, talk to more potential customers;
2. Focus on building a data flywheel.
Jeffrey Paine: Is this a question to answer or just commentary?
Huangchao Chen: You can say your three key pieces of advice.
Jeffrey Paine:
1. Be clear why you're doing what you're doing;
2. Find the global top 1% people in this field, find ways to know and learn from them, and close the knowledge gap;
3. Have you ever slept under your customer's desk for two straight weeks to understand how they truly work? You must be fanatical enough about the problem you're solving.