Yunqi Capital Leads Investment in Jekka: Pushing the Next Frontier of AI | Yunqi Doers

Why is customer service both the starting point and the endgame for AI? Jekka AI, a seed-round portfolio company of Yunqi Capital, answers this with human-grade AI service: for AI to truly enter enterprise workflows, it must first achieve stability, controllability, correctability, and SOP execution — win the floor before reaching for the ceiling.

In this edition of Yunqi Doers, we bring you a recent interview with Aaron, co-founder of Jekka AI. The boundaries of AI, how to reduce hallucinations, why reliability matters more than intelligence, why curiosity will be the scarcest productive force in the future... If you're watching where the next wave of AI applications will actually land, read on.

The following is adapted from ZhenFund.

Where Are the Boundaries of AI?

It's not just about AI winning Olympiad gold medals, conducting independent research, or one viral demo after another. Demos only show the ceiling — an idealized state of smart. What truly determines whether AI is "reliable" isn't its occasional ceiling, but its stability and variance control in the worst-case scenarios.

Jekka has been constantly rebuilding its understanding of human boundaries, service, and intelligence. Founder Aaron says: "Large models are the hardest thing I've encountered in over a decade of entrepreneurship. Nothing else comes close."

Jekka's biggest difference from other AI companies lies precisely in its understanding of AI. Ninety-nine percent of this market is foam, and that foam is often built on false premises — like excessive faith in Scaling Laws, or the belief that "intelligence matters more than reliability."

He prefers to forget still-fuzzy concepts like AGI, and instead work bottom-up, understanding AI's boundaries through one test result after another. Ninety percent of Jekka's team is building test sets, test processes, and testing tools. Even an intern's independent project can trigger new nodes that give the model stronger generalist capabilities. This is a story truly driven by curiosity.

The AI era is different now. We need to rethink customer service. The essence of customer service is serving human needs. As long as we're working, we're serving someone — from simple information delivery to becoming part of the product, the process, the outcome. So "customer service is the starting point of AI, and will also be its endgame."

AI itself needs to be defined. In Aaron's view, AI should be reliable, correctable, memorable, and capable of learning — able to do everything your colleague can do.

People are the most expensive, and also the cheapest. A phone call might be worth $30, or it might cost nothing more than a steamed bun for a morning's conversation.

The most precious thing in the future won't be intelligence or labor, but that one attempt pushed forward by curiosity.

---

AI ≠ Human: Humans Are White-Box, AI Is Black-Box

Part.01

Q: Could you briefly introduce yourself and what you're working on?

Aaron: I'm co-founder of Jekka AI. We provide human-grade AI Agents and underlying models for Fortune 500 companies and tens of thousands of SMBs. "Having the capability to provide human-grade AI through our own underlying models" is our biggest difference from other Agent companies — including in communication ability and independent problem-solving.

Q: What led you to found Jekka?

Aaron: There were two important moments.

The first was deciding to work on AI. When I saw the GPT-3 demo, I knew I wanted to do this. I've always felt that entrepreneurship is fundamentally an expression of a person's curiosity: you're curious about the world, discover things no one has done yet, and want to research them yourself and make them happen.

The second key point was around GPT-4's release. The internet was flooded with demos of "what GPT-4 can do," all looking amazing, but when you actually tried them, the results often fell short of expectations. That gap frustrated me, and I couldn't help but want to try it myself. I couldn't help wanting to deeply understand the principles, so naturally I chose to go deep into the underlying layers.

Q: You mentioned Jekka recognized AI's potential early on. Can you share a personally memorable moment of interacting with AI?

Aaron: We were shocked by the original ChatGPT not because it could write poetry or code, but because of its communication ability. For the first time I felt a non-human presence that could communicate with us in a human way. That moment was deeply striking, and we must constantly remind ourselves: when we talk about AI, are we describing a broad concept, or that point that moves people?

Q: During Jekka's development, what "wow moments" have you had?

Aaron: Wow moments actually run through the entire process. AI is fundamentally something new — from generating beautiful images, answering interesting questions, to even solving problems humans themselves can't answer. All of it is surprising.

But personally, my most wow moment was re-understanding the boundaries of humanity and of myself. For example, I once heard composer Chen Qigang and Dou Wentao discussing "can AI replicate people." Chen Qigang said "AI has no soul, it can't replicate people, the human soul cannot be captured." At the time I thought he was being too absolute.

But later, while debugging Jekka, I curiously asked it to explain this view from Chen Qigang's perspective. It answered: "If I were Chen Qigang, I wouldn't think AI could replace humans either." I found this amazing, and followed up: "Why do you think that?"

Jekka told me: "Because AI cannot experience."

That moment made me realize that AI is no longer merely a tool-like, predictable existence. In some sense it can be independently understood, respected, even regarded as an individual worth discussing.

Q: From initial Q&A to today's experience closer to "mutual understanding," how did Jekka build this "most human-like" AI dialogue capability?

Aaron: First, we can't understand AI purely as a tool. To build human-like AI, we must first answer "what is human." Humans aren't fully rational; they're a mix of rational and irrational, with instincts, emotions, and patterns of language and thought.

In designing Jekka, we weren't just building technology and functions — we also had to give it certain "human traits," like emotions, curiosity, modes of expression. It's this understanding that can truly differentiate human-AI communication from ordinary software.

Looking back, if we really want to create a very human-like AI, we need an interdisciplinary team: both solid technical and programming capabilities, and deep exploration of human thinking, language, and the structure of human nature.

Q: How does Jekka specifically define and implement this exploration of thinking, language, and human nature?

Aaron: We have to step back and define AI first. I don't think AI has a unified standard to this day. Many people consider machine learning to be AI, ChatGPT to be AI, even the earlier BERT model was called AI.

Everyone's understanding of AI differs through this process. Ilya Sutskever (former Chief Scientist at OpenAI) said that being able to predict the next token equals being able to predict human thought. That's one definition of AI's essence. But with different definitions come different AI products.

When building Jekka, I tend to define AI as a form of language. This isn't diminishing it — because I think language itself is one of humanity's great inventions in history. If there truly were a cross-species language, like in the Tower of Babel story, even God couldn't tolerate it.

In my view, AI is not only a cross-species mode of expression, but also a language that crosses entities. It can use human language to express the emotions of anything, even describe relationships between two objects. Once you have a definition, you naturally have boundaries. Within these boundaries, you'll find that because AI crosses entities, it can neither be called "human" nor "non-human" — it's more like an existence between human and non-human.

If we want non-human and human to communicate, we must bridge the gap. In closing that gap, we need to understand human patterns of thinking, purposes, emotions in appropriate contexts, then integrate these elements. Of course this is a technical problem, but if we can think at this level, it becomes easier to know what AI should do.

Q: At the model level, how does Jekka solve the hallucination problem?

Aaron: When I first started the company, I gradually understood this problem through hands-on work. Initially seeing GPT-3, I didn't know its boundaries either — I just felt it already seemed to have human capabilities. After starting to work on it, for about a month we wrote our own prompts and did basic coding, and discovered that AI's communication feel was different from humans'. This involved what people call the hallucination problem.

I was thinking then: what is hallucination? Because both then and now, when people mention hallucination, they often conflate it with "error." AI answers a question wrong, and we call it hallucination. But when humans make mistakes in communication, why isn't it called hallucination? Or why don't people suffer as much from it? I think fundamentally, hallucination and error are two different things.

When humans do things, it's a "white-box" process: observe first, then think about the essence of the problem, then reason, then reach a conclusion. Even if the conclusion is wrong, it doesn't matter, because you can ask "why did you think this" or "why was it wrong." After the other person explains, you can explore together, guard against or reflect on principles and facts, making errors easy to correct and communication smooth.

AI's core thinking process is essentially "black-box" — a string of numbers. Prompt is equivalent to AI's eyes or ears; we describe things to it, it goes through a series of numerical calculations to reach a conclusion. If the conclusion is wrong, you can Google or verify, but you can't debug which parameter was wrong, which set of parameters was wrong, like you would with a program. Nor can you predict whether a correction will trigger new errors in other scenarios.

Understanding this, we realized we can't solve the hallucination problem using AI alone. We need a "human-like" thinking model combined with AI to truly solve this. This is the human-like neural network control structure Jekka self-developed along the way — organically combining AI's generalization capability with white-box human thinking processes to serve customers.

For enterprise services, this is critical — you can't tell a company "this thing might do something dumb in the dumbest possible way." You have to strictly follow enterprise SOPs and deliver 100% accuracy. Aside from Jekka, I haven't seen any better practice yet.

Jekka Delivers Keynote at Asian Development Bank Forum

Q: What do you think AI will be capable of in enterprise services and broader fields over the next 18 months that it can't do now?

Aaron: I think we've already achieved a considerable degree of making enterprise AI meet enterprise standards: fully following SOPs, solving over 80% of problems. These are results accumulated from our hundreds of thousands or even millions of daily conversations and phone interactions.

What can be done in the short term — say, next month — I can roughly say. What things will look like in three to five years, I can also judge. But 18 months is a particularly difficult interval to estimate.

However, one thing I can be certain of: I know what AI can never do. For example, Ilya's claim that "predicting the next token equals predicting human thought." My definition of what AI "cannot achieve" is this: AI can never accurately predict the next token.

It can predict, but it cannot predict accurately. Building AI based on this understanding might save us from many detours.

---

From Sierra to Jekka

Part.02

Q: Speaking of customer service, what's your view on Sierra and Bret Taylor's argument that "the best enterprises will benefit most from this AI revolution"?

Aaron: I have great admiration for Bret Taylor. In the OpenAI ecosystem, where information is most abundant, this technically-minded founder chose to build Sierra — something that sounds distinctly unsexy. Yet at a time when AI Agent usage is generally declining, Sierra is one of the few products still growing. He must have seen something.

I strongly agree with his view. Bret's approach to serving large enterprises comes from first principles: the bigger the company, the greater the customer service volume. From a service perspective, limited resources should go to larger clients.

From my perspective, even the best companies' customers are still underserved today. For example, when I pay for a bank's phone customer service, am I buying customer service? No, I'm buying the service itself. An interesting example: one of our unmanned-economy clients needs Jekka to handle the final customer interaction — to converse with users and complete the final product delivery. You can't have an unmanned vehicle for everything and then still have someone sitting in an office driving it, right?

Q: Sierra doesn't serve SMB customers. What's your take?

Aaron: Companies like Sierra can't do the SMB market not because they aren't smart, but because they can't deliver generalizable, reliable AI services. They're more like the Accenture of the AI era. It doesn't sound great, but that's the essence of their business model — an AI consulting firm in a much larger TAM space.

We believe the true value of next-generation AI lies in enabling every enterprise and every individual to receive the service they deserve. Our technology can serve Fortune 500 companies and also small businesses handling just 100 orders a day. Of course, the go-to-market strategy behind that is a separate matter.

Q: What background initially caught our attention and made us feel there was an opportunity for Jekka's enterprise-grade service?

Aaron: We typically understand AI along two dimensions. One is AI's level of intelligence — OpenAI is clearly trying to push AI toward PhD-level capability to solve complex research-grade problems. The other is AI's degree of generalization: how much it can help you with in daily life, how broad its coverage of use cases.

We're also seeing many people trying to push in this direction with AGI or agentic frameworks. I deeply admire these explorations, but honestly, both paths are extremely difficult. They're essentially research projects requiring simultaneous breakthroughs in both science and engineering to truly expand boundaries.

However, people easily overlook a critical point: AI that isn't that "smart" and isn't that good at calling external tools — but is sufficiently reliable — is the origin point of enterprise applications. For example, if an AI can't even accurately answer "how many r's are in strawberry," or is still debating whether 9.1 or 9.8 is larger, yet is called a major breakthrough — that doesn't seem like reliable AI to me. The hardest problems often hide in the simplest forms of expression.

From a technical perspective, if an AI can't even handle customer service problems well, it can't be AGI. Customer service was the first intuitive application people tried when ChatGPT came out. Because while this scenario is simple and generalizable, it extremely tests reliability. If it can't even follow the most basic SOPs, yet claims to reach PhD level, solve ACM problems, and win gold medals — that's somewhat unrealistic. Perhaps we can use Sierra's development to work backwards and estimate how far OpenAI is from AGI.

Q: Looking at it now, what makes AI reliable? Is it the SOP execution you just mentioned, or are there other standards?

Aaron: I think we first need to understand how people work. For enterprises, "reliability" is the most important standard. When companies hire someone, the first requirement is "is this person reliable?" This assessment usually comes from two aspects:

First, they need basic general knowledge and common sense, with certain reasoning ability. They may not be exceptionally strong, but at minimum they can understand what's being said and read situations. Most basically educated people can reach this level. This part is somewhat subjective, requiring us to first establish appropriate datasets to cover common-sense questions.

Second, they need to work under the company's SOPs. Take car sales as an example — selling Li Auto and selling AITO involve completely different SOPs: one emphasizes experience, the other emphasizes mechanical performance. These aren't things you learn beforehand; they're "second training" after joining the company. A reliable person is essentially a combination of "general knowledge + SOP execution."

This is also what we focus on most when defining AI. If an AI can execute SOPs 100%, it's already extremely valuable in customer service, sales, and similar scenarios. SOP execution is quantifiable: not executing is 0, executing is 1. Getting AI to 100% is the direction we're working toward.

Q: In reality, getting a person to execute SOPs isn't easy. When Jekka communicates with customers, how do you ensure it accurately identifies information and correctly executes requests?

Aaron: This tests your understanding and definition of AI. We acknowledged from the start that AI isn't like humans, so we can better see both the advantages and shortcomings of it "not being human."

For AI, its contextual processing ability far exceeds humans', so long SOPs aren't a problem. But issues usually arise at two points:

First, how does it identify and understand contradictions within SOPs? Because no matter how complete, SOPs always have omissions and conflicts. Second, how does it handle conflicts between SOPs and human common sense? This is the harder part.

For example, a fresh graduate might solve some problems SOPs don't cover using common sense. But if AI lacks this ability, it tends to make mistakes. So in training, we focus heavily on how to identify these conflicts and give the model "white-box debugging" capability.

So-called white-box debugging means when AI makes errors, we can quickly trace the root cause and promptly adjust the model's understanding. This process is both about discovering problems and about continuously digging deeper and rapid debugging. This is precisely what enables AI to expand its capabilities quickly.

Q: Jekka's application scenarios aren't limited to customer service — it's more like talking to a person. How do you continuously optimize the experience so customers feel pleasure from the conversation itself?

Aaron: I've always felt the boundary question of AI itself is an extremely complex topic. The "how does AI develop empathy or compassion" I mentioned earlier — it can't truly achieve this. We've also seen many ToC products try to make AI play a role, like a compassionate professional image, then solve problems through setting. But from our practice, relying solely on this approach tends to be one-sided and can't truly solve problems.

So we use two frameworks to address this:

The first leans more toward rational, convergent logic. Conversations ultimately need to solve the user's problem; they must have a purpose, and this purpose should align with the user's interests. In other words, conversations must have convergence, continuously advancing toward goals, making the world more efficient. This is the core function of how we build models.

The second is more like divergent thinking. AI is like a vast network with an enormous number of nodes needing activation, like brain cells. When we want it to be as comprehensive as possible in the convergent process, we need to activate relevant intermediate nodes through various means, letting it associate with a series of concepts, then combine these concepts to form more comprehensive understanding.

For example, when someone casually says "have you eaten today?" humans immediately understand this statement is actually irrelevant. But getting AI to understand "unimportant" isn't easy — it can't stay at the Q&A level but needs to understand the context behind it: why does this statement carry no information in this conversation? What associated concepts need activation for AI to "get" this conversational logic like a person?

Q: We're seeing many general Agents that can also make phone calls and such. Do you think convergent logic better focuses on problems?

Aaron: This is why I said earlier "the demo doesn't matter." People working on AI now include scientists, entrepreneurs, linguists — people from all industries are participating. AI's final form often relates to all professions. The demo you see might be in a phone, in a browser, in all kinds of scenarios executing a task.

But the problem is, demos only show the upper limit, and perfect human-to-human communication naturally tends toward consistency. What truly creates distance is the lower limit. The difference in lower-limit performance between companies is what matters. As a practitioner, I don't overfocus on a demo's best performance, because that's like a lottery ticket: try ten thousand times, you'll eventually get one stunning result.

But the foundation for building truly reliable AI isn't its occasional upper limit — it's its stability and variance control in the worst scenarios. This is what can carry future applications.

Q: How long do you think before Jekka becomes the default choice? Where people think of customer service, think of AI conversations, and think of Jekka?

Aaron: I don't think it will take too long. Both enterprises and individuals typically go through several stages:

First stage: they feel AI is worth exploring, so they start experimenting.

Second stage: they realize AI is complex, and everyone's understanding differs. Some immediately grasp its sophistication and complexity; others think, "Our internal team could build something similar." But that's a common bias.

In our development process, we've deeply understood that going from demo to a truly deployed bot with 100% accuracy — more reliable than humans — demands extremely high standards in team background, technical capability, and engineering control. Even looking at giants like Google and Amazon, seemingly mature foundational modules like ASR and TTS still show huge variance in output stability. Not to mention concurrency, upstream-downstream coordination, network integration, and other complex issues. If even they can't nail it, the odds of other enterprises solving this in-house are even lower. So ultimately, choosing professional third-party products is inevitable.

Q: Will it become the norm for enterprises to use AI agents to represent themselves, like how everyone has a social media profile today?

Aaron: Before that, we need to rethink "service." Today when people say customer service, they might see it as a small, unimportant function — even worry that AI will replace service jobs. But if you shift perspective, service is actually a luxury. Being able to give every customer a VIP experience — that's the true value of service.

So before AI becomes an enterprise "representative," what's more important is making AI truly deliver service, giving more customers a better experience. For example, even Apple, a company that cares deeply about experience — when you call AppleCare, you might pay and still wait twenty or thirty minutes to get through. You'll realize service isn't taken for granted; it's often treated as a "cost center." So the next step is for AI's true value to democratize service, rather than rushing to build so-called enterprise agent personas.

Q: What pricing model is Jekka considering? For example, like Devin, paying per completed task?

Aaron: The biggest difference between AI and SaaS is cost structure. SaaS has near-zero marginal cost, while every AI computation requires tokens, requires compute — so it's more like infrastructure, like telecom networks. It will definitely be a recurring model: use it daily, pay daily. And not just usage fees — it's tied to actual outcomes too. The eventual logic will likely resemble telecom more closely.

Pay-for-results is a fantasy for many investors and AI entrepreneurs. Because outcomes themselves don't have universal measurement standards. If you can't even pay employees by results, how could you pay AI agents by results? So for most businesses, the solution is: use outcomes to drive user willingness to pay, but charge by usage.

Jekka homepage

Q: As time accumulates, what compounding effects do you think Jekka will have?

Aaron: I think the biggest compounding comes from the AI industry itself. Choosing to work in AI means embracing compounding. First, it satisfies my curiosity; second, AI is the major direction of this era.

Along the way, I've had several reflections. First: what will AI's future actually look like? Of course, I have no ability to accurately predict the future, but I suspect that in ten years, the world will become a world of "extremely low friction."

Why do I say this? Because today, anything you do in a company — once the team grows large enough — involves massive communication and coordination. Solving the problem itself is often not the most time-consuming part; the real time sink is communication. And if we understand AI as a "universal language," it will make exchange extremely easy. Once efficiency improves, the way everything gets done will change. The higher the efficiency, the cheaper labor becomes. You could even say AI as compute is faster and more efficient than humans — that's the intuitive understanding.

But if you think deeper, humans consume enormous amounts of time on communication and collaboration. From a purely biological perspective, a single bite of food can stimulate countless neurons — in some sense, possibly more efficient than NVIDIA's GPUs. So the real bottleneck isn't "computation," it's "communication." When AI changes this, the entire logic of social production will be reshaped.

My second reflection: if we treat AI itself as a language, it will precipitate all human expression and knowledge, bringing us into an era of "maximum intellectual concentration." But in such an era, intelligence itself will also become cheap. For example, past ACM problems — five or ten years ago, only the world's top minds could solve them; today, any AI might "hit the lottery" and solve them. In other words, the scarcity of intelligence is rapidly declining.

So my judgment of the future is that AI will give people more opportunities to try new things. Today's social mechanisms reward results: make money, produce outcomes, and the company survives. But in the near future — perhaps within our generation's lifetime — "time and attempts" themselves will gain higher weight in value creation, not just results. This will have major implications for how society distributes resources.

As for exactly where this leads, who knows? That's also what I find both mysterious and exciting.

The Essence of Communication Is Consistent

Q: Using AI agents for service — what do you see as Jekka's biggest advantage?

Aaron: I think the core advantage is that we have a relatively accurate definition of AI, and can derive from that definition downward, ultimately practicing a product that truly meets user needs. This is actually extremely difficult.

Today, whenever you discuss AI, two questions are unavoidable: how is AI controlled? How are prompts written? Writing prompts and controlling AI itself involves research. And in this process, we have some unique secret sauce that lets it execute SOPs exceptionally well. That's our special expertise. Like Coca-Cola's formula — not something we'd detail publicly.

Just on prompt engineering alone: who should write prompts? The model provider? Or should enterprises hire their own writers? Neither makes complete sense. Having OpenAI researchers write prompts for some e-commerce company clearly isn't realistic — they may understand engineering deeply, but their understanding of specific industries certainly isn't enough to solve the problem well.

Conversely, the vast majority of enterprises also don't have anyone with professional training to write out entire upstream-downstream workflows perfectly in prompts. In this situation, to truly deploy AI, you must establish a complete set of processes, combined with professional training for people.

Q: How does Jekka compete with Sierra, the star company in the customer service space?

Aaron: Facing competition requires reframing the question. We're not talking about customer service SaaS — we're serving every customer of every business in the world. If we can examine AI customer service without bias, the TAM in an AI context is enormous; we're all just players in it.

Moreover, someone in the world has to write prompts. If Sierra writes them itself, then as I described earlier, Sierra becomes like a consulting firm — Accenture, essentially. If clients write them themselves, most clients aren't actually professional at this.

We believe there's a third path.

Q: In cases Jekka has observed, what are merchant users' first demands?

Aaron: It's actually quite simple. Merchants want two things: first, get things done; second, make money. Because business is business. Whether completing tasks or generating profit, both fundamentally depend on a complete set of processes. Large companies have more rigorous processes; smaller ones may be looser. But at root, merchants' core demand is: can AI stably, completely run the process through.

Jekka's Korea and Morocco teams collaborate closely to serve local customers

Q: In 3-5 years, how do you envision agents integrating into daily workflows? For example, what would it look like when you open an interaction interface?

Aaron: I think the first value point is making communication between people faster, more accurate, and more timely. Our product solves the old problem of "poor communication between users and companies." This isn't new, but why have so many clients — even Fortune 500 companies — adopted it in just the past six months? Because the results are immediately visible.

For example, one client might have had a 3-4% consultation conversion rate before using us; within a day of implementation, it jumps to 8%, 10%, even 15%. This isn't because AI is so "smart," but because "timely communication" itself creates enormous value. Anyone who shops online has probably experienced this: you message one merchant, no response; you switch to one that replies immediately, and you place the order. Whether they reply promptly determines the outcome.

So if we're talking about what agents will bring in 3-5 years, I think the first value is solving "immediacy." As long as communication between people no longer creates misunderstanding due to time delays, and can reach agreement as quickly as possible, many subsequent things will follow. Including what I mentioned earlier: intelligence becomes cheap, labor becomes cheap, and society's distribution logic may shift in another direction.

Q: Jekka is going global to serve North American users. Have you noticed differences between domestic and overseas users? Or is the essential demand of communication actually consistent?

Aaron: The essential demand of communication is consistent, but expectations differ. Take North America as an example. Many people say overseas clients are more willing to pay, and that's true. The reason lies in their stage of social development — the cost of hiring a service worker might be 2-3x domestic rates, while domestic rates might be 2-3x Southeast Asian rates. So at the level of human "price," different regions are indeed different.

We often come across similar discussions in books: why is the same burger more expensive in North America than in China? It's not because the ingredients or service itself are different, but because higher wage levels in North America make services scarce. For example, when you call somewhere in North America, it might take a long time before someone picks up, whereas in China you can get a response within one to two minutes. Fast response is a critical service metric domestically, but North American consumers simply don't have that expectation.

But from another perspective, this also means that in North America, whoever can deliver better service can create a massive competitive advantage. One of our North American logistics clients became the first company in the region to offer true 24/7 phone support with on-site problem resolution. None of their competitors could match this, so they established a generational lead. This is the fundamental difference in user needs between North America and China.

Q: Beyond 24/7 response, has Jekka achieved improvements in any other specific scenarios?

Aaron: I think 24/7 response itself is the most valuable yet most underestimated AI use case globally. Its performance varies enormously across countries.

For instance, we had a Japanese client. Initially, we assumed their willingness to pay wouldn't be as strong as North American clients, given that labor is more expensive in North America. But this client told me that in Japan, once you hire someone, it often implies long-term or even lifetime obligations. Combined with language constraints, they can only recruit locally.

But even finding part-time college students for customer service is difficult. Outsourcing to call centers in Dalian or the Philippines is still costly — a part-time customer service agent might run 20,000 to 30,000 RMB per month, which is almost unimaginable domestically.

So the key isn't a static outcome, but the stacking of multiple factors: language density, demographic structure, per capita GDP, and so on. All of these determine how AI's value manifests in different regions.

Q: Many people worry that AI will replace jobs, and customer service is frequently cited as the first to go. What do you think?

Aaron: I'm not a fan of fear-mongering. Some people do peddle anxiety, saying you must learn AI or be eliminated.

But my own view is that humans are currently underserved. Customers aren't actually treated like bosses. At this stage, AI will only elevate customer service from basic to more advanced levels, and user expectations will keep rising.

As I mentioned, many Chinese companies offer 24/7 service, while North America largely doesn't. Once you perfect 24/7, users' next demand might be: can I participate in product design? Can my feedback be better respected? Can I receive more valuable, even non-standardized, long-term services?

If customer service can shift from a cost center to a measurable component of unit economics and ROI, the industry impact would be enormous. Take that logistics client I mentioned: 24/7 availability, 80% of follow-up ratings are five stars, and they're handling 5x the call volume at the same cost. If communication is already this seamless, why do we still need to open a mobile app to place orders?

I've run the numbers for many e-commerce companies: they treat customer service as a cost center and hire minimally. But customer service often generates value equal to 10-20% of pure profit. In theory, if they made customer service the fastest and most comprehensive, profits could increase. Many companies just haven't realized this yet.

That's the pragmatic calculation. But the more essential reason is: humans deserve to be served.

Doing What Most Ignites Curiosity

Q: Would you consider Jekka internally an AI-first or AI-driven team? In building Jekka, how does this mindset manifest in communication or workflows?

Aaron: I think AI entrepreneurship has two basic requirements for teams. First, purely top-down output doesn't work. AI fundamentally requires constant exploration of model or use case boundaries through extensive practice. Unlike traditional software development, where you can hire competent engineers and replicate what others have built.

AI is different — it often hits "blind spots" at certain points. You need it to complete a simple task, like counting numbers, and it simply can't do it, no matter how you debug GPT-5. So we strongly encourage everyone on the team to keep experimenting and to document these attempts.

That's why we build extensive infrastructure. This is another major difference between AI entrepreneurship and traditional entrepreneurship. AI is a product of probability and statistics; its debugging is completely different from software programming. Software debugging is deterministic: input A always yields output B, so you modify if it doesn't. But with AI, you need to test at different time points, with different models, even changing a single word in a prompt. Roughly 90% of our team is building test sets, test processes, and test tools.

Jekka Office Photo Wall

Q: Organizationally, is it engineering-oriented or operations-oriented?

Aaron: AI isn't fundamentally a pure science subject — it's very intuitive, very much like the humanities. Our philosophy is: AI's boundaries are discovered through cases and test results. Often, these findings don't come from engineers or data scientists, but from users in practice or operations staff in debugging. So the key is getting users, operations, technology, and algorithms into the same process to accumulate data and complete testing — this is the core of efficient AI application development.

Q: What's your most important criterion when building a team?

Aaron: As I mentioned, the standard for AI is the minimum standard for people: they need common sense. But common sense is actually very scarce, somewhat counterintuitively.

We also particularly value curiosity. Because once all infrastructure, test data, and tools are systematized, what truly determines outcomes is individual curiosity — seeing what results from translating that curiosity into daily work. This confirms what I said earlier: the most precious thing in the future isn't intelligence or labor, but that attempt driven by your curiosity. This is already starting to manifest internally at Jekka.

Q: Can you share an example of this "curiosity-driven" exploratory atmosphere on the team?

Aaron: Over the past two years, Jekka has mainly focused on two things: first, enabling effective problem convergence, which is core to our underlying technology; second, exploring how to activate different nodes to give AI stronger general knowledge and associative capabilities.

The second point simply can't be achieved through top-down approaches. Once, an intern (later a full-time employee), working independently on an assigned topic, used something like a "word cloud activation" method to replace human-readable prompts with human-unreadable word vectors, feeding different inputs into our testing platform. The models performed better than before. No one could explain why it worked, but it played an important role in the product flow. This is a typical curiosity-driven outcome. Of course, this requires good infrastructure.

Q: You've experienced many ups and downs over more than a decade of entrepreneurship. What experiences do you think are shareable or replicable?

Aaron: I think the biggest change after over a decade of entrepreneurship is mindset. At the beginning, I was easily influenced by everything around me, even forcing connections between external events and my own experience, imagining causal relationships where none existed. But now, I tend to first analyze and understand what I want to do more thoroughly, stick to my own judgment and grasp of opportunities, and care more about what I truly think inside.

The world itself is multi-dimensional, like an amateur production, with opportunities everywhere. For people who love entrepreneurship, the core motivation comes when you see a problem that's "just unbearable to watch" and feel compelled to fix it yourself. Society has so many problems, so many dimensions — you can always find an entry point. So ultimately, follow your heart, and work on what you're most passionate about, what you find most interesting, what most ignites your curiosity.

Q: What's your superpower?

Aaron: I've never been quite sure whether I'm a humanities or science person. In high school I competed in Olympiads and won a national first prize; in my final year I switched to humanities; when studying abroad I started with humanities, then switched to science near graduation; later I even did a PhD in Quant. The whole process was jumping back and forth between humanities and science, so I developed a strong ability to empathize and then find internal coherence after empathizing. I find this quite interesting and relatively unique among people around me.

Q: From what you've described, you both ask deep questions and can execute. Were there pivotal moments in your experience that shaped this way of thinking or these standards?

Aaron: Yes, for example when working on AI you inevitably encounter "text-to-image" or "text-to-text" problems. What is beauty? This question is inherently subjective. But because of my science background, I can't help trying to define what "beauty" actually is. For instance, I studied art history as an undergraduate and would wonder: when clothing folds and body proportions are harmonious, is that beauty? Is it beauty when exaggerated? Does AI-generated content fall into the former category, the latter, or fail to achieve "beauty" entirely? This process forces me to deconstruct AI's capabilities and expressive methods. The deconstruction is neither purely scientific nor purely humanistic.

Q: If you could say one thing to other entrepreneurs, what would it be?

Aaron: If there's one thing worth doing in this world, I think AI is the most worth exploring right now. Entrepreneurship has nothing to do with environment — people start businesses even in the hardest environments. It's an inner drive, an impulse. If you have this impulse, do it. A person lives in this world for that moment when mission and impulse come together. Regardless of success or failure, this transcends the meaning of individual existence.

Q: If Jekka could rank first in keyword search, which word would you want it paired with? Why?

Aaron: I want it bound with "human." At the functional and productivity level, Jekka will certainly become world-class — that's our core mission. But I want it more to have humanity, to be integrated with human existence rather than isolated as an object or tool. It should be a bridge for communication, connecting people with people and people with things, truly contributing to human civilization. This is what I hope Jekka can ultimately achieve.