The pitch, as YC now frames it, is a sentence long. Software budgets are a small pool. Labor and services budgets are the ocean. So rather than selling a tool to a lawyer, be the law firm — with AI doing the work and a much smaller human layer on top. YC’s own framing goes further: the largest companies of the next decade may not be software companies at all.

It is a good thesis. It is also, at this moment, almost entirely American in its evidence base. Every case study in circulation comes from that same batch.

So I brought together two founders in Singapore who have been operating the model since before it had a name — YG, founder of CapGo AI, which does AI-driven SEO and GEO (getting your product surfaced by ChatGPT and other models), and Mao Hua, founder of ccMonet AI, which does AI-driven finance and accounting for SMEs. Different industries, different customers, no knowledge of each other’s businesses.

They arrived at the same structure for the same reason. And in getting there, they ran into three problems the thesis is quiet about.

Nobody Chose This. Customers Forced It.

Neither founder set out to sell services.

YG built a product. Customers refused it. “They’d say: I don’t want a tool. I want you to just do the thing. And they’d pay more for that — they specifically wanted the service.”

Mao Hua’s version is nearly verbatim, from a different room in a different industry. ccMonet started as an AI finance system for business owners to operate themselves. The first question, every time, from every owner who liked the demo:

“If I buy this, do I have to use it myself? I don’t have time.”

Or the follow-on: “Do I need to hire someone just to run your product?”

This is worth pausing on, because it inverts how the thesis is usually told. YC frames AI-native services as a founder-side strategic insight — a smarter way to capture budget. What these two describe is closer to market gravity. They were pushed. The customer did not want leverage. The customer wanted the finished thing.

The Test for Whether You’re Actually AI-Native

Every company now claims to be an AI company, which means the label carries no information. YG offered the cleanest test I’ve heard:

Remove the AI. Does the workflow still run?

QuickBooks and Xero function fine without AI. A traditional SEO agency delivered recommendations long before models existed. Those businesses added AI. They are not AI-native.

A genuinely AI-native service is one that could not have existed before 2023, because the delivery chain is built on model capability. CapGo doesn’t advise on SEO; it generates the pages end to end. ccMonet doesn’t hand you a bookkeeping tool; it delivers financial statements you can file.

The distinction matters commercially, not philosophically. A company that added AI is competing on a cost curve it doesn’t control. A company built on AI is the cost curve.

What the Service Actually Buys You

Mao Hua described a client in Thailand: ten related entities, a closed funding round of roughly $40 million, and financials that would not reconcile. Transactions in Thai baht, US dollars, and Chinese yuan. A twenty-person finance team that kept getting currency units and figures wrong. The balance sheet wouldn’t balance, the report wouldn’t close, and the money couldn’t land.

ccMonet’s product at the time was early — AI plus OCR, extracting line items and doing arithmetic. Unglamorous. But it removed the human error at the base of the stack. The report closed within months and the funding cleared.

“We maybe solved less than half of what they needed,” he said. “But it was the half everything else was sitting on.”

The second example is smaller and I find it more persuasive. Flipping idly through a client’s invoices, they found a reimbursement overstated by 40 yuan — about five dollars.

A traditional accounting firm would never catch that. Not from carelessness — from granularity. Conventional bookkeeping records a total: a 100-yuan receipt becomes 100 yuan. ccMonet records what the 100 is made of. Unit price, quantity, tax treatment, procurement date, payment terms.

Different resolution, different class of error becomes visible. That is the actual product.

Now the three problems.

Wall One: What Qualifies You?

Neither founder has the credential their category implies. YG is not a fifteen-year SEO veteran. Mao Hua is not a Fortune 500 CFO. I asked what gives them standing to sell into these functions at all.

YG’s answer was that the acquisition curve for domain knowledge has genuinely changed. What used to take five years of apprenticeship can be substantially covered in six months of deliberate reading, listening, and building — not to expert depth, but to a level where you can construct the system and improve it against real feedback.

Mao Hua’s answer was better, because it was a design decision rather than a claim. He deliberately did not hire a single accountant in the company’s first year.

His reasoning: hire a professional accountant and the team quietly builds around them. Product decisions default to the finance person will handle this step. By refusing to hire one, he forced the product to become operable by someone with no accounting background at all.

Not hiring the expert is what produced a product that doesn’t require one.

Wall Two: If Batch Is Free, Is Batch a Moat?

This was the question I most wanted answered. Both businesses run on volume — generating pages at scale, processing receipts at scale. But driving the marginal cost of volume toward zero is precisely what AI does. So when everyone can batch, does batching defend anything? Doesn’t the remaining value sit in exactly the things that can’t be batched?

YG’s answer reframed it usefully. Volume is not the moat. Volume that works is the moat.

“It’s like investing. Anyone can place orders at scale. What’s hard is having a strategy that generates alpha. Our SOPs and accumulated knowhow are what determine whether the pages we generate actually get surfaced and actually convert.”

Mao Hua pushed it further, and more bluntly: software itself is heading toward worthless.

He offered two things he’d watched happen. A singing teacher with 30 million followers, vibe-coding her own class registration app. A second-generation dentist at a family clinic, vibe-coding an operations dashboard. Neither has any engineering background. Neither hired anyone.

“Everyone can build it now. Everyone is building it. There’s no defensibility left in the software layer itself.”

Which is, incidentally, the strongest argument for the AI-native services thesis I’ve heard — stated as an observation about the world rather than a pitch.

Wall Three: Why Wouldn’t I Just Use Claude?

I put to them the two things I actually hear in conversation. That GEO is something people assume they can do themselves. And, from the owner of an accounting firm directly: I’ll just throw the receipts at Claude, it’s cheaper than you.

YG laughed and described something that had just happened to him — he’d asked Codex and OpenAI’s models to review his bank statements. They got the exchange rate wrong and made arithmetic errors. Claude got it right. Then he named the underlying pattern: the Dunning-Kruger effect applied to knowledge.

“Buffett wrote down exactly how he invests. You read it and feel like the second Buffett. Producing the result is where all the knowhow lives. GEO looks like publishing articles. What to publish, where, how to structure it, how to get a model to actually take it as authoritative — none of that is visible from outside.”

Mao Hua answered structurally rather than defensively. Take a restaurant group with forty-plus locations: receipt volume across multiple channels, payment terms, revenue from several platforms. You can ask a chatbot any individual question in that stack. You cannot process the stack.

“Isolated problems can go to a model. Systemic problems need a system.”

He also drew a boundary YG doesn’t have. In accounting, AI handles roughly 80%. The final 20% — review and regulatory sign-off — must be a licensed accountant. That’s a legal and liability constraint, not a capability one. YG’s ceiling is different: CapGo’s stated direction is copilot toward autopilot, with the goal that nearly everything is machine-executed and humans handle only strategy and the conversations AI can’t have.

Two AI-native service businesses. Two completely different maximum automation rates, set by regulation rather than by technology. Anyone modeling this category should be underwriting that variable directly.

The Two Problems Nobody Has Solved

YG named the parts that remain genuinely open, and they’re the parts the thesis skips.

Scope creep. “Seventy to eighty percent of the time in a service business goes to aligning with the customer, not doing the work.” Every hour spent clarifying what was actually wanted is an hour AI didn’t make cheaper.

Retention. SaaS accumulates data and switching costs. Services don’t. “If a client is unhappy today, they can change providers tomorrow. There’s nothing holding them.”

His current answer is to keep platform characteristics — data accumulation, workflow lock-in — while delivering as a service. Both legs, deliberately.

Displacement or Absorption

The closing question: do companies like these displace the Big Four and the large agencies, or do they get acquired and become those firms’ AI divisions?

Mao Hua’s view was unambiguous, and it rests on the last technology transition. The incumbents that bolted the internet onto existing operations largely did not produce a new winner. The companies that rebuilt the category natively — the Meituans and Didis of that cycle — are the ones that took the market.

“Traditional firms adding AI don’t produce a new species.”

I’d put it slightly less certainly than he does. The Big Four have distribution and regulatory standing that the last cycle’s incumbents didn’t. But the structural point holds, and it’s testable within a few years rather than a decade.

What’s already clear is narrower and more useful: the model works, it wasn’t invented in a batch, and the customers pulled it into existence before the founders had a name for it.

This article is adapted from 离线时间 EP20., a conversation with YG (CapGo AI) and Mao Hua (ccMonet AI).