ChatGPT has been out for more than three years. The models have improved fast, and with them came a common line: building at the application layer keeps getting harder, because whatever you ship today the model may swallow tomorrow.
So I wanted to check in on the founders who jumped into applications right after the first wave. Where are they now?
Mao Hua is one of them. In March 2023, four months after ChatGPT launched, he took the team that had built voice assistants with him at Tencent and started an AI personal assistant. He shut it down because tokens were too expensive, then turned to AI finance for small businesses. That became ccMonet.
My main takeaway from our conversation: the product changed shape several times, but the logic underneath it barely moved.
Redoing what didn't work ten years ago
In 2016 Mao Hua was building Tencent's voice assistant, Xiaowei. It shipped inside QQ Music, Harman Kardon and Bose speakers, Sony headphones, even HP printers and laptops. But it ran on traditional NLP. It could handle simple dialogue and was, in his words, not smart enough.
When ChatGPT came out in November 2022, his first thought was whether the things he had tried and failed to do at Tencent could finally be done well.
That became PlatoX, a consumer assistant with three modules: time management (screenshot a message or say a sentence and it reminds you later), personal finance (snap a receipt and it's logged), and knowledge management (note anything, ask anything).
It was capable. In July 2023 he was in Los Angeles and a friend sent a Google Maps pin for dinner. PlatoX told him it was a Sichuan restaurant and what people typically spent. The next day he asked it to order for five people, under US$100, with two non-spicy dishes because his son had a sore throat, and otherwise the most popular items. The first attempt came back with three dishes for under US$50. Technically correct, he said, since he had said under 100. After he adjusted, it picked five dishes for about US$90. At the restaurant they didn't open the menu. They showed the waiter the chat.
Knowing tokens would get cheaper, but not surviving until they did
The problem was cost. One message cost about 20 US cents, and solving a task took eight, ten, sometimes twenty messages. Ordering that dinner took close to ten rounds. Mao Hua alone burned more than US$1,000 of tokens in a month testing it, and his team asked him to slow down.
A normal user would cost a few hundred dollars a month. ChatGPT charged US$20. No consumer would pay the difference. With under 5,000 users, PlatoX already had to ration usage.
"I knew tokens would get cheaper. I just couldn't survive until they did."
The company needed revenue, and business customers were the realistic path. Time management isn't a priority for companies, but every company needs finance. His own company's books were outsourced too. So the team pulled the finance module out of PlatoX.
On 7 May 2024, right after the May Day holiday, he took his Hangzhou team to Johor Bahru, rented a house, and locked everyone in for a month. That was the first day of code for ccMonet.
Working backwards from the CFO
ccMonet was never meant to be bookkeeping software. The goal was a CFO: give every small business someone who can do the numbers, forecast, make decisions and prepare for fundraising.
A CFO needs complete and accurate data. That is an accountant's job. So they worked backwards and started with the simplest way to get a receipt into the system.
Completeness comes from cross-checking income documents, expense documents and bank statements. A bank transfer with no receipt means a receipt is missing. A receipt with no bank record means the money hasn't moved yet.
It took about fifteen months to get accounting reasonably complete. Only then did the CFO layer appear: has a supplier's price changed, what's the burn rate, how long will the cash last, who owes you money, and how do you chase it.
Deliberately no finance people on the team
The most counterintuitive decision: early on, ccMonet had no finance staff. The product manager had never studied accounting. The engineers had built consumer apps at Tencent.
That was on purpose, because the product is for people who aren't in finance. The receptionist at a dental clinic issuing a bill. The delivery driver handing over an invoice, which the clinic signs, photographs, and is done with.
"I insisted on not having finance people because I wanted the product to be used by people who aren't in finance."
He sees this as the core difference from SAP, Oracle, or even QuickBooks and Xero. Whether those tools target large or small companies, the people using them are trained in accounting. ccMonet's users never were, and don't need to be.
His example is fixed assets. In traditional software, buying a computer means setting it up as a fixed asset and choosing a depreciation method and period yourself. In ccMonet, you photograph a US$2,000 invoice for a laptop. The system sees you're a restaurant or a clinic, not an electronics trader, treats it as a fixed asset, depreciates it over 60 months by default and calculates the residual value each month. "You don't have to touch it."
The team did eventually add finance staff, starting around August 2025. Reports filed with the government need a licensed accountant to review and sign. AI can generate them, but you can't verify them without one.
It's not just OCR
Recognising the text on a receipt is only the starting point. Southeast Asia has no standard invoice format. Receipts are printed, handwritten or both. In Malaysia one receipt may mix Malay, English and Chinese. After recognition comes understanding, and after understanding comes checking, because financial data can't be wrong. "AI still hallucinates."
A year ago a restaurant owner wanted to move his books from a traditional accounting firm, and Mao Hua went with him to collect the files. The first page in a stack of receipts was a handwritten wage claim for a temp worker: nine days at 14 dollars an hour for eight hours a day. He photographed it and the system flagged an error immediately: the handwritten total was more than 40 dollars higher than the actual sum.
A traditional firm would only enter the number in the bottom right corner. The stack held claims for twenty or thirty staff. The owner was surely wondering how many more had been inflated.
A CFO who used one feature
A month into development, ccMonet had exactly one feature: photograph a paper receipt and get structured data back, including date, document number, payment terms, line items, unit prices, tax and discounts.
A friend in Thailand saw it. He was CFO of a multinational group with more than a dozen companies, in the middle of a fundraise, and unable to produce financial statements. The books never reconciled. Amounts were mistyped and US dollars, renminbi and Singapore dollars got mixed up, all by hand.
He reorganised his 20-person finance team around that one feature. The less experienced staff did one thing: photograph container records, cargo lists, customs declarations and invoices into complete PDFs and upload them. Five stronger people reviewed. The CFO signed off.
In three months the books for all the entities were done, EY's audit passed, and the investors' money arrived.
Features can be copied. What can't?
I asked him directly: in the AI era, anyone can point a coding agent at your product and copy most of it. What's the real moat?
He agreed that copying features is easy. Two things aren't.
The first is accumulated data. A customer's history lives on the platform and can be analysed in many ways, which raises the cost of switching.
The second is a word his team uses about service: warmth. Alongside the AI, real people check in and send reminders. "Being served by AI and being served by a person feel different." The service team has two layers: a data team that steps in when the AI flags a problem, and licensed accountants who sign at the end.
"AI can solve 80% now, and later 90%. Who does the last 10%? People."
The small-business market has its own friction. Finance is outsourced, so switching waits for the end of a financial year. Many owners have worked with the same firm for five years. Even if ccMonet is better and half the price, there's a personal cost to leaving, and the old firm doesn't always cooperate with the handover. All of this slows acquisition.
Three years on
A lot changed. A consumer assistant became a business finance tool. Users went from consumers to small businesses. A team with no accountants added licensed ones. Tokens got much cheaper, and things he couldn't afford then are given away free by many products today.
Three things didn't change.
First, the goal: put the most advanced tools in the hands of people without specialist skills. From Xiaowei to PlatoX to a clinic receptionist doing the books, that line never broke.
Second, he always works backwards: define the end state, an AI CFO for every small business, then solve the foundation, whether the data is accurate and complete.
Third, he is clear about what AI can't do. It gets to 80% or 90%. The last part needs someone to sign and take responsibility.
Models will keep getting stronger and features easier to copy. What carried him through three years wasn't a feature. It was the judgments underneath: who the user is, where the end point is, and who answers for the result. Those don't expire when the model upgrades.
*This article draws from a podcast conversation with Mao Hua, founder of ccMonet.