On 30 September I attended Stripe Tour Tokyo 2026. This year marks Stripe's tenth year in Japan (Stripe's press release). The session I joined was the Startup CxO Lunch, which was invite-only. I also went to the main stage sessions, but I will write about those in a separate post.
I have used Stripe since my previous company, where we ran a SaaS business. Kafkai's payments run on Stripe today as well.
This week we released Kafkai Ver.4. With Ver.4, our customers' AI agents connect to Kafkai over MCP and fetch search data themselves. So the idea of AI agents as users, and as buyers, is exactly what we are thinking about right now. I came into this session with high expectations.
The lunch had two parts, followed by networking. I will go through them in order.
(I am writing this from memory and from my notes, so some numbers may be off. Please keep that in mind.)
The Opening Question: What Should Startups Do in an AI-Driven Economy?
Stripe opened with one question. What should a startup do to succeed in an AI-driven economy?
According to Stripe, agentic commerce, where AI agents shop and transact on behalf of people, will be worth 5 trillion US dollars by 2030. That matches the top of the range in a McKinsey report, which estimates 3 to 5 trillion dollars globally. Stripe called 2026 a turning point. AI has started to act on its own: it launches tools and it completes transactions.
Stripe then asked us to take three questions home with us.
- How do you position your company in an AI-driven economy?
- How do you build AI into your product?
- How do you use AI inside your company?
Part 1: Three Trends Stripe Sees in the AI Economy
Part 1 was a discussion between Emily Glassberg Sands, Stripe's Head of Data and AI, and Kevin Miller, Stripe's global head of payments, risk and support. Hayley Hopwood, who leads Stripe's startup business in APAC, moderated.
Ms. Glassberg Sands said Stripe processes about 2% of global GDP. In its 2025 annual letter, published in February 2026, Stripe reported a total payment volume of 1.9 trillion dollars for 2025, or about 1.6% of global GDP. In the AI economy, the share is much larger. According to Stripe's page for AI companies, more than 88% of the Forbes AI 50 use Stripe. From that position, Ms. Glassberg Sands described three trends.
Trend 1: Many More People Are Starting Companies
New sign-ups on Stripe in 2026 are up 50% on the previous year. These are not companies that disappear quickly. Businesses that signed up in 2026 earn about 1.5 times the revenue of the previous year's cohort over the same period. This trend started in 2025. The annual letter says that companies that joined in 2025 grew about 50% faster than those that joined in 2024. You used to need to be a developer to start a company. Now anyone with an idea can build a business.
Trend 2: From Flat-Rate Pricing to Usage-Based and Outcome-Based Pricing
With traditional SaaS, one more unit of usage cost the vendor almost nothing. AI products are different. Every inference has an LLM cost. So companies are moving away from monthly or annual flat rates towards usage-based pricing, where customers pay for what they use. Some businesses are moving to outcome-based pricing, where customers pay for results. Stripe said its acquisition of Metronome, a billing platform, was a response to this shift.
Because tokens have value, fraud has changed shape too. Fraud used to happen at the point of payment. Now it often happens at sign-up or at the start of a free trial. Stripe Radar has evolved from checking payments to checking the customer. And that customer can be a human or an AI agent.
Trend 3: Who Buys Has Changed
For the whole history of the internet, humans bought from humans. Now every combination exists. Agents buy from humans, agents buy from agents, and humans buy from agents. Ms. Glassberg Sands said this means the economic infrastructure needs to be rebuilt.
What "Agent-Ready" Means
Mr. Miller went deeper into how agentic commerce works in practice.
Traditional checkouts assume a human fills in a form. A few weeks ago, Stripe released a checkout that supports an agent protocol. That protocol is the Agentic Commerce Protocol (ACP), which Stripe developed with OpenAI and released in September 2025. When an agent visits a merchant's site, the checkout switches from the form to an API. It is still early, but the first data shows about 40% higher conversion than form filling. Mr. Miller summed up the bigger direction as "move to API-to-API transactions wherever you can".
The other point that stayed with me was a problem that is still unsolved: discovery. How does an agent find out what it can buy? We launched mcpnavi.jp partly so that agents can read its data. It does not solve the problem completely yet, but have a look if you are interested. (If you want to know why we built mcpnavi.jp in the first place, I wrote about it here.)
Transactions through agents are still a very small share of the total. But Mr. Miller said consumer behaviour has visibly started to change in the last four weeks or so, and it is growing week by week. He expects agents to be involved in most transactions within three to four years.
"Kai", Stripe's Internal AI Platform
Ms. Glassberg Sands also introduced "Kai", Stripe's internal knowledge AI platform (LangChain's case study). The name comes from the Japanese word kaizen, continuous improvement. It has three parts.
(As it happens, we also have an internal system called "Kai" at our company. It is short for "Kafkai".)
- The harness. An agent runtime built around how Stripe works.
- Access to tools. Agents can reach internal and external tools such as Salesforce and Jira, with the same permissions as the person using them. They connect through MCP servers.
- Shared skills. "AI architects" in each department build these skills and stand behind their quality. Evaluation and monitoring keep the quality up.
Within a few weeks of launch, almost every employee was using Kai every week. Today 95% of employees use it weekly, and many use it daily. Agents, not data analysts, now run most of the data analysis inside Stripe.
The last question was: "If you started a company today, what would you do?" Both gave similar answers. Pick a painful problem that nobody has solved well. Set up a good harness and improve fast. Use off-the-shelf infrastructure for things like payments, and put a working product in front of customers as early as you can. Talk to customers almost every day and get their feedback. Mr. Miller said, "The cycle is no longer weeks or months. It is days."
Part 2: VC Panel, "The Next Ten Years of AI and Money"
The theme of Part 2 was "The Next Ten Years of AI and Money: Why Founders Should Move Now". The panellists were Akio Tanaka, Founding Partner at Headline Asia, and Tiffany Kayo, Principal at Coral Capital. Motoya Kitamura, who works on startup and investor partnerships at Stripe, moderated.
Headline started in Germany and now invests in Europe, the US, Latin America, Asia and Japan. Its assets under management are about 600 to 700 billion yen. Coral Capital is based in Tokyo and San Francisco and has invested in about 150 companies.
The questions were prepared in advance. At the panellists' request, the session became a free discussion instead.
AI Has Changed the VC Business Too
Ms. Kayo said, "A fund is a business too, so AI has changed it a lot." Sourcing, finding companies to invest in, used to be done by interns and new hires. Now agents do it. For research, they use tools such as Claude. Exit modelling used to be worth the effort only for Series B companies and later. Now they can try it on seed-stage companies.
Mr. Tanaka's Headline is further ahead. It has an in-house software team of about 22 people. For more than ten years it has run a system called "Searchlight". Searchlight watches about 7 million services on the web, around the clock, using about 220 data sources, and finds companies that are growing. That means Headline can approach founders before the founders come to them. In one example, Searchlight found a company with less than 100 million yen in annual revenue. Four years later, that company is at about 12 billion yen.
They use AI in due diligence too. But both panellists agreed that a human still has to make the final call. Mr. Tanaka put it well. We are at the stage where "the human senior teaches the AI junior". Unless you teach the AI how people with good judgement work, it only produces very generic conclusions.
One point came up again and again: proprietary data matters. Ask a general LLM the same question, and everyone gets the same answer. Headline has built its own software with due diligence data on about 10,000 companies, so it can compare companies that are well run with those that are not.
Which Companies AI Will Replace, and Which It Will Not
Mr. Kitamura told a story from an industry conference. The organiser said, with a straight face, "Now that Anthropic is here, all our portfolio companies may disappear." Mr. Kitamura asked the panellists whether they feel that kind of threat.
Ms. Kayo said the concern is real across the board. But Japan is a country where things change very slowly. It will take longer than in the US before the impact becomes obvious. Japan is probably the only country where electronics stores still sell fax machines as a normal product. That does not make Japan safe, though. Thinking "nothing happened for two years, so my industry is safe" is a false sense of security, Ms. Kayo said.
So which companies will survive? According to Ms. Kayo, the ones at risk are "thin" services that only build the surface. That applies to SaaS and to e-commerce alike. If a week of hard work with Claude Code can replace you, you are at risk. Companies with roots below that layer last longer. Examples are companies with licences, their own supply chains or communities: things that are tedious and take effort to build. But she added that even those advantages do not last forever.
Mr. Tanaka named two strengths that existing software companies have. The first is a UI and UX that customers already know. The second is the customer base and its data. These strengths buy time. What decides the outcome is how fast a company can ship an AI-native business within that time.
Mr. Tanaka also talked about a company spun out of one of Headline's portfolio companies. It builds a "company OS" for small teams, where humans and AI agents work in the same org chart. One company tried it. Fifteen years ago, that company was number one in its industry. By now, three or four people were keeping it running. After about a month and a half on the new system, it was back to number one in Japan in its category. Today about five people run a business the same size as competitors with 50 staff.
Solopreneurs and the Point of VCs
The next question was about solopreneurs, people who start a company alone. With AI, people say one or two people can build a company that earns billions of yen a year. If that is true, what is the point of VCs?
Ms. Kayo said there is no doubt that you can now build a company with limited money and people. On the other hand, companies like Base44 and Lovable grew fast with small teams, and they still became organisations of hundreds of people once they reached a certain size. Building a generation-defining company alone is still hard. She expects more of these small companies to end in an acquisition instead.
Mr. Tanaka started with, "This is something a VC should not say." Then he said this. If two or three people build a company with 10 billion yen in annual revenue and strong cash flow, there is no need to force an IPO. "Lifestyle businesses" used to be small. Now it may be possible to build one at unicorn scale. That is a third way to build a company, outside of an IPO or an acquisition. Ms. Kayo agreed: if you can grow by bootstrapping, without outside money, there is no need to raise from VCs.
One founder in the room was building a defence-tech messaging app on their own. A few years ago, people told them defence tech was not investable. They said the mood has changed recently.
We at Kafkai Giken are a small team too, and we have grown without outside funding. It was a pleasant surprise to hear VCs themselves say that VC is not the only path.
Language Is No Longer a Reason to Stay at Home
On differences between countries, Mr. Tanaka made this point. Japanese startups used to follow a "Galápagos" model: build a position in a domestic market that US companies did not bother to enter. But if a consumer service is not protected by a licence, AI now gets over that wall easily. "Language is no longer a reason not to go global," he said. That line stayed with me. If you think the domestic market is enough, you risk being wiped out by foreign players, the same way the iPhone replaced Japan's feature phones.
Ms. Kayo also said the localisation barrier is lower now. Building a Japanese website or Japan-specific features used to be hard work. Now it takes very little time. If you focus on the Japanese market, you need to think about how deeply you can get into your customers' workflows.
Mr. Tanaka also raised Japan's digital deficit. Most companies in the room pay more to AWS, OpenAI, Google and other foreign companies than they earn from abroad. With the yen weak, he called on the room to build more services in Japan that earn foreign currency in the AI era.
When Will Agentic Commerce and Stablecoins Become Everyday?
Ms. Kayo said some founders she knows overseas already let agents buy office supplies for them. They set the guardrails themselves: which products, which sites, and up to what amount the agent can pay without asking. But there are still two hurdles today. One is technical. The other is trust, the feeling that it is safe. The people who shop through agents now tend not to be price-sensitive. They are fine paying more for convenience. She expects both hurdles to come down in one to two years.
On stablecoins, Mr. Tanaka said, "The places where people really use them day to day are not Japan or the US. They are Latin America and Africa." In countries where holding the local currency is itself a risk, it makes more sense to hold dollars from the start. One of Headline's portfolio companies is a stablecoin-based neobank in Latin America. About a million businesses use it as their main bank account.
The Most Common Question at Networking: "What Is MCP?"
After the sessions came networking. I talked with people from a bank, a travel booking app, a ticketing platform, a company researching AI foundation models, and more.
Each time, I introduced Kafkai the same way. Kafkai collects search data on your own site and your competitors' sites, and hands it to the AI agent you already use, over MCP.
The most common question was about MCP. One person from a bank already had it right: "MCP is the standard that lets AI connect to different services, isn't it?" Then they asked, "Does the site I want to analyse need its own MCP server?" The answer is no. Kafkai runs the MCP server. The site you want to look at does not need to prepare anything. If you know the domain, you can look at your own site or a competitor's. I explain this in more detail in the Ver.4 announcement.
Everyone I talked to had the same problem. With enough effort, you can get a fair amount of data from free tools. The problem is that nobody reads it and decides what to do next. So companies end up hiring outside consultants or SEO agencies. We are a small company ourselves, so I understand this well. Kafkai splits the work: Kafkai collects the data, and your AI agent reads it and decides.
Someone from a travel booking app company said, "We have not automated that far yet. I would like to talk about whether we can." For example: every Wednesday, check ranking changes, fix any article that dropped more than a set amount, and republish it to WordPress. You can build a flow like that with the scheduled tasks feature in your agent. But we recommend that a human makes the final call before anything is published.
I also talked with Stripe's account team. At Kafkai, customers buy credits in advance, and each operation uses up credits. We run this on Stripe. The shift Ms. Glassberg Sands described in Part 1, from flat-rate to usage-based pricing, is the direction we have already chosen.
Three Takeaways for Kafkai
Looking at the day from Kafkai's side, I took home three things.
First, give agents an API, not a form. Mr. Miller said conversion goes up when agents pay through an API instead of filling in a form. Kafkai Ver.4 follows the same idea. Instead of a person clicking through a dashboard, an agent calls Kafkai directly over MCP. Building an entry point made for agents works better than making agents operate screens made for humans.
Second, getting agents to find you is still an open problem. Discovery, how an agent finds out what it can buy, is still a big challenge. This is the same problem as LLMO, the work of getting AI to find your products and services. Even when the buyer is an agent, nothing happens until it finds you. I wrote about this on kafkai.ai in "The Age of AI Recommendations: Can Your Company's Product Pages Pass AI Screening?". One of our newer projects, mcpnavi.jp, is also part of this. It is a directory that makes it easy to find MCP servers to add to your AI agent.
Third, the warning about "thin layers" applies to us too. A service that only puts one screen on top of AI can be replaced in a week with Claude Code. We believe Kafkai's strength is not its screens. It is the data, and the way we classify it with our patented 4C framework. Part 2 kept coming back to one point: without proprietary data, AI gives everyone the same answer. That is exactly the bet we made with Ver.4.
Summary
The Stripe speakers and the two VCs said the same thing. The unit of change has moved from months and weeks to days. And the more work we hand to AI, the more valuable proprietary data and human judgement become.
Thank you to the Stripe Japan team for inviting me, to the speakers, and to everyone I talked with at the networking session.
If you want to connect Kafkai to your AI agent, see how to connect to Claude and how to connect to ChatGPT.
