🏦 Every time a Japanese company has signed an AI deal with Anthropic this year, it's been an IT contractor lining up to resell Claude to its clients. SBI just broke that pattern. On June 2, 2026, the financial conglomerate said it would put Claude across its entire group — banking, securities, insurance, asset management, crypto — and bolt it onto the on-chain finance machine it has spent years building. It's the first time a Japanese financial group, rather than a systems integrator, has gone all-in on Anthropic. And it points at something bigger than a software contract.

A user, not a vendor

Here's the run of recent deals. NEC became Anthropic's first Japanese "global partner" in April, putting Claude in front of about 30,000 staff. Fujitsu signed with both Anthropic and OpenAI in May, across roughly 100,000 employees. Both are integrators — companies whose business is building someone else's AI into systems they then sell to banks, factories and ministries.

SBI sits at the opposite end of that transaction. It is the customer. With around 80 million customers and roughly ¥11 trillion (about $69 billion) in assets under management, it is one of Japan's largest financial groups, and it plans to use Claude on itself.

The arrangement is a three-way one. SBI Holdings and Anthropic set the direction; Ridge-i — a Tokyo-listed AI firm that SBI counts as an equity-method affiliate — does the building, standing up a cross-group engineering team to develop and run the systems. In return, Anthropic gives SBI and Ridge-i priority access to its newest models and security features, an early look at its product roadmap, and hands-on engineer training.

What does "group-wide" actually cover? Four things, roughly. Claude goes to staff across every business line for everyday work — coding, customer inquiries, data analysis, demand forecasting, corporate sales. SBI's securities arm becomes the first Japanese financial institution to trial Anthropic's "Claude Security" tooling. SBI feeds its Japanese-market data into a joint experiment with an Anthropic financial-knowledge AI agent that's still in development. And — the part that matters most — Claude gets embedded into the customer-facing platform itself.

Where AI meets on-chain finance

That last piece is what separates SBI from the integrators, and from most banks anywhere.

SBI has spent the past few years laying rails for what it calls on-chain finance: moving money, securities and other assets onto blockchains. It is launching JPYSC, a yen-denominated stablecoin issued through a trust bank — a structure Japanese law permits, and one that sidesteps the transfer caps placed on other stablecoin types. With its partner Startale it is building Strium, a blockchain designed from the ground up for trading tokenized stocks and real-world assets around the clock. In May it opened talks to take over the crypto exchange bitbank.

Now set a capable AI agent on top of all that. SBI's stated goal is a "global financial concierge" — a personalized agent that reads a customer's situation and offers asset-management advice, analyzes a household budget, or proposes insurance, drawing on real Japanese market data and, in time, acting across tokenized rails. SBI, and chairman Yoshitaka Kitao in particular, has framed finance as moving from human decisions toward AI agents making the calls. This deal is the attempt to actually build that agent.

It's a different proposition from a bank handing Claude to its analysts to speed up research. SBI is wiring AI into the product customers touch — and into financial plumbing that runs 24/7 with no human in the loop. Which is also exactly where the risk lives.

What the US and Europe are doing

SBI is buying into a fast-moving wave, not starting one. Anthropic has spent 2026 pushing hard into finance.

In early May it held an invite-only financial-services briefing in New York — its CEO Dario Amodei appeared alongside JPMorgan's Jamie Dimon — and rolled out a model tuned for financial work plus ten ready-made agents for grind tasks like KYC checks and month-end close. It tied up with FIS, the back-end plumbing behind a large share of global banking transactions, to build a financial-crimes agent. Goldman Sachs reportedly embedded Anthropic engineers for six months to automate accounting and compliance. Bridgewater, Citi and Australia's Commonwealth Bank are all in the mix. By one January survey of large enterprises, Anthropic had edged past OpenAI to roughly a third of the corporate AI market.

But the American pattern is mostly individual institutions buying AI for back-office and analyst work. Europe's is shaped by its rulebook. Under the EU's AI Act, using AI to score someone's creditworthiness or price their insurance is classed as "high-risk," and from August 2026 those systems face hard obligations: documented risk management, human oversight, bias-tested training data, audit trails — and crucially, the institution stays responsible even when the model is bought from a vendor. Fraud and anti-money-laundering tools mostly escape that top tier, but the net effect is a slower, more paperwork-heavy road to deployment.

SBI's group-wide, customer-facing, on-chain ambition is more aggressive than the typical US analyst rollout, and it would run straight into Europe's high-risk regime. So what does Japan's regulator make of it?

The regulator's quiet green light

Japan's Financial Services Agency hasn't written AI rules. It has written a conversation.

Its "AI Discussion Paper," first published in March 2025 and updated to version 1.1 in March 2026, is explicitly not a binding regulation — it calls itself an "initial mapping of issues," built on a survey of 130 firms and continuing talks with the industry. The FSA's stance is technology-neutral: existing financial law applies whether or not you use AI, and the agency says it will adjust its guidance as the technology moves. Notably, it frames standing still as its own hazard, warning firms about the "risk of not challenging" — of falling behind by being too cautious.

Two details make SBI's timing look deliberate. The 1.1 update specifically noted that the industry is shifting from using generative AI for internal work toward embedding it in customer-facing services — which is precisely SBI's move. And among the risks the FSA flagged was over-dependence on a small number of third-party providers. That is the Claude-dependency question, restated in regulatory language. SBI's partial answer is to have its securities arm be the first to stress-test Claude Security, and to keep the engineering in-house through Ridge-i — building the compliance muscle alongside the capability.

What could go wrong

In finance, the cost of getting it wrong runs higher than in almost any other sector.

A chatbot that invents a plausible-sounding fact is an annoyance; an AI agent that misreads a portfolio or mishandles a regulated disclosure is a liability. Hallucination and explainability — being able to show why the AI did what it did — are the exact problems the FSA keeps circling, and they're harder in finance than almost anywhere. There's the cost question at group scale, the dependency question already noted, and the plain execution question: SBI has to make this work across banking, securities, insurance and crypto, each with its own rules. Anthropic itself ships its finance agents as draft-generators that require a licensed human to sign off — a reminder that even the vendor isn't claiming the human can leave the loop yet.

None of that makes the bet wrong. It makes it a bet. SBI is wagering that being early — first Japanese financial group, first to test Claude Security, first to fuse AI with its own on-chain rails — buys a lead that's hard to copy later.

So how does it work where you are?

In Japan, the regulator is nudging finance to adopt AI faster, not slower, while quietly worrying about who everyone ends up depending on. In the US, the big banks are racing ahead under light-touch oversight. In Europe, the rulebook comes first.

Where you live, when your bank or insurer reaches for AI, who's really in charge — the institution, the regulator, or the overseas company that makes the model? And would you let a financial "concierge" you never hired manage advice about your own money?

References