MUFG and Sakana AI's "AI Lending Expert" Enters Verification Phase with Real Cases
Sakana AI ● Covered by 2 sources
Sakana AI and Japan's MUFG Bank just wrapped a six-month test of an AI agent that drafts loan approval documents. Bank staff still check the AI's work, but it's already handling real cases at select branches.
Based on reporting by Sakana AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Loan approval memos are the unglamorous backbone of Japanese corporate banking. Every yen a bank lends has to survive a gauntlet of internal review, and that process — called ringi — has stayed stubbornly manual for decades. Sakana AI, the Tokyo-based lab founded by former Google Brain researchers, spent about six months with Mitsubishi UFJ Bank trying to change that.
The two companies built an AI agent system that does more than spit out a first draft. It pulls together initial case analysis, organizes supporting information, runs financial simulations, and only then produces a ringi document for a human loan officer to review and edit. Sakana AI leaned on techniques from its own research projects, including The AI Scientist and ALE-Agent, to get the system to reason through a process that's historically depended on institutional memory rather than written rules.
About 100 people worked on this, split between MUFG relationship managers, credit reviewers, digital and planning staff, and Sakana AI's own engineers and project managers. Roughly 30 of them formed the core team that actually built and iterated on the system. That's a serious headcount for a proof of concept, and it signals how much both sides are betting on this actually working rather than just producing a flashy demo.
The pilot focused on domestic corporate lending, and the companies say it held up across the bank's main loan categories, speeding up what is normally a slow, multi-step approval chain. MUFG now plans to roll it out gradually to real cases in select branches and departments before expanding further. One of the more interesting admissions here is that the project reinforced just how much of a veteran loan officer's judgment is tacit knowledge — the kind that's hard to write down — and that figuring out how to encode that into an AI system turned out to be as much the point of the exercise as the automation itself.
This is one piece of a broader MUFG-Sakana AI partnership announced in 2025, and both companies say more banking functions beyond lending are already being explored for AI treatment.
My take — AI-written commentary, not fact-checked reporting
A hundred people and six months to get a loan-memo assistant into limited production tells you everything about how slowly AI actually moves inside regulated finance, no matter how good the demos look. I'd bet the real story here isn't the model — it's the unglamorous work of extracting decades of loan officers' gut instinct into something a system can act on, and that's a much harder, much more valuable problem than another chatbot wrapper.
Read more about this at: Sakana AI
