Sakana AI, MUFG Test Loan Expert AI

Sakana AI and Mitsubishi UFJ Bank (MUFG) have moved their collaborative project, the 'AI Lending Expert,' into real-world testing. This initiative aims to streamline the bank's core lending operations using advanced AI agents. The project is led by Masato Ota, an Applied Research Engineer at Sakana AI, and Takato Iida, a Project Manager.

The 'AI Lending Expert' system is designed to support the entire loan workflow, from initial analysis and information gathering to financial simulations and drafting loan proposals. By automating and structuring complex analytical tasks, the AI frees up bank employees to concentrate on client interactions and critical decision points. This move signifies a deeper integration of AI in banking, moving beyond simple data analysis.

Structuring the 'Implicit Knowledge' of Lending

While advanced AI can generate documents, the nuances of financial decision-making require more than just data summarization. Iida explained that critical judgments must be logically defensible, incorporating the often unstated 'implicit knowledge' gained from years of experience and organizational context.

Ota highlighted Sakana AI's approach: "We didn't aim for the AI to produce the final output in one go." Instead, they developed an AI agent workflow automation that meticulously traces human thought processes. This mirrors Sakana AI's research into 'AI Scientists,' where AI iteratively hypothesizes and verifies, building logical frameworks step-by-step.

Initial user feedback from MUFG was direct and critical, pointing out areas where the AI's output was too fact-based or lacked organizational perspective. However, this rigorous feedback, numbering close to 1,500 instances during the evaluation period, proved invaluable for improvement.

To manage this volume of feedback efficiently, Sakana AI employed AI to classify issues, extract improvement points by theme, and continuously refine prompts and rulebooks. They also implemented a system where AI evaluates its own responses, cross-referenced with human feedback, to calibrate accuracy and prioritize areas for enhancement. This rapid, AI-augmented iteration cycle allowed the AI to evolve from a junior-level assistant to a tool appreciated by experienced bankers.

Iida added that MUFG's commitment was crucial, with top management driving the project and a dedicated digital strategy team quickly addressing data, governance, and security requirements.

Sakana AI's Engineering Edge

A key differentiator for Sakana AI was leveraging their 'ALE-Agent' framework, which allows AI to iteratively refine tasks and document its thought process as knowledge. This was adapted to capture and structure the implicit knowledge of experienced bankers into a usable format for the AI. Despite constraints on using the latest high-end models, Sakana AI's engineering prowess in workflow design maximized performance, demonstrating a problem-solving approach beyond mere tool reliance.

The project involved over ten Sakana AI members, including Applied Research Engineers (AREs) and Software Engineers (SWEs), working in close collaboration. A dedicated 'accuracy evaluation specialist' within the ARE team ensured consistent quality. The deep mutual understanding between engineers and business stakeholders was identified as a significant strength.

The partnership model with MUFG was also vital. Unlike standard development contracts, this allowed for a long-term perspective focused on essential value creation, enabling bold course corrections and addressing challenges that extended beyond the immediate project scope. MUFG's active engagement and willingness to pursue ambitious goals were key.

The Future of Work: Human-AI Collaboration

Sakana AI views AI not as a replacement but as a powerful 'buddy' that enhances human capabilities. The focus is on ensuring AI use fosters, rather than hinders, employee growth. Discussions with MUFG centered on defining clear roles for humans and AI.

Iida believes that AI-generated efficiencies should be redirected towards more qualitative aspects of work, such as stakeholder dialogue and deeper understanding. While AI can provide data-driven answers for cost-cutting, the human element—understanding the impact on employee morale, for instance—remains critical for true decision-making. The 'AI Lending Expert' is designed with flexibility to incorporate these human insights.

The vision is a future where delegating routine tasks to AI allows humans to return to their core role: making complex, human-centric decisions. This leads to a virtuous cycle where human growth, fueled by AI assistance, generates better feedback, further improving the AI.

Sakana AI aims to expand the AI's support across broader business processes, creating an AI that learns and evolves organically within the workflow. Their goal transcends mere efficiency, focusing on AI as a catalyst for value creation and organizational growth, positioning themselves as a leader in AI innovation.