Как ИИ-агенты меняют банковское кредитование: опыт MUFG и Sakana AI

Artificial intelligence in the banking sector has long been a topic of discussion. However, discussing prospects is one thing, and truly integrating an AI agent into real banking processes is quite another. This is precisely what Sakana AI undertook, in collaboration with one of Japan's largest financial holdings – MUFG (Mitsubishi UFJ Financial Group). The companies recently published an interview with the project participants, offering an inside look at how it all unfolded.

Зачем банку ИИ-агент?

Why Does a Bank Need an AI Agent?

Lending isn't just about distributing money. Behind every decision lies a vast amount of analytical work: studying the borrower's financial situation, assessing risks, verifying documents, and ensuring compliance with regulatory requirements. All of this is traditionally performed by people, requiring time, resources, and a high level of concentration.

The central idea behind the project was to delegate a portion of this work to an AI agent. Simply put, it's not merely about using AI as a search or summarization tool, but empowering it with the ability to independently perform a sequence of steps: gathering the necessary information, analyzing it, and forming a structured conclusion.

This approach fundamentally differs from the usual scenario where a user asks a question and receives an answer. The AI agent acts as an active participant in the process, not just an on-call consultant.

Банк это не стартап как ИИ внедряют в критические процессы

A Bank Is Not a Startup. And That Changes Everything

One of the central themes of the interview is the stark difference between how AI operates in a lab setting and the demands of a real banking environment.

In a bank, there are stringent requirements for accuracy. An error in credit risk assessment is not just an inconvenience; it represents a potential financial loss and reputational damage. Therefore, AI is measured by a fundamentally different standard here than, for example, in a product chatbot.

The project participants note that banking processes were historically designed without AI in mind. They are optimized for human interaction, featuring specific document formats, established verification procedures, and internal regulations. Integrating an agent that operates differently into this intricate system is a non-trivial task.

Furthermore, the financial sector is particularly sensitive to the explainability of decisions. If a bank employee makes a decision, they can explain their reasoning. The same level of transparency is expected from AI. This places additional demands on how the agent formulates and presents its conclusions.

Как обучить ИИ работать с финансовыми данными и документами

What It Means to «Teach» an Agent to Work with Financial Data

In the interview, the team explains that one of the key challenges was working with financial documentation. Bank documents are more than just text. They contain tables, specific terminology, abbreviations common to the Japanese financial industry, and formats that can vary significantly depending on the type of borrower.

An AI agent needs not only to «read» a document but also to interpret it correctly within the context of a specific task. This required careful fine-tuning and iterative testing – repeatedly refining how the agent comprehends the task and what steps it takes to solve it.

A separate challenge is ensuring stability. AI models can sometimes behave inconsistently in similar situations. In a banking context, this is unacceptable: the process must be reproducible and predictable.

Сотрудничество между банком и разработчиком ИИ: новый формат

A Collaboration That Became an Experiment in Itself

Interestingly, the collaboration format between Sakana AI and MUFG was unconventional. Usually, in such projects, there is a clear division: one side is the technical implementer, and the other is the client with a set of requirements. Here, things were structured differently.

The teams worked in close contact, collaboratively determining what was possible and what was not. According to the participants, this required the banking side to be willing to adopt new ways of thinking, and the technical side to deeply immerse itself in a subject area it had never encountered before.

This approach – where the client and the developer explore the realm of possibilities together – is typical for cutting-edge applied AI projects. There are no ready-made recipes, and this has to be accepted as a given.

ИИ в банке: осторожный оптимизм вместо громких заявлений

Cautious Optimism Instead of Bold Promises

The way project participants discuss the results is telling. There are no triumphant declarations like «AI now replaces analysts.» Instead, there's a balanced assessment: what has been automated, where the agent performs confidently, and where human intervention is still needed.

This is, perhaps, the most realistic view of where applied AI stands in finance today. Rather than replacing people, it redistributes the workload: routine, well-structured parts of the process go to the agent, while non-standard situations and final responsibility remain with humans.

Moreover, even partial automation in bank lending is already significant. If the agent handles data collection and preliminary analysis, the analyst is freed to focus on tasks that truly require judgment.

Значение кейса внедрения ИИ-агентов для всей отрасли

Why This Matters Beyond a Single Bank

The story of this project is interesting not merely as a case study for MUFG or Sakana AI. It reflects a broader trend: large organizations with established processes are beginning to seriously explore how AI agents can be integrated into their workflows – not as pilots for the sake of exploration, but as a real part of the operational chain.

And here, the experience of the banking sector is illuminating: if an AI agent can operate under strict requirements for accuracy, explainability, and regulatory compliance – it signals the maturity of the technology, not just its capabilities in ideal conditions.

Of course, open questions remain. How can such solutions be scaled? How can their support be ensured when regulations change? How can trust in the agent be built among employees accustomed to working differently? These are not purely technical questions; they are implementation questions, and they typically take longer and are more complex to solve than the development itself.

But the fact that the first step has been taken – and taken thoughtfully, without attempting to pass off the desired as the actual – already speaks volumes.