Understanding Japan’s AI Shift in Finance: Challenges and Priorities (Basic Version)

The application of generative AI to operations has proceeded rapidly even among Japanese financial institutions. According to an AI Discussion Paper released by the Financial Services Agency on March 4, 2025, entitled “Preliminary Discussion Points for Promoting the Sound Utilization of AI in the Financial Sector,” over 70% of the financial institutions surveyed said they permitted the use of generative AI broadly among employees (see Figure 1). The paper also found that over 69% of institutions reported that generative AI is “used more actively now than after its implementation” or “continuously used with no significant changes in its usage since its introduction.”

We believe that behind these developments lie concerns over the DX initiatives of Japanese financial institutions being behind the curve, and shortages of personnel. We can also guess that because generative AI, represented by ChatGPT, is easy to use and can be tried out without any investment, the technology was easy to adopt. The experience of operational streamlining through RPAs that has spread in the past ten years may also have had an effect. Looking at the issue by industry, the rate at which RPAs have spread throughout major Japanese financial institutions is believed to have been higher than in other industries, so there is likely to be a high degree of synergy between such technology and these operations.

Taking all of the above into consideration, while it is clear that generative AI has come on leaps and bounds in what it can do as a solution to business challenges, it is critical for companies to properly build up a store of knowledge around such technologies, as, in a sense, it is an extension of past technologies. To build up such a store of knowledge and make more effective use of AI, we believe companies need to do the following things.

While there are cases where even overseas financial institutions have actively made use of AI, due to the speed with which the technology is progressing and spreading, companies have not really gotten results commensurate with their investments, so it is fair to say there are no clear winners yet. Under such conditions of uncertainty, it is important for institutions to try out actual products and services and to get a feel for them through a process of trial and error. Compared to traditional large-scale package products, projects related to AI are amenable to small starts. Generative AI, in particular, compared to traditional AI, is often something that can be used in a way that is user-led, without putting in place any large-scale framework.
Reports published by public institutions and case studies into companies proactively engaged in DX initiatives also serve as useful reference points (see Figure 2). Case studies from other companies often do not match the conditions of your company or are out of date almost as soon as information about them can be acquired. Thus, rather than simply copying such cases, it is critical for companies to understand the aims, background and key points in those case studies, and to think up their own measures in light of their own management environments and business characteristics.

The main approach in traditional system development is to define and migrate data during system development or during large-scale system upgrades, or, alternatively, to newly store data from scratch. However, during AI-enabled system development, the approach becomes one of leveraging existing data. As a result, it is difficult to make effective use of AI if you do not have the data itself. In preparing data, setting specific goals is not something that can be done overnight, but, as there are cases where Japanese financial institutions have many legacy systems making it difficult to use data with flexibility, it is important for companies to begin by rendering data into a form that is easy to use. There are several methods available for making data easy to use including preparing integrated infrastructure such as data warehouses and data lakes, or making data easy to link through an internal API. When preparing data, as the amount of data held by financial institutions is massive, it is not practical to organize all data using various different apertures. For this reason, companies need to work on preparing data by considering several use cases and defining the data needed to realize those. Companies will also need data governance structures to protect the quality and security of their data.

The Waterfall model of development, in which IT departments design and develop systems based on requirements defined by users, is still widely prevalent in the development of large-scale systems such as core systems. If we put the focus on making use of AI and data, however, defining requirements as in the Waterfall model is impractical due to the need to progress with an awareness of the characteristics and limits of the technology. It is thus important for companies to increase the number of people they have who are skilled in operations, IT and data, and who can thus think of the design and implementation of operations as a unified whole. It has long been pointed out that the knowledge barrier between operations and IT has become lower, and there are often cases of people moving between operations and IT departments. But in making use of AI, personnel should also be skilled in data on top of those domains. Such personnel also need to understand the characteristics of AI, due to the need for humans to be present to improve the accuracy of AI and to ensure the ethical side of its operation, as part of the “Human-in-the-Loop (HITL)” idea in AI development. As developing such personnel takes time, companies will also consider recruiting experienced personnel externally as an option. People skilled in IT and data are, however, in demand across all companies, so recruiting them is no mean feat. Companies thus need to take the medium- to long-term view and work on developing people internally in parallel to external recruitment.
Companies will also need to change the nature of their organizations to make effective use of these personnel. Broadly speaking, there are centralized approaches and autonomous, distributed approaches to this, but hybrid forms may also be possible. Where businesses are small or business models are simple, tacking the issue speedily through a centralized approach is ideal. However, where that is not the case, autonomous and distributed models or hybrid models will be the core approaches. Each approach has its pros and cons, so companies should revise their organizations in forms that are in line with the nature of their businesses and their cultures.