Zhou Tianhong, CIO of China Merchants Bank: Three major breakthroughs and four impacts of the big language model

Written by

Iris Vance

Updated on:June-25th-2025

We are currently in the midst of a wave of artificial intelligence, with large language models as the main players.

The large language model has achieved three major breakthroughs in the field of artificial intelligence over the past 70 years:

First, it has achieved very powerful natural language understanding and generation capabilities. When large language models communicate with humans using natural language, it is already difficult to distinguish whether it is a machine or a human.

Second, the big language model has a strong ability to understand and generate unstructured data, including text, images, voice, video, etc., among which text is the strongest. Human civilization has accumulated a lot of knowledge over thousands of years. In the past, information systems were unable to utilize this knowledge. They could use some of it, but the level was not high. However, the big language model has achieved a very high level in understanding and applying knowledge.

Third, great breakthroughs have been achieved in general reasoning capabilities.

For the banking industry, large language models can have four major impacts:

First, the service model has changed. As a service industry, the core concept of banking is customer-centric. Serving customers requires investing in employee resources, but employee resources are limited. Based on the three breakthroughs of the big language model, our information system has become more powerful, so we have the opportunity to truly provide personalized services to all customers.

Second, the change in working mode. In the past, complex working scenarios, especially face-to-face customer service scenarios, were mainly staffed by human employees. With the help of large language models, they can now be gradually transformed into "human + digital" scenarios, where human employees and intelligent employees work together to serve customers.

Third, the change of interaction mode. Especially for retail business, most of the retail business of China Merchants Bank is conducted on mobile APP. In the past, the graphical interactive interface was used to interact with customers. The graphical interactive interface is very intuitive, but it also has shortcomings. The way to organize the business functions of the graphical interactive interface is the menu, but the menu structure is very complex. It constitutes a very wide and deep tree. In this tree, except for a few high-frequency functions that customers often use, they know how to find the corresponding menu path, but they don’t know most of the function paths. When they need to use them, they will have problems and difficulties. How to solve it? The solution is search. When using "search", users have to input and jump several times before they can find the place they need. Overall, the user experience is not very good. With the help of the large language model, we can shift from a pure graphical interface to an interactive mode, and the natural interactive mode may be accepted by more and more customers.

Fourth, the change in data analysis. For banks, we all emphasize the use of data. A high-level bank must have a high proportion of employees working with the help of data. How do business personnel use data? They use data through reports. In the traditional model, when business personnel have new report requirements, they need to raise requirements, such as a new report or a modification to the original report. Raising requirements is a research and development process. The requirements are raised to the technology unit, and the technology unit arranges people to develop. In this process, there may be some back and forth, such as the requirements are not clearly raised, or the developers understand them incorrectly. Development, production, and delivery are a research and development process. The big model has a strong natural language understanding ability and the ability to learn knowledge. In this way, the big model can learn what data, fields, and tables a unit has, and it will know how to organize the data of this unit in the table and where the data is stored. Based on the ability to understand natural language, the big model can clarify the user's problem, know what kind of data is now, write a program, and directly complete the report development work in tens of seconds. Therefore, the original work process of several days or weeks has become a quasi-real-time process.