Practice and thinking on the application of large models in China Guangfa Bank

At present, artificial intelligence big model technologies represented by ChatGPT and DeepSeek have made breakthrough progress and are deeply integrated into the core areas of commercial bank innovation and development, helping commercial banks to continuously build a smart financial ecosystem covering multiple scenarios such as intelligent risk control, precision marketing, and intelligent customer service. Against this background, China Guangfa Bank adheres to the spirit of the Central Financial Work Conference, actively promotes the construction of digital finance, and promotes the continuous improvement of basic capabilities such as technology research and development, operation and maintenance services, data governance, and intelligent applications. In 2025, China Guangfa Bank will focus on promoting the "Starry Big Model Project" throughout the bank, building enterprise digital intelligence capabilities, promoting "artificial intelligence +" banking application innovation, and creating intelligent and inclusive financial services.

Generative AI breaks through the limitations of traditional technologies through the progressive mechanism of "pre-training-fine-tuning-reinforcement learning (RL)", and emerges with "intelligence" that surpasses human knowledge reserves and learning speed. In 2025, the large reasoning models represented by DeepSeek R1, with technical innovations such as reinforcement learning and mixed expert models (MOE), have significantly improved complex logical reasoning and mathematical analysis capabilities. Its low-cost training reasoning and open source technology characteristics have laid the foundation for universal application.

Based on the application of general large models, combined with technologies such as knowledge retrieval enhancement (RAG), supervised fine-tuning (SFT) and reinforcement learning, large models can continuously learn the private domain knowledge of enterprises, thereby improving the complex decision-making effects in business scenarios. Enterprises need to focus on data reserves and knowledge engineering construction, master training resources and have algorithm capabilities, form a virtuous cycle of "technology research-industry application-optimization and improvement", and promote the continuous improvement of organizational capabilities.

Agents are rapidly evolving in terms of autonomous task decomposition, tool calling, and closed-loop processing capabilities, making enterprise data no longer a "digital fossil". Agent products represented by Manus and DeepResearch continue to emerge, and by integrating multiple tools and operating capabilities, they have achieved end-to-end comprehensive task processing. The deepening application of multi-agent systems and reinforcement learning technologies has further improved the application maturity and stability of the "super AI portal". The model of "humans define problems - AI solves problems - humans evaluate results" will enhance the job capabilities of "digital assistants" and "digital employees", thereby continuously improving the management and operation efficiency of enterprises. In terms of external services, the human-computer interaction method is changing from "graphical user interface (GUI)" to "natural language as interface (LUI)", which will lower the threshold for the use of financial services, provide customers with "AI financial steward" services, promote paradigm changes in financial services, and reshape the industry's ecological landscape.

The global manufacturing landscape and supply chain are facing profound adjustments. The wave of digital economy is accelerating the reshaping of business models. The macro environment in which the banking industry is located is undergoing tremendous changes, which makes the banking industry urgently need to transform to more refined and agile management capabilities. With the continuous improvement of model level, the continuous enhancement of computing power, the significant reduction of inference costs and the rapid innovation of application scenarios, the banking industry has become one of the industries with the highest proportion of large model technology implementation. Artificial intelligence technology has therefore become an important driving force for the development of digital finance. Its nonlinear enhancement effect has transformed it from a simple efficiency tool to a core strategic asset of banks, driving the digital transformation and upgrading of banks.

Faced with the wave of industry changes triggered by big model technologies represented by DeepSeek, China Guangfa Bank has established the construction direction of "All in AI" and comprehensively promoted the construction of the "Fanxing Big Model Project". Through the four major initiatives of "awareness enhancement, capability enhancement, practice deepening, and security assurance", it has promoted the in-depth application of big model technologies throughout the bank from four aspects: comprehensive application, foundation enhancement, knowledge engineering, and talent training. It has improved quality and efficiency internally, promoted experience and strengthened services externally, enhanced business innovation capabilities, seized new opportunities for the development of artificial intelligence technology, and promoted digital transformation to a higher level.

China Guangfa Bank will introduce open source big models in 2023, launch intelligent question-answering and knowledge retrieval tools, complete big model architecture planning and expand application scenarios in 2024, and complete the internal deployment and online use of DeepSeek-R1 big model in February 2025. At present, China Guangfa Bank has implemented more than 50 big model business scenarios, covering multiple business sectors such as retail and corporate, focusing on the three core directions of office assistant, job assistant, and capacity improvement, enabling business development from the four dimensions of marketing, service, risk control and operation.

(1) Retail marketing is more precise

The traditional retail customer marketing model is mainly activity-driven, that is, marketing activities are designed first, and then the target customer group is screened according to customer tags. However, there are problems such as scattered customer tags and difficulty in comprehensively identifying customer needs and preferences. The big model can be combined with customer life cycle management to make customer marketing management more accurate and effective. In terms of customer portrait construction, China Guangfa Bank uses the big model to process nearly 10,000 customer communication records per day, extracts information on six major categories of product opportunities from them, generates customer intention tags, and helps account managers to more comprehensively grasp customer needs; in terms of behavior prediction, through the customer behavior sequence big model, analyzes customer behavior and attribute feature data, and predicts the probability of key customer behaviors in asset improvement, loan delinquency, etc. Compared with traditional machine learning models, after adopting the big model, the marketing success rate has been greatly improved, and customer behavior prediction and customer group screening are more accurate; in terms of personalized strategy, based on the tag information of key customers, the big model is used to analyze and generate customer core features and customize personalized marketing plans for them, truly realizing "thousands of faces for thousands of people" precision marketing.

(2) Customer service is more efficient and warmer

The intelligent agent assistance system of China Guangfa Bank has covered 6,500 manual seats, providing functions such as customer emotion recognition, real-time quality inspection, speech recommendation, call summary, automatic query and form filling, which has improved the efficiency of telephone salesmen by 7.9% and effectively shortened the length of customer service calls. In the credit card interest and fee account inquiry scenario, facing thousands of transaction details, it takes a long time for customer service staff to handle and answer questions. With the help of intelligent agent task planning and interface calling capabilities, the system can automatically complete interest and fee calculations and generate query results, greatly shortening customer waiting time. In the online customer service robot scenario, for complex questions such as complex long sentences and context references from customers, with the help of large model semantic understanding capabilities, the accuracy of intention recognition has been increased to more than 95%. In order to reach a large number of long-tail customers more efficiently and improve financial management professional capabilities, China Guangfa Bank will continue to optimize intelligent custody and tool assistance functions, strengthen product promotion, repurchase upon expiration, investment consulting and other services, enhance customer trust, and provide a more convenient and warmer customer service experience.

(3) More efficient credit management

In the corporate credit business, China Guangfa Bank needs to regularly conduct credit investigations and reviews for thousands of customers. In the traditional model, front-line account managers and credit reviewers need to spend several working days to complete information collection and report writing. Through the automation of information collection and the embedding of large models to generate core content such as industry analysis, China Guangfa Bank has significantly improved the efficiency of report writing. In the future, China Guangfa Bank will continue to expand external data through Internet search, support richer document generation functions with more powerful data analysis capabilities, further reduce the burden on front-line personnel, and help business expansion.

(4) Smarter back-office work

China Guangfa Bank has built an AI center portal, integrating DeepSeek reasoning large models, multimodal large models and other capabilities, integrating AI tools and interactive interfaces, providing knowledge questions and answers, Internet search, copywriting, office PPT generation and other functions, and developing general products such as report writing, data analysis, and voucher document recognition to comprehensively improve office efficiency. The office assistant will continue to expand transactional functions such as document distribution, supervision, meeting and schedule management, while enhancing the content generation capabilities such as meeting manuscripts, report writing and publicity design.

The implementation of large-scale intelligent applications is inseparable from solid basic technical support and application R&D capabilities. China Guangfa Bank continues to strengthen the construction of large-scale model basic capabilities, formulates a "one core and multiple" architecture plan, conducts in-depth research on technologies such as model fine-tuning training and intelligent agents, and independently develops a data science platform to realize intelligent computing power management, providing platform-level support for the development and application management of large models. In the future, China Guangfa Bank will further build an efficient and empowering large-scale model basic platform and tool system, accumulate and control core technologies such as training and fine-tuning algorithms, and consolidate the technical foundation.

(1) Fine-tune training and build enterprise models, and combine multiple methods to reduce model hallucinations

The illusion of big models will cause the reliability and stability of generated content to be insufficient, which is the key bottleneck restricting its application and the biggest concern of financial institutions in expanding the scope of business applications and directly serving customers. For example, when generating a credit risk assessment report, if the big model changes the data details or the analysis is wrong, it may lead to the conclusion deviating from the actual situation and causing losses; in the process of serving customers, if the big model generates errors or misjudges key service information, it may cause the risk of customer complaints.

To address the hallucination problem of large models, China Guangfa Bank has been continuously exploring the direction of enterprise-level large models, starting from complex scenarios such as software engineering, data governance, and intelligent agent intent recognition, and has carried out a series of practices: First, it has built a tool chain for fine-tuning the collection, annotation, and cleaning of training data to form a data engineering management method that is convenient for collection, efficient storage, and high-quality preservation; Second, in the case of sufficient computing power reserves, it has enhanced the ability to follow task instructions by fine-tuning SFT, carried out secondary pre-training experiments to learn private domain knowledge, and studied reinforcement learning technology to improve reflection and domain reasoning capabilities, creating a vertical large model that understands internal business better and has stronger job task capabilities; Third, it has established a feedback reflux mechanism for internal data to form a data ecological cycle of "capability output-data reflux-model optimization."

In simple business scenarios, there is no need for complex fine-tuning training techniques, and good results can be achieved by combining prompt words and RAG technology: first, set the key nodes of the business process in advance, and let the model perform specific tasks that it is good at, and humans are responsible for review and approval to ensure the quality and risk of the generated results; second, through external knowledge retrieval enhancement technology, create an enterprise-level knowledge project covering all business lines, connect to the Internet search to obtain rich real-time information, and provide knowledge support for large models.

In general, based on comprehensive consideration of input costs and output benefits, "simplicity is the best way" is the best way to reduce the probability and risk of hallucinations.

(2) Continue to introduce new technology architectures to enhance the capabilities of large model platforms

In order to enable R&D personnel to quickly build intelligent applications, China Guangfa Bank is building a comprehensive large-scale model platform that integrates intelligent agents and training reasoning capabilities. The platform integrates functional modules such as prompt words, RAG, workflow, plug-in library, and application interface. Based on a unified platform framework, it supports R&D and business personnel to quickly get started in a low-code manner, forming a convenient and easy-to-use intelligent agent application development model. At the same time, China Guangfa Bank actively introduces new technical architectures such as multi-agent system (MAS) and model context protocol (MCP) to build end-to-end task closed-loop processing capabilities, so that intelligent agents can dynamically analyze newly acquired knowledge, adjust strategies in a timely manner, improve the scalability and generalization of the platform, provide a one-stop smart solution, and promote the upgrade of intelligent application development from the "small workshop" model to the "assembly line" model, meet the needs of personalized business applications and long-tail scenario tool development, and accelerate the implementation of "AI+banking" application scenarios.

In order to meet the strong demand for AI business, China Guangfa Bank continues to expand the scale of large-model information creation computing power. However, intelligent computing equipment is expensive and requires huge investment. The large-scale development of intelligent applications is often constrained by the shortage of computing resources. Therefore, the large-model training and inference management platform for refined management of intelligent computing power has become a key construction direction: first, to achieve intensive management and flexible scheduling of computing power through resource pooling; second, to establish a computing power coordination mechanism, formulate a standardized application allocation process, do a good job in the refined management of inference cluster resources, and improve the utilization rate of computing power resources; third, to improve the enterprise model evaluation specifications, and realize the rapid introduction and economic adaptation of large open source models with large and small parameters.

In the AI ​​era, data engineering and knowledge engineering are the two core driving forces behind the digital and intelligent transformation of enterprises, providing high-quality scenario data and knowledge bases for fine-tuning training and application of large models.

China Guangfa Bank's data engineering is driven by improving data quality and aims to deepen data application. It continuously enriches and improves internal and external data based on the big data platform, and vigorously promotes data governance in combination with regulatory requirements and industry trends, providing high-quality data resources for big models and forming a data value cycle chain.

The planning and construction of the knowledge project of China Guangfa Bank is centered on the logical architecture of enterprise knowledge, covering financial general knowledge, enterprise integration knowledge and business domain knowledge, and building enterprise knowledge base products based on business application scenarios. Among them, enterprise integration knowledge realizes the integration and acquisition of cross-domain knowledge in a fusion cognitive way; business domain knowledge is organized, stored, divided into permissions and applied according to departments and professional categories to achieve knowledge ambiguity elimination and refined domain management.

In terms of AI talent system construction, China Guangfa Bank has built a three-tier AI talent system based on the different needs and audiences of the bank, and promoted it in a layered and classified manner: at the general knowledge level of the bank, it promotes intelligent office tools to all employees, integrates internal and external information portals, improves the corporate knowledge system, and conducts skills training on the use of large models to help all employees quickly master and use them conveniently, thereby improving daily office efficiency; at the business line level, with the goal of job performance improvement, it has established an "AI Center" to coordinate business needs and solutions, promote the deep integration of business and technology, and establish a large model "pioneer team" to explore business scenarios, comprehensively evaluate the scenario value from the aspects of application value, technical feasibility and data foundation, and promote business innovation; at the core technology level, it has established an artificial intelligence laboratory to build a team that masters the core technology of large models, conducts in-depth research on key technologies of platforms, algorithms and applications, enhances the large model business scenario development capabilities of various application R&D teams, and proficiently uses large model R&D tools to achieve efficient operation in the links of solution design, project management, development and testing, online operation and maintenance, and security management, promote high-quality delivery of software products, and provide strong support for business development.

In terms of AI talent training, China Guangfa Bank adopts a strategy that combines external introduction with internal training. On the one hand, it actively introduces mid- to high-end talents with rich experience in artificial intelligence technology to enrich the technical talent resource reserve; on the other hand, it regularly organizes "Artificial Intelligence Day" activities, carries out artificial intelligence technology training courses, competitions and laboratory technology exchange activities, creates internal communities, subscription accounts, communication groups and other communication media, helps all employees quickly master the latest technologies and tools, continuously improves their application capabilities, and forms a strong technical atmosphere.

Big model technology is both an "accelerator" and a "touchstone" for the intelligent development of the banking industry. China Guangfa Bank will anchor its course in exploration and pragmatism, and continue to expand the breadth and depth of artificial intelligence technology applications by increasing investment, optimizing mechanisms, and deepening cooperation, injecting strong momentum into intelligent transformation.

First, we will keep up with industry trends and increase investment in technology research and development . We will focus on strengthening investment in large model development, algorithm optimization, tool platform construction, etc., accelerate the construction of intelligent platforms, improve the technology development ecosystem, improve the efficiency of intelligent application incubation for technology and business personnel, and enhance independent innovation capabilities; at the same time, we will continue to consolidate computing power infrastructure, optimize computing power resource layout, and improve computing power utilization efficiency.

Second, we will focus on core scenarios and improve the path of AI empowerment . Focusing on the construction of intelligent assistants, we will create general office assistants covering all employees, as well as exclusive assistants for positions such as credit, marketing, customer service, risk control, and R&D, to achieve multi-task, full-process, and end-to-end job coverage, and promote the deep integration of big model capabilities into all business areas of the bank.

The third is to promote internal transformation and optimize management mechanisms and processes . Explore organizational collaboration mechanisms under the new situation, coordinate the application of artificial intelligence technology in various business areas, deepen the integration of technology and business, continuously improve the efficiency of technology application transformation, and continuously enrich innovation results.

Fourth, strengthen risk prevention and control, and focus on data security and privacy protection . Strengthen data security protection to ensure the safety and controllability of user data in the collection, storage, and processing stages; at the same time, strengthen research on model security technology to ensure the security, reliability, and compliance of artificial intelligence applications.

Fifth, strengthen external cooperation and build an open innovation ecosystem . Actively cooperate with leading technology companies, universities and research institutions, introduce cutting-edge concepts and technological achievements, and promote the joint research and development and application of artificial intelligence technology.

In the future, China Guangfa Bank will continue to uphold its original intention of providing finance for the people, continue to promote the big model to become an intelligent consultant throughout the front, middle and back ends, create digital employees in more business areas, and actively build a smart bank with artificial intelligence technology as the core driving force.