GuShiio.com Wen| Author, WEMONEY Research Laboratory| Wang Yanqiang
When traditional banks encounter DeepSeek, the new favorite of technology, what kind of sparks will they collide?
As DeepSeek continues to become popular, many bank officials have successfully connected to DeepSeek. It is understood that currently, there are Industrial and Commercial Bank of China, China Construction Bank, Postal Savings Bank, Bank of Jiangsu, Bank of Beijing, Bank of Chengdu, Bank of Chongqing, Zhongyuan Bank, Qingdao Rural Commercial Bank, Hai ‘an Rural Commercial Bank, Chongqing Rural Commercial Bank, Weizhong Bank, and Xinwang Bank. Fifteen banks including Bank, Baixin Bank and Sushang Bank have announced access to the DeepSeek series of models.
This wave of AI+ finance is no accident. Under the pressure of digital transformation, traditional banks are reshaping their competitiveness by deeply binding technology partners. They must not only break through in the survival battle of reducing costs and increasing efficiency, but also build a moat in value wars such as personalized services and real-time risk control. The role of technology providers such as DeepSeek has gradually escalated from being a mere exporter of tools to a strategic co-conspirator in the intelligent transformation of financial institutions.
CITIC Securities Research News pointed out that for traditional financial institutions, embracing AI change will be a must. The core values are expected to include: first, reducing costs and increasing efficiency, reducing repetitive human investment, and freeing employees to focus on high-value tasks; second, controllable risks, real-time monitoring market and operational risks to avoid major losses; third, experience upgrades, providing personalized and instant services, and enhancing customer stickiness; fourth, innovation-driven, creating differentiated competitive advantages through AI technology and seizing market opportunities.
Multiple banks are connected to DeepSeek, and each application scenario has its own focus
DeepSeek’s open source model means that it can be modified like Lego bricks, which not only lowers the threshold for using large enterprise application models, but also provides the possibility of multi-scenario intelligent applications.
Among state-owned banks, ICBC is the first in the industry to introduce DeepSeek series of open source large model bases based on ICBC Zhiyong, a self-developed and full-stack independently controllable large model platform. It has built more than 10 typical scenarios such as financial report analysis assistants and AI wealth stewards. Promote intelligent upgrading of business processes and effectively improve work quality and efficiency.
By deploying the DeepSeek model, China Construction Bank optimized the credit approval process and improved the identification accuracy of non-standard materials to the peer benchmark level. At the same time, it built a related risk map to strengthen anti-fraud capabilities and improve the accuracy of risk labeling. After the intelligent customer service system integrates DeepSeek semantic understanding technology, it significantly improves the response efficiency of complex consultations, and accelerates core system iteration with the help of code assistants.
Postal Savings Bank relies on its own large model postal intelligence to deploy and integrate the DeepSeek-V3 model and the lightweight DeepSeek-R1 inference model locally as soon as possible. First, we apply the DeepSeek model to Xiaoyou Assistant, adding logical reasoning functions to enhance precise service efficiency, and through in-depth analysis and other functions, we accurately identify user needs and provide personalized and scene-based service solutions.
Among urban businesses, Bank of Beijing joined forces with Huawei to take the lead in introducing and deploying the Deepseek series of large models as early as the end of 2024 to explore the application of the Deepseek model in the financial field. Currently, it has been piloted in multiple key business scenarios such as AIB platform Beijing Bank Research, Beijing Bank Think Tank, customer service assistant, and Beijing Bank Map.
Bank of Chongqing also recently announced through its official Weixin Official Accounts that it has completed the localized deployment and verification testing of the DeepSeek-R1 model, achieving more accurate semantic understanding, logical reasoning and multi-round dialogue capabilities.
Among rural commercial banks, many banks also connect to DeepSeek. For example, Qingdao Rural Commercial Bank revealed that it has localized deployment of Smart Qimi, an enterprise-level AI model service platform based on the DeepSeek model, and applied it to scenarios such as digital people in outlets and hall, and text verification of training textbooks to promote the further development of digital finance across the bank. Intelligent upgrade.
Chongqing Rural Commercial Bank also announced that it will use the capabilities of Tencent’s Cloud Model Knowledge Engine to launch AI Xiaoyu, an intelligent assistant application based on the DeepSeek model, on corporate WeChat, becoming the first financial institution in the country to access the DeepSeek model application.
In addition, among private banks, some banks are also the first to access DeepSeek. For example, since May 2024, Xinwang Bank has applied the DeepSeek model in system R & D scenarios to build a research and development knowledge question and answer assistant and a code continuation assistant respectively, shortening the time spent on front-line engineers reviewing technical data during the research and development process.
Significant cost reduction and efficiency increase, and more accurate risk prevention and control
Against the background of pressure on interest margins and intensified industry competition, how to reduce costs and increase efficiency has become the primary issue facing the banking industry. Through automated and intelligent solutions, DeepSeek can help banks improve work efficiency, reduce costs, and optimize business strategies., thereby maintaining a leading position in market competition.
Based on the data of Bank of Jiangsu, the bank applies the R1 reasoning model and combines the resolution and processing capabilities of the mail gateway to realize full-link automated processing of mail classification, product matching, transaction entry, and valuation table analysis and reconciliation. The identification success rate reaches 90%. Above, it has initially achieved centralized business operation. Based on the average manual operation level, it can save 9.68 hours of workload per day.
After Chengdu Bank of China connected to DeepSeek, it optimized the intelligent credit process through a small model + large model framework. The comprehensive identification rate of credit materials has increased to more than 85%, and the report generation time has been compressed from a few days to one hour, greatly improving approval efficiency; The intelligent knowledge base covers 29 business scenarios. Combined with the optimized RAG technology, the user Q & A adoption rate of 70% for user questions and answers, reducing the cost of manual knowledge retrieval; At the same time, it deploys DeepSeek based intelligent Q & A assistants in the bill business field, integrates RAG and vector database technology to accurately answer business process, regulatory consultation and other questions, reduce operational risks and improve service efficiency.
Baixin Bank also said that with the strong support of DeepSeek series of models, the capabilities of intelligent code assistants have been renewed and upgraded. First, code completion has been used; currently, 80% of R & D personnel have been covered. Many R & D personnel have reported that the application of code assistants has significantly improved R & D efficiency; Second, technical Q & A. Relying on DeepSeek’s rich knowledge base and huge code base, intelligent code assistants can not only provide various technical knowledge, but also support scenarios such as code generation and code interpretation.
In the key link of risk assessment and early warning, DeepSeek’s multimodal fusion analysis (text/image/transaction flow) improves the accuracy of risk prevention and control.
At present, Bank of Jiangsu relies on the smart Xiao Su Da Language Model service platform to locally deploy and fine-tune the DeepSeek-VL2 multimodal model and the lightweight DeepSeek-R1 inference model, and uses recognition results combined with external data to intelligently detect and verify contract information. High-risk transactions issue early warning to effectively prevent potential credit risks. After optimization using the DeepSeek model, the identification and early warning response speed is increased by 20%.
After WeBank connected to DeepSeek, it embedded the DeepSeek model into the risk control review system to cover the entire pre-loan to post-loan cycle, significantly improving credit approval efficiency and anti-fraud monitoring accuracy, and effectively reducing manual review costs; at the same time, its anti-fraud defense capabilities and credit risk monitoring efficiency have been systematically optimized, further strengthening risk early warning capabilities and assisting credit risk prevention
Sushang Bank has also integrated DeepSeek series of model technologies to build a four-in-one intelligent decision-making system of data + algorithm + computing power + scenarios to improve risk control levels. At present, this system has been used in more than 20 business scenarios such as credit risk control and anti-fraud monitoring. The efficiency of due diligence report generation has been increased by 40%, and the accuracy of fraud risk tags has been increased by 35%.
It is worth noting that although the DeepSeek model has strong general capabilities, it still needs to focus on the following risks: First, data security and privacy protection. Banks ‘business data often involves customer privacy and trade secrets. How to use these data to effectively train models while ensuring data security has become the primary consideration. The second is model reliability verification. In credit approval and contract review scenarios, DeepSeek still has logical loopholes and factual deviations.
Overall, DeepSeek promotes large models from closed source to open source, significantly reducing the threshold for localized deployment. The localized deployment model not only meets the requirements of financial institutions for data sovereignty and response speed, but also provides flexible space for AI technology to deeply adapt to segmented business scenarios, marking that the digital transformation of the banking industry has entered a precise implementation stage.
