Large language model technology is becoming the core engine driving digital transformation in the financial industry, with institutions of all sizes fully launching AI application strategies.

Industrial and Commercial Bank of China announced the addition of over 100 new application scenarios, while China Construction Bank, Bank of China, and CITIC Bank each declared having deployed AI applications across hundreds of scenarios by the end of June.

In just six months, AI applications have blossomed throughout the banking sector.

According to the 2025 interim reports of listed banks, approximately 90% of the 42 A-share listed banks disclosed their AI technology applications and implementation results.

On August 26, the State Council issued the "Opinions on Deeply Implementing the 'AI+' Action Plan," proposing to promote widespread application of new-generation intelligent terminals and AI agents in software, information, finance, business, legal, transportation, logistics, and commerce sectors. By 2027, the adoption rate of new-generation intelligent terminals and AI agents should exceed 70%; by 2030, this indicator should surpass 90%.

The financial industry, with its high degree of digitization, has positioned commercial banks at the forefront of this "AI+" wave.

According to the "2025 Financial Industry Large Model Application Report" released by Tencent Financial Research Institute, in the first half of 2025, there were 79 large model-related contract projects based on publicly disclosed information, covering banks, securities, insurance, trusts, and asset management. Among these, banking sector contracts numbered 44, accounting for over half.

At a recent media exchange meeting, Tencent Cloud Vice President Hu Liming described financial institutions' exploration of AI technology applications as "flourishing." "After DeepSeek's open-source release, foundational model capabilities reached thousands of households, enabling large, medium, and small financial institutions to rapidly develop applications on foundational models at relatively low costs."

Earlier this year, Ant Group Digital Technology Vice President Yu Bin also stated in media exchanges that financial large model deployment has moved from early exploration to a critical turning point, with financial institutions' concerns shifting from "whether to follow" to "how to implement."

Global consulting firm BCG released a report in May titled "The AI Test Has Arrived for Banking," stating that AI is comprehensively reconstructing the banking landscape, driving the industry from technology application toward fundamental logic transformation. In this process, financial institutions represented by banks will also be reshaped by AI, with latecomers potentially overtaking and pioneers possibly losing digital advantages accumulated over the past decade.

Clearly, the competition to reshape banking through AI has begun.

"Financial large model deployment has shifted from infrastructure construction toward supporting business upgrades, with some financial institution leaders beginning to consider how to reconstruct business operations and even institutional development strategies for the next three to five years," Yu Bin noted. "Anxiety and excitement, challenges and opportunities coexist, which aptly describes the current state of entire institutions."

In reality, financial large model deployment remains in "the first kilometer of a marathon." Hu Liming acknowledged that underlying foundational large models are still rapidly iterating and evolving, while large models generating value in core financial scenarios still face many challenges.

**Application Scenario Expansion: Some Banks Exceed 1,600**

At the beginning of 2025, DeepSeek released the open-source model DeepSeek-R1, which swept global app store download rankings with its powerful reasoning capabilities and exceptional cost-effectiveness, also igniting Chinese financial institutions' enthusiasm for developing AI applications.

"Financial industry large model applications have basically formed a tiered development pattern with banking leading, securities and insurance following, and trusts and asset management exploring," the aforementioned financial industry large model application report stated. "Entering 2025, industry application construction pace has significantly accelerated, with institutions of all sizes fully launching large model application planning. Large model technology is becoming the core engine driving financial industry digital transformation."

Recently, as commercial banks intensively disclosed their interim reports, more details of the banking industry's large-scale AI deployment have emerged.

For state-owned major banks and joint-stock banks, AI applications have become standard. In their 2025 interim reports, multiple banks announced substantial growth in business scenarios empowered by large models in the first half.

Among state-owned major banks, Industrial and Commercial Bank of China stated it added over 100 new application scenarios including AI wealth assistants and intelligent research assistants in the first half; China Construction Bank announced cumulative empowerment of 274 internal scenarios including credit approval, intelligent customer service, and "Bangde" personal relationship manager assistants, an increase from 193 scenarios reported in 2024; Bank of China also declared using large model technology to empower over 100 scenarios across the bank.

Among joint-stock banks, China Merchants Bank stated it had deployed 184 scenario applications across retail, corporate, risk control, operations, and office fields; CITIC Bank said it actively explored AI-empowered new paradigms in customer marketing, management decision-making, operations, and risk control, constructing over 1,600 intelligent service scenarios.

Notably, multiple bank technology personnel stated that the industry currently lacks unified standards for recognizing AI deployment scenarios, and the number of deployed scenarios also relates to each bank's statistical methodology.

Beyond scenario expansion, the number of banks disclosing AI applications is also growing. In 2025 interim reports, a significant trend is that more small and medium banks began disclosing their AI application details.

Take Changshu Bank, the first listed bank to disclose its 2025 interim report, as an example. This rural commercial bank has assets of approximately 400 billion yuan and first-half operating revenue of 6.062 billion yuan.

In its interim report, Changshu Bank stated it actively promoted localized deployment and scenario-based implementation of large model technology, completing evaluation and deployment of full-capacity DeepSeek and Qwen3-235B foundational models, and successfully launching six large model assistants including code generation, knowledge management, and credit review, covering multiple business areas such as intelligent customer service, office collaboration, development support, precision marketing, and intelligent risk control.

Additionally, multiple small and medium banks including Chongqing Bank, Bank of Jiangsu, and Ruifeng Bank disclosed their large model deployment and AI application progress in interim reports. Bank of Jiangsu stated that by the end of June 2025, it had deployed nearly 60 intelligent scenarios.

"Previously, banks truly investing in AI were mainly joint-stock and state-owned major banks. After DeepSeek, we see all types of banks investing," Tencent Cloud Commercial Banking Solutions General Manager Cao Jun stated.

Furthermore, bank wealth management subsidiaries are also actively deploying AI. BOC Wealth Management Chairman Huang Danggui stated in a public speech earlier this year that wealth management companies will actively embrace technological innovation in channel construction, product development, asset allocation, and risk management, integrating technological innovation into high-quality development of the wealth management industry.

In application scenarios, AI large models are gradually extending from edge businesses like customer service assistants to core business scenarios such as trading, marketing, and risk control.

For example, Bank of Communications announced comprehensive AI application in risk control, building a digital comprehensive risk management system covering "all group institutions, all product varieties, all risk categories," achieving risk control transformation from "human prevention" and "technical prevention" to "intelligent control."

Behind institutional and scenario expansion are tangible efficiency improvements brought by AI large models.

In the first half, Bank of Communications' model strategies deployed in personal mobile banking improved loan approval rates and withdrawal rates by 67% and 83% respectively compared to original modes; telephone banking channels' addition of customer "event-based" marketing strategies improved productivity by nearly 80% compared to regular strategies, with online-operated consumer loan balances increasing 153% year-on-year in the first half; personal mobile banking applications of AI mode improved personalized service precision, with conversion rates improving more than 2-fold.

Changshu Bank promoted business efficiency improvements of 20% through multi-scenario deployment of business intelligent agents, significantly strengthening intelligent service capabilities.

China Merchants Bank used AI technology to improve business processing efficiency and service levels, saving 4.75 million labor hours in the first half while bringing approximately 390 million yuan in economic benefits through procurement substitution and intelligent recommendation services to customers.

ICBC's financial large model "ICBC Zhiyong" undertook work equivalent to over 40,000 people annually in 2024.

"Financial institutions' exploration of AI applications presents a flourishing situation," Hu Liming stated. Currently, in code writing, employee assistance, and customer service areas, large model technology applications are relatively mature with significant efficiency improvements.

**Solving Financial Large Models: Data and Scenarios Most Critical**

Although large model-empowered banking business scenarios are gradually increasing, performance in broader and more complex scenarios remains far from transformative.

Personnel from different types of financial institutions including banks and securities firms stated that deploying AI Agents was a very important task in the first half for their institutions.

AI Agents refer to software or hardware entities capable of autonomous action, able to interact with environments and acquire information, making autonomous decisions and executing tasks through logical reasoning and self-learning. Simply put, they are "digital employees," with large models serving as their "brains."

AI Agents are viewed as key links in large model deployment, representing the "last mile" of large model implementation, with important significance for promoting intelligent upgrades in the financial industry.

An AI business leader from a technology platform stated that a typical current situation is financial institutions' different departments developing dozens of different AI Agents, but most agents are only experiential, with low accuracy in business applications and even unusable in complex scenarios.

"Currently, most AI scenarios are still exploratory in nature, relatively extensive," Hu Liming stated. "However, scenarios like code assistants, enterprise knowledge bases, and securities AI advisors are all relatively mature in actual deployment."

"When many institutions use general large models, they feel they can't achieve desired effects, sometimes feeling large models are mismatched - clearly having conventional financial logic, but not following financial compliance requirements when making analysis and decisions," Yu Bin stated. General large models not understanding finance is a major pain point for financial institutions deploying large model applications.

General large models have strong reasoning capabilities but insufficient knowledge reserves in vertical fields like finance. In practical applications, bank technology personnel often need to conduct secondary training of general large models with financial data to reach usable standards.

Large models trained for financial business scenarios are also called financial large models. Notably, multiple professionals stated that financial large models are not a single large model but a suite of models composed of different sizes. During digital employee work processes, different sized models may be used at different stages. Interim reports show multiple banks are currently improving financial large model output accuracy through combining large and small models.

Multiple interviewees stated that achieving deep application of large models in financial business requires financial institutions to build financial large models suitable for themselves, with data and scenarios being key.

"If financial institutions want to use large models to serve their customers, data aggregation, organization, and accumulation become very important," Cao Jun stated. Now we can see many banks increasingly emphasizing data importance.

Banks with better accumulation in data quantity, quality, and diversity often achieve better results under the same scenario and large model technology conditions.

China Everbright Bank Vice President Yang Bingbing previously stated in an exclusive interview that the prerequisite for intelligence is digitization. At the bank's recent interim results conference, Yang Bingbing specifically mentioned that each of the bank's three digital task forces is equipped with 75 data personnel specifically responsible for data organization and analysis.

The importance of scenarios is also very prominent.

On one hand, scenario selection significantly affects application effectiveness.

Taking debt collection scenarios as an example, according to an IT personnel from a small to medium bank, the bank previously developed a debt collection summary function using large model technology to analyze debt collection call content, discover collection clues, and generate summaries to serve collection personnel. In practice, they found that in personal consumer loan scenarios, almost no one looked at summaries; but in enterprise customer loan collection scenarios, this function was widely used by collection personnel and significantly improved recovery amounts.

The reason is that enterprise customer loans usually involve larger amounts, complex personnel and processes, and longer recovery times. Summaries can help collection personnel quickly grasp historical collection situations, formulate new collection strategies, and avoid detours, thus significantly improving application effectiveness. "Scenario selection is indeed different from what we imagined," the aforementioned IT personnel reflected.

On the other hand, what are business scenario requirements? How should scenarios be designed? These all depend on financial institutions' own deep understanding of business scenarios.

A bank technology department head stated that in helping business departments achieve scenario implementation, the bank has formed a methodology.

First, based on scenario difficulty, determine what work scenario implementation requires. Some simple scenarios just need knowledge base attachment, while more complex scenarios may require large model fine-tuning. As scenario complexity further increases, reinforcement learning, secondary training, large-scale training, or even AI Agent technology may be needed. This mainly tests financial institutions' AI technology application maturity.

Second, large model deployment in business scenarios also requires business departments to prepare sufficient training samples. Sample format, quantity, and logical reasoning chain design all need to match scenario requirements. These all require continuous experience accumulation in practice.

"Overall, after foundational large models like DeepSeek became open-source, technical gaps between financial institutions of different sizes may be shrinking, but in deep business scenario applications driven by high-quality data, gaps will continue to exist for some time," the aforementioned bank technology department head said.

**Organizational Structure Upgrades: Business and Talent Reshaping**

"At the beginning of this year, many banks were discussing how to solve C-end user service problems; now they're talking about how to improve wealth management business and make insurance claims more efficient," Yu Bin stated.

Clearly, financial large model construction has moved from infrastructure construction into the business upgrade phase, with some institutional leaders considering AI-based business reconstruction.

Multiple interviewees believe that financial institutions deploying AI large models involves not just purchasing IT systems but business process reshaping, strategic reconstruction, and driving organizational upgrades.

First, AI reshapes banking business models and service paradigms. Bank employees will see substantial reduction in repetitive labor, with more responsibilities shifting toward AI model supervision, management, and optimization.

A typical application is banks using Robotic Process Automation (RPA) to reduce grassroots burden. As of June 2025, Bank of China's RPA covered over 3,300 scenarios; Postal Savings Bank of China deployed over 2,000 RPAs with over 4,000 automated processes online.

At the 2025 World Artificial Intelligence Conference, Bank of Communications displayed its digital avatar-driven remote financial service model. The bank applied audio-visual and AI technology to construct a remote video service system, building a "four-in-one" new service model integrating and coordinating remote video seats, digital employees, relationship managers, and customer service managers. This model breaks traditional branch time-space limitations, providing customers with one-stop comprehensive financial services while improving service continuity and overall customer service capabilities.

"Bank-customer interaction modes are transforming from passive service to proactive, real-time, and personalized service," the aforementioned bank technology department head stated. In this process, large models will continuously strengthen intelligent, refined, and interactive customer service experiences, truly achieving "both old friend warmth and banker rigor."

Meanwhile, large models' capability to comprehensively analyze massive user data across multiple dimensions will effectively expand bank employee service radius. In wealth management scenarios, AI-empowered wealth managers can now serve 2,000-3,000 customers instead of the previous 200, representing at least a 10-fold improvement.

Second, commercial bank organizational and talent structures face reshaping. With deepening AI applications, banks' demand for AI-related technology talent is rising. New positions like data scientists, algorithm engineers, and AI trainers are emerging within banks, while traditional back-office operator positions are gradually decreasing.

In recent years, commercial bank recruitment has seen numerous fintech-related positions. Annual report data shows that by the end of 2024, technology field employees at six state-owned major banks totaled over 100,000.

Analysis of commercial bank recruitment position requirements shows that compared to previous years, 2025 bank technology positions have shifted from generalized names like "information technology positions" toward "artificial intelligence field positions" and "AI experts," with technical requirements upgrading from "understanding machine learning basics" to requiring proficient related skills and AI practical scenario adaptation.

For example, Industrial and Commercial Bank of China established headquarters "AI+" special recruitment positions in its 2026 campus recruitment plan, prioritizing artificial intelligence and big data related majors; Ping An Bank's campus recruitment announcement launched a "technology special session" cultivating talent through a "finance + technology" dual helix system, explicitly focusing on artificial intelligence and information security fields.

In March, China Merchants Bank released a Digital Finance Training Camp (2026 Campus Recruitment Advance Batch) recruitment announcement emphasizing "AI + Finance" direction themes, with outstanding performers receiving direct admission to 2026 campus recruitment advance batch.

"Some traditional position bank employees need to complete skill and work method changes," Cao Jun stated. "We clearly see that more and more bank internal employees are very willing to incorporate AI into their career development planning. Future banks will have increasingly more people who understand AI."

The aforementioned BCG report shows that today, a typical bank might have 15% of employees in front office, 10% in risk control departments, 10% in operations departments, 20% in corporate business departments, 20% in middle office, and 25% in technology departments. "With the development of generative AI and AI Agents, banks focused on utility services may retain more operational personnel, while other banks may prioritize consulting and technical personnel. Everything depends on each bank's respective AI vision and business strategy," the report stated.

To address related changes, banks' organizational structures also need timely adjustment. Cao Jun stated that AI large model deployment requires close collaboration between technology and business departments to truly embed AI capabilities in business scenarios.

Taking China Everbright Bank as an example, to promote digital transformation in inclusive finance, supply chain finance, and retail credit, the bank established three digital task forces, each composed of both business and technology personnel. Previously, the bank adopted a "technology secondment system," seconding technology department employees to business departments and subjecting them to business department assessments. An Everbright Bank representative stated this aimed to break departmental barriers between technology and business.

Additionally, AI applications will make bank decision-making mechanisms more intelligent.

"Through data-driven decision-making, improve decision-making scientificity and foresight, reducing errors caused by empiricism," the aforementioned bank technology department head believed. Future bank decision-making levels and processes may be reduced, emphasizing data-driven approaches with changes in departments participating in decisions.

**Regulation and Compliance: Human Oversight Extremely Important**

Notably, as a heavily regulated industry, the financial sector's compliance with AI large models and other emerging technologies receives high attention.

At the Tsinghua PBC School of Finance Global Finance Forum held in May, People's Bank of China Financial Research Institute Deputy Director Mo Wangui warned that AI technology applications may generate three new types of risks:

First, model hallucinations - finance requires high precision, professionalism, consistency, and stability; hallucinations make some areas unusable; second, algorithm black boxes - may fail to meet penetrating supervision requirements for key businesses, disadvantaging risk management tracing and responsibility determination; third, potentially amplifying traditional risks, such as strengthening pro-cyclical behavior, over-dependence on few technology companies, and consumer rights protection related risks.

Mo Wangui suggested financial institutions be more prudent: first, whether business scenarios and specific technologies match - blindly applying many complex advanced technologies to simple business scenarios may become complicated and bring unnecessary risks, so compatibility is important; second, minimize over-dependence on large technology companies, develop personalized solutions suitable for themselves, avoiding herd effects; third, human-machine collaboration may require more attention from financial institutions, particularly incorporating large model applications into internal overall risk control and compliance mechanisms, emphasizing business process reengineering.

"Personally, I feel that some key business processes, decision-making business processes and task nodes require human intervention, and human-machine collaboration and human intervention will increase controllability," Mo Wangui stated.

Yu Bin summarized that commercial banks currently deploy AI large models mainly through four paths:

First, starting with AI technology platform construction, gradually completing foundational capability building including data governance, computing power construction, AI platforms, and financial tool sets. Based on this, run through business chains via general scenarios like intelligent customer service to further improve infrastructure.

Second, transforming existing mobile banking apps into AI mobile banking, shifting customer interaction modes from "people seeking services" to "services seeking people."

Third, constructing AI Agents based on business scenarios, usually starting with subdivided industry scenarios where AI is most needed to maintain "moats."

Fourth, bank-wide phased construction. Through systematic planning, comprehensively reconstructing business processes with large models, building AI Agent clusters, driving business innovation and experience upgrades.

Yu Bin stated that state-owned major banks with strong R&D capabilities basically choose to build their own computing power, construct AI infrastructure, and build intelligent agent development platforms independently. Joint-stock banks and leading city commercial banks have more diverse implementation models covering all four aforementioned paths. Smaller regional banks initially tend to test waters in some scenarios with smaller investments, gradually advancing AI large model deployment in layered and classified approaches.

Regardless of path, multiple interviewees stated that currently, financial institutions mainly strengthen security protection in large model applications through introducing compliant foundational models, improving data quality, strengthening model evaluation, sensitive word filtering, and human oversight.

Currently, most AI applications are mainly used for internal services, empowering internal employees, with very few directly facing customers. Even in the relatively mature intelligent customer service field, AI currently serves more as an assistant to customer service personnel, with human oversight still controlling key links.