Large financial institutions are pivoting from isolated AI Agent trials to comprehensive, industrialized deployment frameworks, with recent procurement data highlighting a significant shift in capital allocation. In early September, Zheshang Bank disclosed the results of a tender for foundational Agent software and hardware, with a bid price approaching 65 million yuan, marking the largest publicly recorded financial institution Agent project this year. When combined with earlier investments in wealth management and supply chain Agent systems, Zheshang Bank's total public procurement for Agent technology has surpassed 74 million yuan.
This pattern is not unique, as Bank of Shanghai, Shanghai Pudong Development Bank, and Hengfeng Bank have all recently finalized substantial purchases, channeling funds into Agent workstations, development platforms, and operational monitoring tools. Banks are clearly spending more on Agent technology, but crucially, these funds are now being directed toward deeper, more foundational engineering capabilities.
At major industry conferences like the China International Fair for Trade in Services and the INCLUSION Conference on the Bund, senior bank executives are now focusing on how to scale Agent development, evaluate post-launch performance, and manage these systems continuously. Some fintech companies are even labeling their Agent platforms as "super factories," aiming to integrate development, deployment, evaluation, and management within a single unified framework.
Where the money is flowing: the broader shift in perspective points to a new overarching goal for banks, moving beyond deploying isolated Agents to establishing a robust base capability for the repeated production, invocation, and management of these AI entities. In real-world applications, Agents must access knowledge bases, tools, and business systems to execute tasks. For the banking industry, this implementation presents a specific challenge: institutions must not only verify output results but also meticulously control the entire task execution process.
One data architect at a joint-stock bank explained that not every scenario suits direct Agent handling. For instance, tasks involving intent recognition and contract element extraction, characterized by non-standard inputs and standard outputs, are relatively straightforward to implement. Conversely, complex processes like credit approval and financial planning require embedding the model into existing workflows, with control exercised through established rules, permissions, and manual intervention points.
This evolution is fundamentally altering the traditional research, development, and delivery methods of bank technology projects. Historically, business units raised requirements for the tech department to develop. Now, as Agents enter production, business rules, standard operating procedures, and experiential judgment must be translated into logic that models can recognize and execute, necessitating deeper business-side participation in testing and refinement. The data architect noted that in this new development model, business experts must be deeply involved in codifying knowledge and rules, while technical staff are responsible for converting this expertise into executable processes and tools, extending collaboration beyond simple requirements delivery to joint development and continuous adjustment.
As Agents move from pilot phases to mass adoption, the development, evaluation, and operational management tasks previously handled by individual projects are now creating a demand for standardization. Recent procurement activities reflect this shift. Industry data shows that in the first half of this year, the financial sector saw 406 winning bids for large-model AI projects, a year-on-year increase of 110%, with banks accounting for nearly half of both the project count and total investment. Since the latter half of last year, there has been a noticeable rise in Agent-specific application projects, indicating that bank AI construction is moving from model deployment and computing power expansion towards more specific application layers.
Notably, Zheshang Bank's near-65-million-yuan foundational Agent software and hardware project covers computing power, knowledge engineering, models, and the Agent runtime environment, with its scope extending to the foundational environment for future Agent development and operations. Concurrently, Shanghai Pudong Development Bank is procuring an Agent creation platform, focusing on the batch creation and management of Agents, while Hengfeng Bank is independently building Agent observability capabilities, extending its investment to post-launch operational monitoring.
These various project types target different components of the Agent production system, demonstrating that banks are extending their investment focus from specific applications to the supporting infrastructure for development and operations. This shift is also reflected in the responsibilities and hiring requirements of technology departments. A source from a fintech company noted that as Agent numbers grow, bank tech departments must provide unified development and runtime environments, managing challenges related to resource invocation, Agent coordination, and robust governance.
Recent job postings confirm this trend. Bank of Ningbo has listed roles for Agent workflow construction and technical framework R&D, while Ping An Bank is recruiting AI Agent evaluation engineers with responsibilities spanning tool invocation, multi-step reasoning, and safety and performance testing. These roles indicate that as banks enter the mass-production phase of Agent deployment, their engineering skill requirements now extend to workflow design, tool integration, evaluation methodologies, and operational management.
While wealth management, credit, and customer service Agents continue to be developed according to their distinct knowledge domains, processes, and risk boundaries, a relatively unified methodology for their development, validation, and operational management is gradually emerging. As Agent adoption scales, this production methodology itself is becoming the new foundation for bank AI investment.
Reworking core processes: integrating Agents into a bank's core business presents the most complex challenge of re-dividing tasks within existing workflows. The credit sector serves as a prime example. This year, Bank of Suzhou has successively advanced projects in document recognition, due diligence, AI-assisted approval, loan disbursement review, and post-loan management. Similarly, Bank of Shanghai is deploying Agents in the marketing, due diligence, and post-loan phases of its corporate credit business.
These projects span different stages of a credit issuance. Tasks like material processing, data extraction, and financial analysis are relatively standardized and more easily handled by machines. However, as you move towards risk judgment and approval, the reliance on business experience increases significantly. Determining whether a set of financial indicators is anomalous, which questions warrant further investigation, or what specific indicators mean across different industry cycles often lacks a fixed answer, with much of this judgment residing in the accumulated expertise of seasoned credit officers.
Some fintech companies are now attempting to dissect and codify this experiential judgment to re-embed it into the credit process. Jin Mei, General Manager of the AI Innovation Center at Digital China Information Service Company, explained that their CreditMind solution reorganizes historical due diligence reports, approval opinions, and loan performance data to extract the logic and expertise used by senior credit officers. This knowledge is then utilized in new approval tasks. In this new credit workflow, the system first processes client materials and performs basic analysis, then consults historical cases and expert experience to supplement risk alerts, with final decision-making remaining with human approvers.
The value of such initiatives lies in transforming some of the judgment previously stored only in individual experience into Skills that Agents can invoke, reintroducing it directly into the approval workflow. However, only relatively explicit judgment logic can be effectively codified; decisions that truly depend on specific client relationships, industry context, and long-term experience must still be made by humans. Within this framework, Agent integration into credit processes involves breaking down original tasks into finer components: standardized information for machines, extractable experience for Skills, and critical risk judgments for humans.
This breakdown is expanding to a broader scale. Shanghai Pudong Development Bank has reported that in the first half of this year, it deployed over 440 AI applications, with 220 Agents achieving engineered implementation. The same period saw the rollout of 108 AI pipelines, where the "digital workforce" handled work equivalent to more than 2,500 person-years. The concept of "AI pipelines" reveals more about process-level changes than a simple count of individual Agents: different AI capabilities are reorganized according to business sequence, with machines performing certain tasks continuously before handing off to humans at points requiring judgment, authorization, and accountability.
Zheng Bo, Chief AI Scientist at Shanghai Pudong Development Bank, defines the arrival of financial AI in production by three criteria: integration into core business processes, taking on key tasks previously done by humans, and improving customer experience, operational efficiency, or risk management. For banks, the depth of Agent integration increasingly depends on how finely processes can be divided. The clearer the boundaries around what can be reliably automated, what experience can be codified into Skills, and what judgments must remain human, the more likely Agents are to become truly embedded in core business operations.
Managing post-production operations: once Agents are integrated into business workflows, bank management concerns extend beyond simple system uptime. They must ensure business outcomes are consistently achieved, determine if continued investment is justified, and ensure processes can be smoothly recovered in the event of failure. The first priority is continuous calibration. An Agent passing tests at launch does not guarantee its performance will remain stable. Business rules, knowledge content, and customer language all evolve, and new edge cases emerge in real-world use. Banks must continuously diagnose whether performance issues stem from a knowledge gap, process design, or model capability, and then make targeted adjustments.
Wang Jun, a professional from a fintech company, noted that traditional software is primarily maintained, but Agents require continuous "operation." This involves not only whether the system is online but also whether output meets business requirements, where problems originate, and how to correct them. This makes post-launch evaluation more critical. Zheng Bo identifies production capacity, value operation, and organizational evolution as key factors for competition in financial AI over the coming years. Chen Haiqiang, Chairman of Zheshang Bank, states that technology investment must become a profitable component of cost reduction and efficiency improvement, assessing actual output through metrics like usage frequency, workload undertaken, processing efficiency, and the results of customer conversion and risk management.
These outcomes influence future resource allocation: Agents that are reliable and create business value can be iterated upon, while applications that are persistently low-use or have limited effect must be adjusted or retired. Another operational priority for banks is business continuity. In credit, for example, if an Agent is involved in data reading, due diligence analysis, and approval workflow routing, banks must be able to pinpoint exactly where a task stopped during an anomaly, which results have already been passed down the line, whether any approval actions were triggered, and who is responsible for taking over. For tasks involving transactions or system operations, they must confirm whether instructions were executed to prevent duplicate actions during manual takeover or new operational risks from unclear statuses.
Wang Jun mentions that while the industry has begun discussing the implementation of "kill switches" for Agents, the greater challenge lies in determining how business continues after the Agent is halted. These requirements are now entering banks' construction plans. Hengfeng Bank's development of standalone Agent observability capabilities reflects a growing focus on post-launch operational status and anomaly detection. Only with a clear understanding of what an Agent is processing and its current stage can a bank intervene promptly when an anomaly occurs. Continuous calibration determines whether an Agent becomes more accurate over time, value evaluation determines whether it is worth retaining, and anomaly handling determines whether banks will be confident in delegating more work. As Agents become permanent nodes in business processes, operational capability becomes the new critical differentiator for successful large-scale implementation.