When Industrial and Commercial Bank of China (ICBC) stated in its annual report that its "AI digital employees handle an annual workload equivalent to 55,000 person-years," readers might pause to consider: what does 55,000 person-years truly mean? A rough estimate based on average commercial bank salaries suggests this corresponds to tens of billions of yuan in labor costs. However, a thorough review of the annual report reveals no matching reduction in costs or surge in profits.
Over the past year, a "digital competition" centered on AI efficacy has quietly intensified during the bank reporting season. China Merchants Bank claimed in its report that AI saved over 15.56 million hours of manual labor during the period. China CITIC Bank stated that its intelligent models enhanced efficiency by over 17,000 person-years. JPMorgan Chase CEO Jamie Dimon was more direct, stating an investment of $2 billion yielded approximately $2 billion in benefits.
These figures are impressive, but when reporters visited bank frontlines, they heard a different story.
A corporate client manager at a major state-owned bank in western China admitted that daily frontline operations still rely on traditional methods. Counter services, document review, corporate material organization, ledger statistics, and client liaison are largely done manually. "The so-called AI is mostly basic辅助 tools, like simple copy generation and basic data sorting. It doesn't amount to 'replacing manpower.' Metrics like labor substitution and intelligent cost reduction mentioned in the head office's annual report have very weak penetration at the city-level secondary branches. We don't feel any noticeable cost reduction or profit improvement."
On one side, annual reports showcase dazzling conversions involving tens of thousands of "virtual employees." On the other, frontline staff respond with a "we don't feel it"冷淡. Where does this温差 come from? As the banking industry's trillions in tech investment face scrutiny over returns, the true ledger of AI efficacy is being re-examined.
Delving into specific daily work scenarios, the efficiency gains from AI are not empty talk.
A person from the asset management department at a city commercial bank's head office in western China painted a compelling picture: "Take a client manager's daily work. Recently, using a large model to analyze financial statements—throwing in basic data and getting precise results in seconds for what a manager might need a whole day to organize, calculate, and analyze. It's确实方便和高效了很多."
This efficiency leap provides tangible relief for client managers who "market to clients during the day and write due diligence reports at night." However, this asset management professional then revealed a deeper anxiety behind the efficiency: "Microscopically, it improves efficiency. In the long term, will we still need so many client managers? After all, gradually, a single manager can use AI to manage more clients. Whether client managers or other positions, there's more anxiety—industry decline, internal competition pressure, the risk of随时被替代... In the future, we might only need people to拉资源. Before, working hard could offer some self-preservation."
This voice from the frontline of a western city commercial bank reveals the duality of AI's penetration in banking: efficiency gains at the tool level are真实可感, but substitution anxiety at the organizational level is equally real. When "seconds" can complete work that used to take a "day," the other side of the efficiency equation inevitably points to a recalculation of manpower needs.
This anxiety is not unfounded. Data shows that in 2025, the total financial technology investment of the six major state-owned banks reached 130.091 billion yuan. Over a longer周期, from 2021 to 2025, the cumulative tech investment of just the six major banks was about 600 billion yuan. Including leading joint-stock banks and city commercial banks, the industry's cumulative tech investment may approach one trillion yuan. Behind this real monetary investment lies an expectation for output and efficiency.
The problem, however, is precisely that the individual sense of efficiency提升 has not yet translated into a perceptible change at the organizational level.
A资深银行业研究人士 pointed out that frontline bank perception of AI currently shows a clear "两层分化": at the tool level, applications like智能客服, report analysis, and copy generation确实提升单点效率; but at the job position level, the daily working模式 of the vast majority of frontline employees has not fundamentally changed. The root cause is that bank AI applications are still primarily "点状嵌入" and have not penetrated to the底层重构 of business processes. It's like putting a better whip on a carriage rather than replacing it with a car—the whip确实更好用了, but the carriage is still a carriage.
If the perception at frontline branches is "having tools but no颠覆," then institutions further down the chain feel "it has barely begun."
The head of the discipline inspection and comprehensive management department at another city commercial bank in western China直言: "Currently, AI applications are almost imperceptible in our bank, and we haven't used terms like '人工替代'." He特意举了一个颇具象征意义的细节: "For example, for some comprehensive work, documents from the head office's highest level this year explicitly禁止 using AI to generate研讨材料讲稿等."
In his view, the落地 of tech applications is not simply about tool deployment but adapting a整套组织能力. "前沿感知 and前沿应用 of cutting-edge technology必然 appear at society's前沿, but require前沿的思想,前沿的人才,前沿的管理模式和运营基础. The focus in relatively偏远地区根本不在这些事情上, at least not currently."
This坦诚表述 from the western city commercial bank reveals a reality obscured by grand narratives like "trillion-yuan investment" and "person-year conversion": there is a significant梯度落差 in tech investment and AI application within the banking industry. While state-owned major banks and leading joint-stock banks高调宣布 breakthroughs in 500 or 800 AI application scenarios, many small and medium-sized banks, especially regional banks in central and western China, are still in the "we don't feel it" stage.
A深层问题 arising from this梯度落差 is: when the industry discusses AI efficacy using the standards of leading banks, does it忽视 the真实处境 of a broader群体? The aforementioned research person believes the "温差效应" of AI application in banking exists both vertically between head office reports and frontline branches and horizontally between large banks and中小银行. These two温差的叠加 constitutes the most值得警惕的认知偏差 in the current AI narrative of the banking industry.
A closer look at bank frontlines reveals that AI application is not "everything can be AI" but operates within a complex framework of compliance and risk considerations,寻找边界.
The head of a secondary branch business department at a state-owned major bank in western China现场演示了 the use of the "智能客服" in the mobile banking app. She stated: "You can try asking the AI客服 a few刁钻的问题. For people with higher education and personal文化素养, they can actually solve their疑问 themselves in the mobile bank. This一定程度上释放了一线员工的人力." In her view, this is the most直观红利 that major bank tech strength brings to the frontline.
However, AI applications for internal bank employees are推进 at a more审慎的节奏. This负责人描述道: "Internal employee AI applications are still being不断优化. Currently, using them is similar to市面上的大模型. You can ask it related business questions, consult the menu for unfamiliar procedures, or inquire about unclear regulations. But it doesn't保证完全正确. If it's wrong, feedback is sent to the backend for further optimization."
The qualifier "不保证完全正确" points to the core掣肘 of bank AI applications. Banking operations have extremely low容错率, involving critical links like client funds, regulatory compliance, and risk pricing. Any model "幻觉" could lead to serious consequences.
This is also why the aforementioned city commercial bank discipline inspection head特意 mentioned that "usable areas, like智能客服问答, report draft generation, regulation retrieval, and simple copy drafting, are being逐步铺开; unusable areas, like深度分析 involving confidential data, key credit decision-making, and regulatory reporting material generation, are strictly restricted."
The aforementioned state-owned bank client manager also has a清醒认识: "Currently, AI is尚不成熟, and banking operations are特殊性质. Many involve confidentiality, with特别考虑 for risk control in this area. So, we haven't reached the point of using AI to替代一部分人,一部分工作."
IBM's Institute for Business Value report "2025 Banking and Financial Markets Outlook" confirms this judgment: as of 2024, only 8% of banks were systematically developing AI technology, while 78% still employed tactical, fragmented methods. Even in technologically领先的机构, the breadth and depth of AI落地 are far from reaching the scale of "systematic substitution."
Although substitution is far from arriving, changes are already occurring at the job level—bank frontline employees are quietly assuming a new role: AI trainers.
The aforementioned major bank secondary branch head described the daily operational logic of this role: "Frontline branches接触的层面比较基础. Counter staff mainly consult it on基础问题, like specific操作方法 and业务制度. Middle and back-office staff have higher维度—some are programmers implementing需求, others are工作人员让AI建立模型并跑需求. Client managers肯定需要它建立模型, and branch managers肯定需要它干更高级的事."
She further stated that the bank has also deployed AI tools for document writing, capable of快速检索并整合信息 into "文字+表格" outputs. However, a key环节 in the usage process is manual verification. "If it's wrong, feedback goes to the backend for further optimization."
This means that bank employees have not become "the被替代者" due to AI's appearance. Instead, in practical operations, they act as the model's "data annotators," "effect feedback providers," and "scenario trainers." This role shift indicates that the真实路径 for releasing AI efficacy is not "machines replacing people" but "people training machines, machines assisting people"—a渐进过程 requiring大量人力参与 and持续迭代.
The aforementioned research person评论称 that the phenomenon of bank frontline employees assuming the "AI trainer" role恰恰说明 AI in finance is not an "即插即用" efficiency tool. Its effectiveness高度依赖 the feedback and corrections from users, and this process itself consumes大量人力资源. This also部分解释了 why, despite巨额科技预算投入 and hundreds of AI场景上线, frontline体感 remains "workload并未明显减少"—because while AI消解部分工作任务, it also creates new work需求: testing, feedback,校正, and兜底.
Taking a纵览全局 view, banking AI application stands at a关键转折点: shifting from "daring to invest" to "knowing how to calculate."
The aforementioned research person believes the core困境 in current banking AI efficacy assessment is not a lack of data but the absence of a公认的价值核算框架. Currently, various banks use different统计口径. Metrics like "person-year," "hour substitution," and "scenario quantity" are not mutually通用, making横向比较 difficult and direct mapping to财务指标 challenging. Some banks convert AI efficacy into人力成本节省, but this conversion assumes "every task completed by AI originally required an equivalent amount of manual labor"—an assumption本身值得商榷.
Taking an indicator like "saving 15.56 million hours" as an example, it might describe a situation where, despite业务量增长, the bank did not add corresponding staffing, rather than directly reducing headcount for变现. This效益 of "avoiding增编" is not reflected in the profit and loss statement, but it is真实 for optimizing the bank's long-term cost structure.
An industry insider noted that, from a more宏观的视角, the attitude towards tech investment among leading banks is also changing.公开数据显示 that after reaching a peak in 2023, the scale of tech investment at several leading joint-stock banks连续回落 in 2024-2025. Is this a natural回调 after前期投入趋于饱和, or is bank patience with AI investment回报正在收窄? The answer might be a bit of both, but the larger背景 is: with持续承压 on净息差, the space for AI "试错" in banks has significantly narrowed.
Comments