The Boundaries of Bank AI Applications: Non-Customer-Facing Scenarios First, High-Risk Decisions Still Left to Humans

Main text

"Of these three poster versions, which one do you think looks better?" the head of a bank's publicity department asked, after sending Jiemian News a set of candidate templates for the bank's Double Ninth Festival promotional poster.

"The first one, it's warmer."

"No. 1 was made by me; 2 and 3 were made by a design company." The person in charge showed off her artificial intelligence (AI) application results to the Jiemian News reporter, not without pride.

Another head of a bank's publicity department told Jiemian News, "Using a large AI model to make promotional posters takes only a few minutes. In many cases we no longer need to look for outside suppliers, and procurement costs have been cut by more than 80%."

On June 18, 2026, the National Financial Regulatory Administration issued the Guiding Opinions on the Safe Development and Application of Artificial Intelligence in the Banking and Insurance Industries (hereinafter referred to as the Opinions), clarifying the basic requirements for financial institutions' AI applications in terms of governance architecture, data security, risk management and other aspects.

On October 19, ICBC President Liu Jun, speaking about banks' AI applications at the 2026 Bund Summit, went straight to the core of the issue: AI can increasingly resemble an "employee," but at the level of legal liability, risk bearing and final decision-making, it is still not a true "responsible entity."

A corporate account manager opens the "one page before the visit" intelligent assistant and enters a company name; information on business registration, equity, public opinion, litigation and more is automatically aggregated, and an interview outline and risk alerts are generated along with it. What once took hours of pre-visit preparation is now completed in 8 minutes;

When tellers at branches have questions about policies or product rules, they directly consult the bank's in-house AI assistant. CCB disclosed that its AI assistant coverage for branch questions has reached 99.42%, with more than 100,000 daily visits;

The Bank of Shanghai's AI robot for interbank certificates of deposit automatically completes the inquiry, quotation and subscription process, having handled more than 2,700 inquiry messages and helped raise over 10 billion yuan in funds.

In credit assistance scenarios, China Merchants Bank disclosed that about 90% of the basic workload of pre-loan due diligence reports for small businesses is completed by AI, that processing time for guarantee business has been compressed from more than 30 hours to an average of 2.72 hours, that bid guarantees support issuance within minutes, and that post-loan AI risk warnings come on average 45 days in advance.

A person from a joint-stock bank gave Jiemian News an example: the bank built a consumer rights protection intelligent review platform based on a large model, with "AI intelligent review upfront, manual review as the backstop" as its basic architecture, conducting intelligent review of credit products, agency wealth management products and marketing materials. AI is responsible for identifying non-compliant wording, matching laws and regulations and generating revision suggestions, but reviewers can still modify or supplement the AI's opinions, and the final professional judgment remains in human hands. More than a year after the platform went live, AI opinions have covered 90% of review scenarios, and review time per document has been compressed from 30 minutes to under 5 minutes.

Multiple banking industry sources told Jiemian News that, judging from current implementation, non-customer-facing areas such as intelligent customer service, marketing recommendations, risk warnings and operational efficiency remain the areas where banks can more easily scale up.

Ping An Bank's third-quarter report disclosed that by the end of September it had implemented more than 500 large-model application scenarios, launched 164 operational review intelligent agents covering 67 business scenarios, and improved average operational efficiency by about 57%; its digital employee platform has more than 10,000 registered users and over 7.3 million cumulative calls; in the third quarter, average daily token consumption was 21.7 billion, nearly three times the previous quarter.

A management figure at a state-owned bank summed it up: "The common feature of the above scenarios is that even if AI makes a certain degree of error, there is usually still room for manual verification or subsequent correction, and it does not directly touch major risk decisions. By contrast, there are high-risk links such as credit approval, fund trading, investment advice and price management. The common feature of these businesses is that once a judgment is wrong, what is affected is no longer just service efficiency, but may directly involve customers' property rights and interests, bank asset quality and even financial market stability."

"Regulators have always encouraged more experimentation with AI applications in customer-facing business, but cautious about large-scale promotion. As far as I know, some leading banks have developed customer-facing financial large models based on their own banks that can complete the entire process of services such as customer wealth management allocation and purchase recommendations, but they have not yet received regulatory approval," a banking source told Jiemian News.

A banking industry source told Jiemian News that banks have in fact already begun to tier AI according to business risk. Low-risk scenarios can adopt more automated approaches; in medium-risk scenarios, AI first gives suggestions, which are then reviewed by humans; for high-risk business, humans should make the final decision, with AI taking on more of the work of information organization, data analysis and assisting judgment.

In the past, when banks used information systems, the chain of responsibility was relatively clear. The system was merely a tool; employees operated according to rules and processes, and ultimate responsibility still fell on people. But once AI begins to have the ability to autonomously generate, judge and even call systems to execute tasks, the traditional relationship between "tool and user" is becoming complicated.

At the 2026 Bund Summit, Liu Jun proposed that after AI enters financial business, the issues of "identity" and "responsibility" must first be resolved. He believes that in traditional banking business, employees cannot be exempted from liability because "the system was not developed by them," but the AI business chain may simultaneously involve algorithm developers, programmers, deployers, users and end users, and the algorithm itself has a certain black-box nature, making it more difficult to determine and verify responsibility.

A bank software developer told Jiemian News that, taking contract compliance review as an example, if a bank does not have a self-developed underlying foundation model, and instead an external large-model vendor provides the base model, an outsourced algorithm team does the fine-tuning, the bank's compliance department labels the training samples, the technology department handles in-bank deployment, and compliance reviewers accept the modification suggestions given by AI, then this chain involves a great many parties with rights and responsibilities, and it is very difficult to investigate exactly which link went wrong.

The "Opinions" issued by the National Financial Regulatory Administration on June 18 explicitly proposed that financial institutions should adhere to "whoever uses it is responsible," compacting the main responsibility of financial institutions as financial service providers and users of artificial intelligence technology, while also emphasizing classified and tiered management. For those using generative artificial intelligence for public-facing services or high-risk scenarios, financial institutions need to report to the National Financial Regulatory Administration or its dispatched agencies.

A banking industry figure told Jiemian News that for banks, AI procurement is actually only the first step; more important is clarifying the chain of responsibility after the model enters the business, including who is responsible for data, who is responsible for the model, who is responsible for business rules, who reviews AI output, who approves going live, and who has the authority to stop the model's operation when anomalies occur.

This also explains why the banking industry currently prefers to advance AI in relatively controllable internal scenarios. The core reason banks remain cautious about AI is not that model capabilities are insufficient, but that financial business has far higher requirements for certainty, stability and a closed loop of responsibility than ordinary commercial scenarios.

In fact, in addition to the difficulty of identifying the responsible parties in business applications themselves, the divergence in banks' current technology routes is also a "ceiling" that application implementation needs to consider prudently.

A banking industry figure in the eastern region said that in choosing large-model applications, banks have roughly split into two paths:

Some banks choose to cooperate with large technology vendors to establish a unified large-model foundation, and then fine-tune it for different financial scenarios;

Other banks adopt a "let a hundred flowers bloom" approach, with each business line piloting different models on its own, and considering unified migration only after one to two years of market validation.

The former approach is more efficient but may lead to vendor lock-in; the latter has relatively dispersed trial-and-error costs but may result in system fragmentation and duplicated construction.

The aforementioned person acknowledged that for large banks, this actually adds another layer of new AI risk: technology dependence risk. If a bank's core business is heavily tied to a certain model vendor, then future changes in the model's own iteration capability, architecture, pricing system, data interfaces and even technology route may in turn affect the bank itself.

Reshaping organizational structure and business processes

Agricultural Bank of China Chairman Gu Shu proposed at the 2026 Financial Street Forum annual meeting that digital technology is driving the financial business model from "serial operation and standardized supply" to "intelligent collaboration and scenario customization," risk management from "experience-based judgment and responsive handling" to "real-time perception and proactive intervention," and the financial ecosystem from being centered on "people + channels + capital" to a new ecosystem in which data and algorithms play a more important role.

The aforementioned bank management figure said that previously bank digitalization was more about moving offline processes online; in the AI era, a digital employee can call data across multiple systems, form judgments and push the process to the next link. This clearly means changes in bank business processes and organizational structure.

A figure from a major state-owned bank told Jiemian News that in the past, in the risk control line, work such as risk warnings, material organization, report writing and sorting out regulatory calibers often required a large amount of manual work, a considerable part of which involved work with clear rules and standardized processes but that was time-consuming. Although core credit judgments still need to rely on practitioners' experience, large models have clearly compressed the time for these links.

A banking industry figure believes that the changes that may occur in many middle- and back-office positions in banks in the future are not simply "job disappearance," but changes in the task structure within positions. AI may undertake the vast majority of standardized affairs, while people focus more on a small number of complex, high-risk matters and matters requiring communication and coordination. This is also why the banking industry's AI application cannot simply be measured by "how many people are replaced."

A banking industry figure in the eastern region also believes that the impact differs greatly among different banks and among different departments within the same bank. Taking banks with a centralized management model at the head office as an example, AI has a greater impact on head office personnel, while the impact on branches is relatively limited. The core functions of branches lean toward marketing and customer acquisition; a large number of functions that can be empowered by AI, such as risk control, compliance and management statistics, are concentrated at the head office, so middle- and back-office positions at the head office are more prominently affected by AI. "A large amount of repetitive work such as risk warnings, plan writing and caliber sorting that used to require multiple people and several days can now produce a first draft in a short time with the help of large models, greatly compressing working time. But this does not mean that core risk approval positions will be completely replaced: key credit granting and core risk judgments still rely on practitioners' professional experience."

The aforementioned state-owned bank management figure also believes that AI will indeed replace some simple repetitive work, but customer managers' ability to maintain customer trust, make professional judgments and provide emotional value will not disappear because of this; instead, it may become more important as basic affairs are taken over by machines.

At present, some banks have begun to shift from "model building" to "enterprise-level governance."

China Construction Bank Chairman Zhang Jinliang disclosed that the bank's AI technology has covered more than 900 application scenarios, but at the same time emphasized that financial institutions should coordinate development and security, improve the full-lifecycle governance system, and strengthen independent controllability of core technologies.

The bank software developer told Jiemian News that as the number of scenarios grows, what banks ultimately need to establish is something like "full-lifecycle management" for bank employees: stress testing and risk assessment before launch, continuous monitoring of model performance after launch, auditing of data, models, prompts, knowledge bases and invocation permissions, and establishment of manual takeover and emergency stop mechanisms under abnormal circumstances.

Multiple industry insiders said that the truly mature AI model for the banking industry in the future may not be to let AI completely replace people, but to let AI undertake the work that machines are better at, so that people can refocus their energy on judgment, responsibility, trust and value choices that machines cannot easily replace.

As AI enters banks at an ever-faster pace, what truly determines the industry's competitiveness may no longer be who has more models, but who can take the lead in establishing a clear set of boundaries: letting AI fully automate where automation is appropriate, effectively assist where it should assist people, and always leave the decision-making power in human hands where people must bear responsibility.

(Source: Jiemian News) Middle of the main text Embedded advertisement