How bank agents can get stuck in real business scenarios

```

The Agent boom is sweeping across the entire banking industry.

At the June 18 Lujiazui Forum, Agricultural Bank Chairman Gu Shu, Bank of China President Zhang Hui, and several other major bank executives shared their views on financial Agents;

Almost at the same time, Alibaba Cloud Vice President Zhang Chi also stated at the China International Finance Exhibition that financial Agents have truly ushered in their first year.

From major banks' technology plans, large tech companies' product strategies, to fintech firms' application explorations, Agents are becoming a new lever for banks’ digital transformation.

But on the frontline of banks, the actual use of Agents is far from matching the enthusiasm conveyed in forums.

Shuhui’s research into several large and medium-sized banks found that, due to compliance, data security, and complex business processes, most banks’ Agents are still mainly in R&D, office, or testing environments, with distance from core business processes.

In the banking context, Agents include at least three layers: tools for office, programming, and knowledge Q&A; tools for business assistance like customer service, marketing, and risk control; and process-type Agents that can be embedded in production flows, call systems, and participate in task execution.

Currently, the fastest progress remains in the first category; the latter two are still far from large-scale adoption.

This forms the core gap in banks’ Agents today: strategic narratives are unfolding, but real business is still outside the door.

Business Side Thresholds

Looking at process breakdowns, banking businesses such as credit, wealth, customer service, marketing, risk control, and operations all have space to be reorganized by Agents;

But the heart of financial business is not just efficiency, but authorization, auditing, and accountability; the business-side threshold only truly emerges when the Agent interacts with real data, is embedded in business processes, and participates in task execution.

This is also the fundamental reason banks remain cautious.

Chen Hua (pseudonym) from the technology department of a joint-stock bank revealed to Shuhui that the company allows R&D personnel to freely deploy various Agents in test environments, but once it involves real scenarios, the review authority is extremely strict.

“The test environment is certainly free—it’s just an issue of running fast or slow,” Chen Hua said, “but for security reasons, these Agents’ review authority at most extends to office settings—they cannot enter the business side.”

Multiple personnel from technology departments of state-owned and joint-stock banks also told Shuhui that few Agents have truly entered core business processes and participated in task execution; most trials remain in office, R&D, customer service assistance, or low-risk pilot scenarios.

This caution stems from the complexity inherent to the financial industry itself.

Dr. Jin Mei, General Manager of the AI Innovation Center at Digital China, summarized this complexity as three words: “Strict, Dense, Valuable.”

First is strict regulation—sensitive data must not leave the bank environment, solutions must be private and controllable, and every judgment must be reproducible and auditable;

Second is dense business links—financial scenes have tightly interlocked business steps; any problem can lead to risk;

Third is expensive human capital—the judgment accumulated by experienced employees over years is difficult to fully encode into rules.

Of course, there are at least some attempts on the business side.

An individual from the technology department of a state-owned major bank revealed to Shuhui that the company already has a small number of business-type Agents in use, but mainly in personal finance and internal office scenarios, and they resemble customer service assistants more.

This person acknowledged that Agent development is still led by the technology department, “The process is for business departments to submit requirements based on actual situations, but R&D still takes the lead on projects.”

Compared to the business side’s caution, breakthroughs on the R&D side come a bit faster.

Multiple state-owned and joint-stock bank technology department staff told Shuhui their banks already use or plan to purchase programming Agents from Alibaba and Tencent. Such Agents mostly appear as embedded plugins, can browse code context, identify bugs, and assist in code generation.

It’s easy to understand why R&D side breakthroughs occur more readily.

Programming scenarios have clear chains and direct feedback; they are naturally easier to Agent-ify than financial business. The competition between OpenAI and Anthropic in the coding domain also shows code generation is one of the first directions for Agents to scale.

But R&D breakthroughs can’t be directly extrapolated to credit, wealth, risk control, and other core businesses. The latter involve real customers, fund flows, and division of responsibility—thresholds are much higher.

Beyond Scenarios

The business side is a test ground for capability: whether banks can allow Agents into real business processes depends on whether they already have the infrastructure to carry them.

Since DeepSeek’s breakout in 2025, banks’ willingness to deploy local models and explore Agents has clearly strengthened.

Shuhui noted that multiple state-owned big banks have started highlighting large model applications in annual reports.

“Scenarios” has become a high-frequency term in this process:

For example, in 2025, a certain state-owned bank has already pushed large models into over 500 scenarios; another major bank’s large model tech enabled 398 scenario applications, penetrating wealth management, inclusive finance, risk management, tech R&D, and more.

But hundreds of scenarios do not equate to mature Agent infrastructure.

Fintech staff Huang Yili (pseudonym) explained to Shuhui that some banks’ scenario counting is rather broad.

For example, the same speech optimization agent distributed to private banking, wealth centers, and customer service can be counted as N scenarios.

This may inflate the apparent penetration rate of large model applications.

A greater number of scenarios does not mean the model has richer capabilities, nor does it mean applications have entered core business chains;

Agent implementation requires more than just a few Q&A assistants, but connections among model, data, systems, permissions, processes, auditing, and evaluation mechanisms.

The more crucial data—usage frequency and activity—remain undisclosed. Compared to scenario numbers, these better reflect whether AI tools are truly accepted by frontline staff.

Huang Yili notes most banks, for data security, choose private deployment of models.

“But some managers wanting model privatization pick smaller models by parameter size,” she said. “Banks choose safety over usability, at the cost of decreased usage rates.”

Huang Yili revealed that in frontline pilot surveys, many employees said they rarely use in-house developed speech optimization Agents, finding their experience inferior to free external tools like Doubao or DeepSeek.

Thus, banks’ Agent progress remains slow.

Strategic statements are active; the business side remains cautious. The increase in scenario numbers still needs validation by actual experience.

All these gaps point to one issue: it’s no longer just model capability that limits Agents entering the business side.

The real difficulty is shifting from model deployment to organizational collaboration.

Organizational Dilemma

In the past, developing software products, banks had relatively clear paths: tech departments interfaced with tech vendors, researched business department needs, then analyzed, developed, delivered products, which business departments then used.

But Agents carry business judgment and experience—and these evolve constantly.

A fintech staffer summarized the difference as “maintenance” versus “operation”: traditional software is like maintenance, focusing on stable operation; Agent is like operation, emphasizing continuous feedback, training, and calibration.

They pointed out the cost investment phases differ between traditional software and Agents:

The former’s costs are concentrated in construction, with later costs mainly for maintenance and fixes; the latter can quickly develop prototypes, but true maturity depends on continual business feedback, user training, and iteration.

This person summarized: “Good Agent development requires business, product, and R&D to sit together from day 0.”

But some frontline bank staff’s resistance to Agents reflects practical difficulties most directly.

Huang Yili said, “When business departments lead and collect requirements, some employees’ enthusiasm is low. Tech departments are eager to design Agents, but returned questionnaires describe needs in only a few words.”

This feedback is not just about tech acceptance, but like a defensive response brought by job-value reassessment.

Agents don’t just ease client managers’ burden—they also rewrite how client managers prove their work’s value.

The observation of a joint-stock bank tech department manager gives a more specific answer.

The manager noted that, after introducing retail multi-agent systems, client managers’ time spent on information sorting and policy searching dropped significantly; system-generated asset allocation drafts reduced repetitive paperwork.

But behind efficiency gains lurks deeper organizational challenges—some staff welcome the convenience, while others worry about substitution of their own value;

More fundamentally, as systems take over information integration and basic plan generation, client managers’ core value should shift to review, optimization, and relationship maintenance—but many banks' assessment metrics haven’t adjusted accordingly.

“Existing bank training systems and KPI assessments still require phone volume and report volume without adjustment,” the manager said. “This leaves employees with idle time but unsure where to focus effort or unwilling to change established habits.”

The manager said the company is designing new “human-machine collaboration” training programs and assessment metrics, but finds it difficult to do so in isolation.

This means Agent implementation is not just about technical rollout—it also touches job definitions, evaluation systems, and incentive mechanisms, beyond the tech department’s scope alone.

Replication Challenge

In this dilemma, collaborative innovation offers a possible breakthrough.

Jin Mei said their team co-created business-side Agents with a regional bank: the bank provided seasoned wealth advisors to judge AI answers and new employees to verify effects; both sides polished with simulated data before on-site deployment.

In her view, financial business feedback chains are long; to let Agents iterate and form a rapid closed loop, business and tech must interact frequently.

But even when co-creation succeeds, Agents face new dilemmas: difficulty replicating.

Shuhui found many small and medium banks want business-side Agents but organizational structures don’t support time-consuming co-creation—they most often ask: “Others’ Agents have launched, can I just buy them directly?”

The answer is often no.

A successful Agent is usually deeply embedded with a specific bank’s business rules, data structures, risk preferences, and permission system.

If it runs well in Bank A, it doesn’t mean it can directly migrate to Bank B;

Even within the same bank, copying from retail wealth to corporate lending often requires remodeling and re-validation.

A joint-stock bank tech manager calls this the “scenario wall.”

The manager said the company successfully applied a retail multi-agent architecture in wealth and fund scenarios, but faced difficulties when trying to copy it to corporate lending.

For example, the rules, documentation, and decision chains in corporate lending differ sharply from retail business: input materials switch from customer profiles to financial statements, contracts, and account flows; the system must also connect to core accounting, lending approvals, etc., in more complex flows;

Existing agents can hardly migrate directly; business logic and data models need rebuilding.

This manager believes banks still lack a “meta-language” to describe business capabilities across scenarios—an expression system to abstract different business processes into reusable modules.

This also explains why many third-party companies deliver Agents as project-based, system integration or custom services, not standard product sales like traditional SaaS.

A fintech staffer said their team once developed an Agent for business loans to help client managers assess merchant qualifications, but such capabilities are usually just submodules in the overall project, not sold separately.

For banks, true value isn’t a generic Agent shell—but long-term adaptation to specific business constraints.

But this mode naturally limits scale.

Large banks have resources for co-creation but long process chains and high cross-department collaboration costs; small and medium banks have urgent needs but may lack organization and budget for complex co-creation. As a result, Agents are misaligned at both ends of supply and demand.

To truly break through the scale dilemma, industry-wide joint effort is needed.

Gu Shu pointed out that definitions and standards for Agents are not unified in the industry, making it hard to scientifically assess each institution’s application level.

Based on this, he suggested first establishing “standard components”: make single-function, fixed-process Agents into standard products, and focus on developing agents with autonomous planning and decision capabilities.

Only by standardizing fixed-process, low-risk abilities—query, summarization, reminders, document verification—can Agents move from stand-alone projects to reusable modules.

Governance First

Collaborative innovation can change working methods; industry standards can promote scale—but the premise for all this is: Agents must be effectively governed.

Currently, financial institutions lack a unified answer to this question.

When a transaction is executed after Agent recommendation and human confirmation, and a dispute arises, how is responsibility divided among model suggestion, human confirmation, and business approval?

If regulators query decision reasoning, can banks reconstruct the Agent’s inputs, outputs, tool calls, and human intervention in audit logs?

This is not just a technical issue, but a legal and governance one; the industry lacks a unified decision audit standard for Agents that enables dialogue between compliance and tech departments.

Gu Shu pointed out at the forum that financial applications of Agents may pose risks of “black box, hallucinations, autonomous decisions,” requiring governance strategies tailored to each.

For banks, whether Agents can enter core business ultimately depends on whether the governance framework is established first.

Jin Mei emphasized this framework includes at least four things:

First is permission boundaries: clarify what the Agent can query, call, and whether it can trigger transactions or approvals.

Second is responsibility boundaries: clarify how responsibility is divided among AI suggestion, human confirmation, and business approval.

Third is audit boundaries: record Agent input, output, call chains, and human intervention points, ensuring key processes are traceable.

Fourth is evaluation boundaries: establish acceptance standards suited to different scenarios—answer accuracy is insufficient to judge Agent effectiveness.

Thus, governance is not just a compliance framework, but will also drive adjustments in roles, assessment, and incentive mechanisms.

In other words, governance is not to restrict Agents, but to qualify them to enter the business side.

For banks, the true watershed for Agents isn’t at press conferences, but in real business processes.

Only when Agents can be authorized, audited, held responsible, and continuously trained and used by frontline staff, do they truly enter bank business.

Risk Disclaimer and Exclusion ClauseThe market has risks, investment must be cautious. This article does not constitute personal investment advice, nor does it take into account individual users' particular investment goals, financial situation, or needs. Users should consider whether any views, opinions, or conclusions herein fit their own specific circumstances. Investment based on this is at the user's sole responsibility. ```