On June 1st, at the GTC conference in Taipei, Jensen Huang summed up the next phase of the AI industry with a single sentence: "The era of practical intelligent agents has arrived." In his view, artificial intelligence has moved beyond the stage of simple question-and-answer large language models and entered the era of Agentic AI, capable of autonomous observation, reasoning, planning, and tool usage. In the past, we were used to opening software, clicking, and typing in the operating system; in the future, people will only need to express their intentions, and intelligent agents will use models, frameworks, memories, tools, and runtime to break down tasks, allocate resources, and complete results themselves. Will this impact banks? If so, the issue isn't whether banks should use large models, but a more fundamental one: If intelligent agents become the default way for people to access various services, will the entry point between banks and customers be redistributed? The ultimate goal of banking services is still people. No matter how technology changes, financial capabilities such as accounts, payments, credit, wealth management, and risk control must ultimately serve real customer needs. However, the path people take to reach a bank may no longer be taking out their phones, opening a bank app, and then searching for functions layer by layer. In the future, a user might only say to their AI agent, "Help me figure out how to allocate this month's funds." A business owner might only tell their financial AI agent, "I have a payment due next week; find me the lowest-cost short-term financing option." At that point, whether a bank remains in the customer's lap might depend on whether it can be understood, invoked, and trusted by the AI agent. AI agents are not chatbots; they are a new financial gateway. Understanding AI agents as "smarter customer service representatives" or "conversational financial advisors" underestimates the capabilities of AI. In his speech, Jensen Huang emphasized that intelligent agents are not a single, large model, but a complex system composed of an execution framework, working memory, long-term memory, and a toolchain. It doesn't take over a single interaction link, but rather the entire chain of "understanding needs—decomposing tasks—calling tools—completing results." What does this paradigm shift mean for banks? Over the past decade, the main battleground for the digital race in the banking industry has actually been the mobile app. Whoever has the more user-friendly interface, the shorter the process, the fewer clicks required for online account opening, and the richer the wealth management channels, is more likely to retain customers within their system. Mobile banking has almost become the "front desk" for banks to serve customers. However, if in the future users no longer actively open bank apps, but instead directly say to the intelligent agent on their phones: "Help me compare which credit card offers the best deal," "Should I use this month's spare cash to pay the mortgage or buy wealth management products?" "My company needs to pay for goods next week; are there any better short-term financing options?", the entry point between banks and customers will shift from bank apps to intelligent agents. Huang Renxun mentioned that in the first few months of 2026, the number of code commits on GitHub nearly tripled, as tens of millions of developers worldwide began to involve AI agents in writing, modifying, and debugging code. The same logic will also extend to financial scenarios. Today, programmers use AI agents as development assistants; tomorrow, individual customers may use them as family financial managers, and corporate CFOs as cash flow management assistants. At that point, a bank's competitors won't necessarily be just another bank's app, but rather the AI agent platform that users default to using. This presents a very real problem: if a bank's account, credit, payment, and wealth management capabilities cannot be recognized and utilized by these AI agents, the bank's position in the customer's decision-making chain may shift further back. Customers may no longer start from the bank's own entry point, but instead use AI agents to compare, filter, and judge. This is similar to how many users today rarely actively open certain service apps, but instead make decisions through search, maps, payment platforms, and lifestyle service platforms. The services still exist, but the entry point has changed, and so have the traffic and pricing power. Banks need to shift from "developing apps" to "creating callable financial capabilities." What should banks do if the entry point changes in the future? Should they serve intelligent agents? That would be going against the grain. Banks' ultimate customers are always people; intelligent agents are merely new intermediaries. What's truly worth considering is whether, when customers access the financial world through intelligent agents, the core capabilities of banks can be securely, compliantly, and efficiently embedded into this new pathway. Jensen Huang made a crucial judgment in his speech: software companies won't disappear because of intelligent agents; on the contrary, they will see new opportunities as intelligent agents extensively utilize tools. The prerequisite is that the software must be presented in a way that intelligent agents can invoke. Applying this logic to the banking industry means that capabilities such as fund transfers, credit granting, wealth management, payments, foreign exchange hedging, and cash management cannot simply reside on product pages within an app. Instead, they must be encapsulated into standardized tools that are identifiable, callable, and auditable. Take wealth management as an example. In the future, users may no longer ask bank customer service, "How much return will this wealth management product yield?" but instead directly ask the intelligent agent, "What should I buy now?" The intelligent agent will then compare and recommend products across banks, products, and timeframes based on the user's income, expenses, liabilities, cash flow, risk appetite, and existing holdings. At this point, whether a bank can enter this decision-making chain depends on whether its product data, risk rating, fee information, redemption rules, and return performance can be accurately read and compared by the intelligent agent. This is precisely where the most sensitive issues lie. Suitability management, misleading sales, and liability attribution—every step treads a tightrope of regulatory scrutiny. What banks need to do is not simply let AI recommend products, but to establish an explainable, traceable, and accountable intelligent investment advisory process. The intelligent agent can participate in analysis and suggestions, but why the customer ultimately buys, whether the risks are clearly explained, whether the risk level is met, and whether a complete record is kept must all be subject to compliance. Of course, banks themselves are already deploying intelligent agents. From the various intelligent agents implemented by ICBC based on its "ICBC Smart Surge" large-scale model platform, to China Merchants Bank's DevAgent intelligent agent; from China Construction Bank's "AI application for the entire credit approval process," to Postal Savings Bank's "Postal Assistant" money market trading robot—intelligent agents are acting as "digital colleagues" within banks. These applications are important because they improve the efficiency of R&D, approval, transactions, customer service, and risk control. However, it's important to recognize that these still fall under the category of "internal efficiency improvement." The essence of "creating callable financial capabilities" is to decouple and encapsulate the bank's core business capabilities, transforming them into standardized modules that can be securely invoked by external intelligent agents. This is precisely the key to whether banks can remain firmly embedded at the bottom of the customer's decision-making chain after the entry point has been reconstructed by intelligent agents. The real hurdle lies at the foundation, not in the interface. For example, if a bank wants to be "called by an intelligent agent," the foundation for this isn't the front-end chat window, but the construction of a complete intelligent agent infrastructure. The first hurdle is data. In his speech, Jensen Huang discussed the memory system of intelligent agents, saying, "It processes short-term memory, known as working memory, as well as long-term memory, just like humans. Therefore, a memory management system is extremely important." For banks, this means that customer data, transaction data, product data, risk data, and external data must break down departmental silos to achieve unified governance, access control, and real-time access. If data remains scattered across core systems, credit systems, CRM systems, and various Excel spreadsheets, intelligent agents cannot form effective long-term memory, and thus cannot provide continuous and comprehensive financial planning for customers. The second hurdle is security and access control. Banks cannot allow AI agents to directly complete all high-risk transactions for customers. They need a layered authorization mechanism: balance inquiries can be automated, bill processing can be automated, transfers must be subject to secondary confirmation, investment transactions must undergo suitability verification, and large-scale financing and complex transactions must undergo manual review. AI agents need secure containers and frameworks to operate. Banks also need their own financial-grade security sandbox. It doesn't reject intelligent agents from entering, but rather ensures that intelligent agents act within controlled boundaries, and that every step can be audited, traced back, and held accountable. The third hurdle is computing power and cost structure. Huang Renxun repeatedly states, "Computing power equals revenue," and "Tokens are profit units," essentially saying that the cost logic has changed in the era of intelligent agents. For banks, AI is no longer a one-time annual technology budget investment, but rather a production system that continuously consumes computing power, models, data, and security resources. Every customer inquiry, product comparison, risk control analysis, investment advisory advice, and contract review consumes tokens. The competition among banks is no longer about "having AI," but about whether they can support more frequent, stable, and secure intelligent agent calls with lower computing power costs. The final hurdle is model governance and accountability boundaries. Who is responsible if the intelligent agent's suggestions are wrong? Who bears the risk of misjudgment? Where are the boundaries of customer authorization? How is model output recorded? How do regulators inspect it? There are no ready-made answers to these questions, but the answers determine whether a bank's intelligent agent can move from the demo room to the production environment. After all, a bank is not an ordinary software company; behind every "tool call" lies real funds, real credit, and real risk. The interface can be created quickly; the underlying structure is the real dividing line. In his speech, Jensen Huang said, "In the past, we were used to launching applications, clicking, and typing. Now, instead, we explain our needs and intentions to AI, which generates code or uses tools to produce the necessary results. This is how computers will work in the future; this is Agentic AI." For the banking industry, it's not about "banks serving intelligent agents," but rather about banks finding new ways to serve people in a world dominated by intelligent agents. Banks will not disappear. The essence of finance remains: credit intermediation, risk pricing, capital allocation, and payment clearing all require licensed institutions and cannot be replaced by any intelligent agent out of thin air. However, the way banks interact with their customers will be rewritten. And while intelligent agents are not customers of banks, they may become the gateway for customers to reach banks in the future.
Weatherly
Kikyo
Catherine
Anais