Agentic AI rewires financial infrastructure, competition and customer behaviour

Written by Hugh Zeng

Apr 1, 2026
8 min

Discussions at The Asian Banker Shanghai International AI Finance Summit focused on how emerging AI agents are reshaping the economics of infrastructure, the boundaries of competition and the way customers access financial services.

In March 2026, the Shanghai International AI Finance Summit convened a session on building an AI-driven financial ecosystem. Dong Longfei, senior vice president of Moore Threads, and Zhao Jianxin, general manager of Infrastructure Cloud Solutions at Jing Dong Cloud, were among the speakers. Their exchange came at a telling moment: within a 48-hour window only weeks earlier, ByteDance, Tencent, Kimi, Jing Dong and Zhipu had each launched products connected to the OpenClaw framework. The race to deploy autonomous agents had moved from concept to execution.

What makes this most significant is not the pace of product releases alone. It is the nature of the shift underway. AI is no longer functioning primarily as a conversational interface. It is becoming an execution layer, capable of breaking down complex tasks, calling external tools, coordinating across systems and collaborating with other agents. For financial institutions, that changes three things simultaneously: infrastructure costs, competitive structure and customer relationships.

Token economics are reshaping the cost base

The most immediate consequence of agentic AI is on infrastructure costs. As agents take on more complex, multi-step work, token consumption rises sharply. Jensen Huang, chief executive of NVIDIA, has said that agents performing complex tasks consume roughly 1,000 times more tokens than traditional generative models, and that continuously running agents can push that figure to one million times.

For banks, that transforms AI from a modest operational expense into a core and recurring cost. In the earlier phase of AI adoption, most institutions made limited API calls for customer service or internal support functions. Agentic AI operates differently. Once agents run continuously across multi-step workflows, inference becomes a standing operating expense. The commercial viability of large-scale deployment now depends on both hardware efficiency and model efficiency.

Zhao Jianxin identified four obstacles enterprises face when deploying AI agents: accuracy, stability, cost and task complexity. Rising token consumption makes each harder to manage. Jing Dong Cloud has responded by prioritising inference optimisation and recently open-sourced JoyAI-LLM Flash, an instruction-tuned model that uses one-quarter to one-fifth of the tokens required by comparable models to complete equivalent tasks. “For completing the same task, the tokens needed are the smallest,” Zhao said. That is not merely a technical boast. It is what makes large-scale agent deployment commercially plausible.

Dong Longfei made the complementary hardware argument. Computing costs will continue to fall, he said, but hard architecture must shift from “training-first” to “inference-first” designs. NVIDIA’s introduction of the Latency Processing Unit (LPU) addresses the high-frequency execution demands of agentic workloads. The infrastructure required to run agents at scale is evolving quickly, and the cost structure is changing with it.

The rise of domestic models reinforces this dynamic. Data from OpenRouter shows that Chinese domestic models accounted for 2% of token consumption at the end of 2024 and 39% by early 2026, while their API prices run at roughly one-sixth those of overseas alternatives. Two domestic models, Step and MiniMax M2.5, already account for about 50% of total token consumption. For cost-conscious financial institutions, this opens a practical route to agent deployment without unsustainable operating expenses.

Competition is moving from rivalry to symbiosis

The second change is structural. The near-simultaneous launch of OpenClaw-related products by ByteDance, Tencent, Kimi, Jing Dong and Zhipu was not simply a product race. Each institution was competing to secure the AI entry point, the traffic channel and the ecosystem position that will matter most in the next phase of digital finance.

Dong Longfei described the emerging relationship between banks and technology companies as one of “symbiosis” rather than competition. His point was straightforward: traditional banks must open their API interfaces if AI agents are to operate effectively within the financial ecosystem. “Agent represents machine-to-machine communication, not human-to-machine communication,” he said. “Machine-to-machine communication is essentially done through APIs, but all banking systems are closed. This is a crucial step for banks to move from closed systems to openness.”

He went further, suggesting that some banks may eventually become agent proxies rather than direct service providers, and that institutions most easily discoverable and accessible by agents may gain business even if customers never interact with their app or website directly. A closed institution risks becoming invisible to the systems that increasingly mediate financial decisions.

Zhao Jianxin added a technical dimension. Multi-agent frameworks are necessary for complex financial tasks, and long-term memory is essential for agent efficiency. Through scenario-specific fine-tuning, reinforcement learning and model distillation, Jing Dong has already deployed agent applications across a range of internal functions. For banks, the key point is that technology exists. The remaining barrier is the institutional willingness to expose core capabilities through well-governed APIs.

Lin Yonghua, vice president of Beijing Academy of Artificial Intelligence (BAAI), offered a useful reframe. Financial institutions do not need to compete on model scale, she argued. They should focus on building professional domain knowledge and skills that solve specific problems reliably. The global open-source community has produced hundreds of thousands of Skills (OpenClaw's modular, callable agent capabilities); what remains scarce are those that are certified, domain-specific and operationally dependable. That is where banks can establish genuine authority.

Customer behaviour is shifting from search to delegation

The third change is the most visible to end users. Agentic AI is altering how people discover, evaluate and transact financial services and doing so across three dimensions.

The first shift is the entry point. On-device agents are becoming the primary gateway for user intent. Xiaomi has embedded its proprietary AI agent, Xiaomi Miclaw, into phones, televisions and cars. Alibaba’s Qianwen integrates a transaction gateway that allows users to place orders and initiate services with a single instruction. Jing Dong has taken a different route with JoyInside embodied intelligence, which places “high-EQ brains” into smart appliances, robots, AI toys and other devices. Zhao Jianxin said Jing Dong does not charge hardware partners for the integration. “The core is to get the entry point in,” he said, so that services can reach users across a wide range of devices. JoyInside now works with nearly 100 home appliance and furniture brands and more than 40 robotics and AI toy brands.

For banks, the implication is that the customer relationship is no longer anchored exclusively in a bank-owned app or website. The primary interface may be a voice assistant, a connected device or a third-party agent. Institutions that make their services easy for these agents to invoke, through clean APIs and well-defined Skills, stay relevant. Those that do not risk being disintermediated.

The second dimension is consumption behaviour. Zhao Jianxin argued the future of AI is, on one hand, to deepen reasoning and task-completion capability, and on the other, to apply that capability in embodied intelligence for homes and factories. In finance, that means users will increasingly delegate decisions instead of making every choice directly. They will authorise agents to act within defined budgets, preferences and behavioural patterns. The business-to-agent (B2A) model is emerging as a distinct commercial category. Agents, as Zhao noted, make purchasing decisions based on data rather than advertising, making them more analytically rational than most human consumers.

This changes product design fundamentally. A mortgage offering is no longer a brochure or a landing page. It becomes a set of structured API endpoints that an agent can call to verify eligibility, calculate repayments and initiate an application. Pricing must be transparent and machine-readable. Customer acquisition depends less on brand visibility and more on being the service an agent selects on a user’s behalf.
The third dimension is transactional integrity. Dong Longfei raised a systemic concern: if 80% of financial institutions use the same underlying model for risk pricing, even a tiny hallucination could trigger a “terrifying one-way market flash crash”. His proposed safeguards include model diversity, AI circuit breakers, human-in-the-loop controls and privacy-preserving technologies. Zhao Jianxin identified multi-agent collaboration as essential for managing complexity in high-stakes scenarios.

For banks, this raises a fundamental question about relationship ownership. When an agent mediates every customer interaction, the institution holding the account may no longer control where business flows. The agent’s preferences, the model it runs on and the services it is configured to call will increasingly determine transaction outcomes. Maintaining relevance in that environment requires more than open APIs. It requires becoming the preferred, trusted service within the agent ecosystems customers already use.

This is not a future scenario; it is being built now through inference-optimised hardware, open APIs that allow agents to call bank services and a steady shift in user behaviour from active search to delegated action. For financial institutions, the strategic question is no longer whether agentic AI will matter. It is whether they are positioned to compete in an environment where the customer no longer searches and compares in the conventional sense, but delegates.

Keywords: