Finance stands as the pinnacle of a 'trust system' within human commercial society. The access and permission divisions, operational audit mechanisms, and circuit breaker systems honed in the most rigorous environments collectively form a stringent set of industrial-grade standards for agents.

Looking ahead, these mechanisms could seamlessly extend to areas such as dispatch and billing in autonomous driving, automated supply chain procurement in smart manufacturing, and even fund allocation in intelligent governance. From this vantage point, all the challenges and regulations established by the financial sector in deploying agents today are essentially forging a 'trust template' for agents in the entire intelligent commercial landscape of tomorrow.

Author | Dou Dou

Editor | Pi Ye

Produced by | Industry Insight

Since 2026, the financial sector has emerged as one of the most dynamic industries for agent deployment.

In recent months, large model vendors, cloud service providers, and fintech companies have simultaneously propelled finance to the forefront of agent implementation.

In June, Alibaba Cloud unveiled its financial-grade, general-purpose intelligent agent platform, 'Dianjin', at the China International Financial Exhibition. This platform focuses on task execution across investment research, advisory services, credit management, risk control, claims processing, marketing, and customer service—going beyond mere question-and-answer and summarization capabilities. That same month, ByteDance launched Kouzi Coze 3.0, packaging industry-specific capabilities in finance, law, and healthcare into callable professional skills. Subsequently, institutions like GF Securities and Guosen Securities introduced financial skills on the Kouzi platform, transforming capabilities such as stock selection, fund selection, market analysis, financial reporting, and macroeconomic analysis into tools directly usable by agents.

By August, Baidu Wenku and Baidu Netdisk jointly introduced Kuku AI, revealing its Chinese name and prioritizing financial investment research as a key office scenario. The product integrates financial data such as listed company information, stock quotes, financial reports, and research reports, catering to the production of Word-based research reports, financial analysis PPTs, Excel financial models, and other outputs closely tied to the daily work of financial practitioners.

September witnessed even denser activity. Tencent WorkBuddy Financial Edition was officially released, offering four specialized workstations for corporate finance, retail finance, investment research and advisory, and individual customer management, tailored to the needs of banks, asset management, and insurance industries. Alibaba's Qianwen Open Platform integrated over a dozen financial intelligent agents covering securities investment, funds, futures, and insurance. Ant Digital also launched Agentar Financial Edition, providing financial intelligent agent expert teams, industry skills, MCPs (Multi-Channel Platforms), evaluation tools, and governance capabilities.

Collectively, these moves indicate that financial agents are evolving from a showcase of technical prowess by major firms into an independent industrial sector.

On the demand side, similar dramatic changes are underway. Most notably, the proportion of AI application procurement is rising, no longer confined to customer service, knowledge bases, and office support but shifting toward decision-making and auditing.

This trend is evident in bidding market data. Compared to the first half of 2025, the number of large model projects awarded in the financial sector in H1 2026 grew by 110%, with disclosed award amounts increasing by 82.8%.

Among these, application-type projects accounted for 61%, with intelligent customer service & digital humans, intelligent auditing & analytical decision-making, and knowledge Q&A & knowledge platforms ranking as the top three by project count (60, 39, and 23 projects, respectively), collectively surpassing computing power projects. In contrast, computing power projects slightly outnumbered application projects in H1 2025.

Notably, in September, the financial sector's first intelligent agent safety-related standard was officially released, providing clearer guidance for financial agent applications and sparking serious industry discussions on core issues like permissions, responsibilities, auditing, and risk boundaries.

Yet beyond the excitement, pressing questions remain: Why finance? How deeply have financial agents penetrated business processes? And what will truly change in finance when AI gains system-calling and execution capabilities?

I. Finance: The Most Valuable 'Agent Industry Blueprint'

A Gartner forecast predicts that by the end of 2027, over 40% of Agentic AI projects will be canceled due to escalating costs, unclear commercial value, or inadequate risk control. It also notes that many so-called agents on the market are merely repackaged AI assistants, RPA (Robotic Process Automation), and chatbots lacking true autonomous task execution capabilities.

Despite 2026 being widely hailed as the 'Year of the Agent,' 'acclaim without adoption' remains the norm in real-world industrial deployment. Ultimately, the longevity of the agent narrative hinges on its ability to transition from demo scenarios into core enterprise business flows.

Finance provides the perfect testing ground for this transition.

First, financial operations are inherently data- and rule-driven. Financial reports, announcements, market data, research reports, business records, credit records, insurance policies, transaction logs, and approval rules have long been entrenched in financial systems. Tasks like credit assessment, investment research, claims processing, and risk control all require searching, comparing, summarizing, and making preliminary judgments across vast materials—precisely the strengths of agents, which excel at aggregating dispersed information and driving next steps along business processes.

Second, financial institutions possess both the budget and motivation to sustain AI investments. In 2025, China's financial AI full-stack cloud market reached RMB 20.76 billion, up 50.0% year-on-year, outpacing the 23.9% growth of the broader financial cloud market. Rising bidding volumes and values indicate a shift from pilot projects to centralized procurement among financial institutions.

More pressing is the internal operational pressure faced by financial firms. In Q2 2026, commercial banks' net interest margin stood at 1.41%, still low, with cumulative net profits declining by 0.6% year-on-year in H1. With narrowing interest spreads, the number of clients and transactions handled by relationship managers, risk controllers, and operations staff directly impacts operational efficiency.

Securities firms face similar challenges. In H1 2026, the average brokerage commission rate for listed securities firms dropped to 0.027%, down from 0.031% in 2025, forcing these institutions to seek growth in investment research services, wealth management, and customer engagement.

Against this backdrop, the value of agents becomes highly targeted. They can help banks reduce due diligence and approval times, enable securities firms to expand research coverage, improve insurance claim audit efficiency, and enhance wealth managers' service reach.

For tech vendors, finance is equally indispensable.

Over the past two years, large model and cloud providers have invested heavily in computing power and R&D. Now, facing significant depreciation pressure on their computing infrastructure, these tech firms must transition from selling 'Tokens' to selling 'digital employee seats.' Across industries, finance offers the most abundant IT budgets and stable willingness to pay—a critical 'blood source.' More importantly, financial business logic is the most complex, and compliance thresholds the highest. Once multi-agent coordination, trustworthy reasoning, private deployment, and security governance are proven in finance, the solutions can be adapted to other sectors.

Thus, the tech industry's rush into finance is no accident. For financial institutions, it's an efficiency lifeline; for tech vendors, a dual test of commercialization and engineering capabilities.

II. Self-Built, Contextual, Foundational: The 'Wild' Growth of Financial Agents

Beneath the buzz of tech firms flocking to finance, different players are pursuing starkly divergent strategies.

First are state-owned and leading banks, which handle the most sensitive customer, account, and transaction data while shouldering the highest compliance burdens. Rather than procuring standalone tools, they prefer building their own foundations, centralizing management, and then extending capabilities to business lines.

Take ICBC as an example. Its 2025 annual report disclosed the 'Gongyin Zhiyong' large model technology system, integrating over a dozen mainstream models and conducting in-house secondary training to create enterprise-level foundation models better suited to finance and ICBC. The bank also built an intelligent agent creation platform to enable rapid agent development by employees. To date, 'Gongyin Zhiyong' has been deployed in over 600 scenarios, with intelligent inquiry-based trading in financial markets exceeding 96% penetration and audit efficiency quintupling compared to pre-adoption levels.

However, this approach comes at a high cost. CMB's H1 2026 interim report revealed RMB 4.679 billion in IT spending (2.99% of revenue), with over 10,000 R&D staff (8.98% of total employees)—approaching the investment intensity of a leading tech firm. Heavy investment, systematic infrastructure, and self-reliance define this path.

Unlike banks, securities firms lean toward packaging professional capabilities into callable modules.

For firms like Huatai, CICC, and CITIC Securities, core assets have long been research, trading, and client service capabilities. Historically, these relied heavily on analysts', traders', and advisors' personal expertise; now, firms are codifying this experience into skills, MCP tools, and investment research agents, enabling AI to recognize, invoke, and reuse them.

For instance, CICC's 'Dianjing MCP' integrates CICC-specific data, models, financial market data, and news into an AI assistant, expanding its analyst skill matrix and research tool library. CITIC Securities' intelligent investment research service encapsulates its research team's methodologies into skill packages, with its inaugural non-bank financial analyst skill derived from research frameworks refined from 40 in-depth reports.

This shift reflects evolving competitive dynamics in securities. Previously, firms competed on researcher headcount, industry coverage, and client reach; now, the race is to deposit research methods, data assets, and trading tools into machine-callable capability systems.

Further down the chain are traditional financial IT vendors like Hengsheng Electronics, Digital China, Yusys Technologies, and Zhongguancun Science and Technology.

These firms may lack the strongest foundation models but possess critical system interfaces, business process understanding, and private deployment experience essential for financial agents to operate in real-world settings. For agents to evolve from 'advisory' to 'executive,' they must interact with account, trading, credit, and risk control systems—long maintained by financial IT vendors, forming their competitive moat. To some extent, these vendors solve the challenge of 'how agents grow limbs.'

General-purpose tech giants like Tencent, Alibaba, Baidu, and ByteDance excel at providing foundation models, cloud infrastructure, agent development platforms, and ecosystem capabilities, packaging models, tools, workflows, MCPs, and security governance into platforms. However, the true difficulty of AI adoption in finance lies not in models but in private data, core systems, compliance workflows, and business accountability. While big tech can supply powerful foundations and toolchains, they struggle to dominate all financial scenarios alone. To penetrate deep into business processes, they must ally with financial institutions, data service providers, financial IT vendors, and ecosystem partners.

In summary, a monolithic dominance of financial agents is unlikely. Leading banks will insist on self-reliance, securities firms will skill-ify investment research and trading capabilities, financial IT vendors will connect core systems, and general-purpose tech giants will provide models, computing power, development platforms, and governance tools. Mature financial agent products will inevitably be ecosystem combinations of 'general-purpose model foundations + financial-specific data + industry workflows + core system interfaces + security governance.'

III. The Tipping Point for Financial Agents

How far have financial agents progressed?

Consider coverage and usage intensity. CCB disclosed at its 2025 results briefing that AI assistant coverage reached 99.42% in branch issue responses by end-2025, with over 100,000 daily visits. ICBC's annual report noted over 500 AI applications deployed across 30+ business areas, with AI digital employees handling work equivalent to 55,000 person-years. CMB's 2025 report mentioned AI saving 15.56 million labor hours in efficiency gains, while CITIC Bank disclosed over 1,700 intelligent service scenarios, boosting efficiency by over 17,000 person-years company-wide.

These figures show AI has transitioned from small-scale pilots to daily tools within banks, securities firms, funds, and insurers.

Externally, financial agents appear to have entered a full-scale eruption phase.

Yet digging deeper into business operations reveals that high AI coverage ≠ true financial agent proliferation. The most mature areas remain low-risk tasks like knowledge retrieval, document processing, and content generation. AI primarily handles repetitive tasks such as querying regulations, summarizing minutes, and drafting reports to enhance information processing efficiency.

Strictly speaking, this resembles Copilots more than autonomous agents.

The subsequent phase centers on process-transforming Agents, where AI transcends textual processing to infiltrate specific business process stages. Take BOCOM's enterprise-grade general-purpose agent platform as an example; it has utilized AI to Refactor (restructure) processes such as account opening, credit approval, authorization, and international settlement. By the first half of 2026, this process-centric deployment had advanced further: in cross-border finance, its shipping trade review agent processed over 500,000 transactions daily, enhancing efficiency by more than 70%. In financial markets, agents handling intelligent foreign exchange inquiries and money market pricing executed transactions totaling over RMB 1 trillion cumulatively.

Parallel transformations are unfolding in the insurance and asset management sectors. In claims processing, AI can analyze image materials, interpret policy terms, compare medical records, and identify missing documents or preliminary audit issues. Regarding public fund investment research, Zhongou Fund's investment research agent integrates over 100 Skills, encompassing data scraping, industry analysis, financial modeling, and report writing, thereby boosting analysts' information retrieval efficiency by over 60%.

At this juncture, Agents commence orchestrating cross-system, cross-role operations, completing the entire cycle from data reading and tool invocation to judgment generation, succeeded by human review.

However, achieving this milestone remains rare. Big data underscores this point: according to the China Large Model Project Monitoring and Insight Report, among 47 financial large model projects involving intelligent agents in the first half of 2026, they constituted approximately 11.6% of all financial large model awards. In contrast, agent-related projects accounted for roughly 16.7% of large model awards across all industries—falling below the industry average.

Why is this the case? The financial sector exhibits zero tolerance for errors. Currently, Agents operate solely in scenarios with well-defined rules and straightforward review processes. They are adept at gathering information and formulating strategies but are never entrusted with independently granting credit, pricing, or transferring funds.

Presently, financial Agents find themselves at a precarious tipping point: the auxiliary layer has expanded, the process layer is broadening, yet the core decision-making and autonomous execution layers remain stagnant. The common thread among deployed scenarios is stable data sources, easily reviewable outcomes, and comprehensive audit trails within business processes.

IV. AI Meets Funds: The Financial Abyss of Agents Beckons

As one of the few industries where AI's 'value' and 'errors' instantaneously translate into tangible financial outcomes, the commercial infrastructure of the Agent era is being prematurely unveiled as Agents commence linking with bank accounts, fund transfers, and trading instructions.

Reflecting on the technological evolution over the past three decades, finance has assumed vastly different roles across various cycles.

During the internet era, finance was positioned as a 'connector and channel.' Its primary mission was to transition physical branches onto computer and mobile screens, enhancing the efficiency of connecting individuals with their finances. In the cloud computing era, finance evolved into a 'business container and computing power purchaser.' Its foremost task was to support the elastic scaling required for high-concurrency, massive transaction volumes during events like Double 11, ensuring the stability and agility of the underlying IT infrastructure. Now, in the current AI era, finance is transforming into the 'credit and clearing foundation for an intelligent society.'

This evolution is entirely logical. As algorithms and AI agents increasingly handle operations autonomously, who will oversee accounting, identity verification, budget allocation, and risk boundary definition for these intangible intelligent entities? The answer remains finance.

In this process, as AI delves deeper into 'monetary affairs,' the fundamental operating logic of industries will undergo a qualitative shift.

Firstly, the delivery standard will evolve. Previously, when large models entered financial institutions, competition centered on benchmark scores, generation quality, and terminology accuracy. However, with AI agents participating in financial flows, institutions' primary concern shifts to, 'Who assumes responsibility in the event of a mishap?' This will compel vendors to transition from selling models or APIs to providing a comprehensive 'braking system,' encompassing permission controls, hard-rule interceptions, operation logs, audit trails, business circuit breakers, and even the advent of new liability insurance.

Secondly, interaction partners will transform. Financial products were traditionally designed around 'humans'—individuals viewed screens, made decisions, and confirmed actions, with risks mitigated through facial recognition, verification codes, and counter checks. However, as AI agents gain autonomous execution capabilities, the future logic of financial interactions will be entirely upended. Banks may engage with financial agents embedded in a company's ERP system, which automatically applies for supply chain loans when inventory dwindles. In secondary markets, buy-side research agents may directly interface with sell-side real-time data, swiftly completing logical verification and order placement.

This implies that existing APP interfaces and manual review processes will recede into the background. Financial systems must be reconstructed from the ground up, incorporating high-frequency machine-to-machine interaction interfaces, security authentication frameworks, and real-time settlement channels.

More significantly, finance will export its risk management paradigms to a broader spectrum of industries.

Finance represents the pinnacle of the 'trust system' in human commercial society. If a multi-agent collaborative system can safely and flawlessly execute the entire loan approval, disbursement, and mortgage registration process without human intervention—while maintaining comprehensive audit trails—then the agent access controls, permission hierarchies, operational audits, and circuit breaker mechanisms honed under these stringent conditions will become universal industrial-grade standards.

In the future, these mechanisms could seamlessly apply to autonomous driving's scheduling and billing, smart manufacturing's automated supply chain procurement, and even intelligent government's fund allocations. From this vantage point, all the challenges and regulations established by the financial industry in deploying AI agents today are, in essence, crafting 'trust templates' for the entire intelligent commercial landscape of tomorrow.

Over the past two years, the entire industry has been preoccupied with computing power and large models, essentially addressing only the 'intelligence supply' facet of AI. However, the impending wave of financial AI agent deployments in 2026 confronts a far more critical and profound challenge: when silicon-based intelligence gains the authority to execute financial transactions, how exactly should humanity restrain it?

This critical examination at the vault door has only just commenced.