The financial industry confronts an "intelligent paradox": leading institutions pour billions into developing large models, while smaller players rush to adopt AI tools. Yet, despite the substantial investment, these efforts struggle to make a significant impact on core businesses. The "high investment, low penetration" dilemma stands as a major obstacle to industry transformation.

Behind this dilemma lies a profound mismatch between general large models and the unique characteristics of the financial industry. The industry's risk descriptions heavily rely on rules and regulatory requirements, and business delivery boundaries are often unclear. These factors inherently limit the application scope of large models; the need for full traceability under strict supervision slows down the pace of change; and the industry's tendency to "reinvent the wheel" hinders small and medium-sized institutions from sharing technological benefits. These three barriers have erected a high wall for AI penetration.

As technology hurtles towards reality, the intelligent transformation of the financial industry desperately needs a new pivot. Enter financial agents, which can achieve the closed loop of "perception-reasoning-planning-execution-evolution." With precise scenario adaptation, manageable compliance paths, and collaborative application models, they are poised to dismantle the dilemma, ushering in a new era where the financial industry transitions from an "investment quagmire" to a "value leap."

To illuminate the application and landing scenarios of financial agents, AntChain and 01Caijing jointly released the "2025 In-depth Application Report on Financial Agents" in mid-to-late July (click "Read the original text" at the end of this article to download the full report). Additionally, they launched the "Penetrating Scenarios, Solving Problems with Intelligence - Breaking the Trap of 'High Investment, Low Penetration' in Financial AI" large model in-depth application live broadcast week (replay available through the 01Caijing video channel). The focus is on transformation pain points, agent development, AI reshaping wealth management, AI enhancing risk control and compliance, and technical credibility and scenario adaptation. It systematically dissects the path from concept verification to large-scale application of financial AI, offering clear guidance for the industry to overcome its challenges.

01 In-depth Diagnosis: Why Are Financial Large Models "Praised but Not Popular"?

The financial industry has a solid foundation for AI applications, with large models currently a significant trend. Ma Zhenxiong, head of Agentar at AntChain, notes that financial institutions are advancing intelligence on multiple fronts, manifested in expanding computational power and breaking through scenario innovation, encompassing process efficiency improvements, interaction experience upgrades, operations and maintenance optimization, automation of non-core businesses, and even cautious explorations of AI in core businesses.

However, this high-intensity investment and technological iteration have yet to significantly translate into users' perceived changes in financial services. In other words, the application of financial large models often garners praise but fails to gain widespread adoption, falling into the "high investment, low penetration" trap.

During the recent live broadcast week on the in-depth application of large models, jointly launched by 01Think Tank and AntChain, relevant academic and practical experts highlighted three key issues:

Insufficient industry adaptability. Zhang Ning, director of the Fintech Research Center at the Central University of Finance and Economics, revealed the primary constraints: the financial industry's risk descriptions rely on rules and supervision, featuring "non-pure probability" (including "punitive probability results"); task "delivery boundary clarity" is insufficient; and "language value density" only partially meets standards. When only one of these three criteria is met, it naturally compresses the application space of large models.

Zhang Ning also pointed out that financial large models currently exhibit two levels of application: one is "centralized application" at the institutional level, focusing on self-built infrastructure and internal process optimization (e.g., integration into OA systems); the other is "distributed application" closer to users, where frontline personnel like financial advisors and insurance agents utilize large models to reach customers' "nerve endings." The latter develops faster and is more prevalent, with frontline personnel and customers using them more frequently than within institutions. Some customers even use these tools to enhance their professionalism, thereby exerting a "whipping" effect on financial institutions.

Strict supervision forms a compliance barrier. For instance, financial customer service needs to be fully recorded and archived, with all customer interactions traceable to prevent errors. While the compliance value is significant, Chen Kani, director of the Digital Finance Laboratory at the Hong Kong University of Science and Technology, noted that these requirements objectively slow down the pace of change. Thus, model optimization should aim to effectively identify risks while enhancing the user experience for most people.

Industry collaboration shortcomings. Chen Kani further pointed out that each bank operates independently, akin to "reinventing the wheel." Unlike the medical industry, which openly shares experiences (including lessons), the financial industry lacks collaboration due to data privacy restrictions, making it challenging for small and medium-sized institutions to reuse large banks' model resources, thereby limiting the scale effect of AI applications across the industry.

Faced with multiple challenges, including insufficient industry adaptability, stringent regulatory constraints, and collaboration shortcomings, AI investment in the financial industry urgently needs to find new directions and paths.

02 Choice and Turn: Adjusting the Path for AI Investment in the Financial Industry

Amid the "high investment, low penetration" dilemma, banks, securities, insurance, funds, and other institutions face strategic tests for AI investment: the surge in leading institutions deploying self-developed large models persists, while smaller institutions, eager to try, find themselves in a dilemma – following suit risks unclear ROI and resource pressure, while waiting may fear missing out on strategic opportunities due to technological gaps.

In this regard, Ma Zhenxiong believes that the financial industry's AI investment will continue but will shift focus to scenario innovation. He emphasizes that building a financial large model is not only necessary but also involves manageable investment.

The AI investment of financial institutions has quietly shifted. Over the past two years, it was in the "infrastructure phase," focusing on building basic capabilities like computational power and platforms. Nowadays, it increasingly emphasizes "scenario innovation," following a gradual path from easy to difficult: starting with simple Q&A, internal office tasks, and non-core businesses, gradually extending to improving R&D and operations and maintenance efficiency, and ultimately aiming for business process reconstruction and core business transformation for customers. This is a process requiring gradual progress but with a firm goal.

The unique nature of financial scenarios underscores the necessity of building industry-specific large models, primarily reflected in the following four core capability requirements:

(1) Precise intent recognition and routing: Accurately understanding the financial business area and scenario to which a question belongs, a task challenging for general models;

(2) Reliable tool/knowledge invocation: Core applications require the model to stably and accurately recall tools and professional knowledge to avoid divergence;

(3) Accurate financial entity recognition: Extracting key entity information from financial texts with precision;

(4) Compliant professional expression: Ensuring answers are rigorous, compliant, and well-structured, reflecting financial professionalism (e.g., avoiding individual stock recommendations, bearish market forecasts, etc.).

These four "financial-grade capabilities" cannot be achieved through general models and must be trained into industry-specific large models using specific data atop the base model. This is the only path for financial AI to take root.

The key to manageable investment lies in the choice of technical path: there's no need for costly pre-training of massive models from scratch. A more feasible solution is to infuse financial-specific capabilities through post-training based on mature open-source base models (like Tongyi Qianwen). The required resources primarily focus on human resources, high-quality samples, medium-scale computational power (white card level), and fine-tuning platform capabilities during the post-training phase, effectively avoiding exorbitant pre-training costs and keeping overall investment manageable.

Zhang Ning also supports the financial industry's continued investment in AI. He suggests that large and medium-sized financial institutions should firmly anchor technological advantages in real business needs to achieve demand-driven technology implementation, striving to transform large model capabilities into their core competitiveness and brand advantages to build core competitive barriers.

For small and medium-sized financial institutions, it's crucial to carefully weigh the input-output ratio and adopt a phased investment strategy, scientifically dividing investment into different stages. It's also important to avoid spreading resources thinly across all scenarios and instead focus on deeply penetrating advantageous scenarios, selecting breakthroughs based on differentiated needs to achieve precise penetration.

After clarifying that investment needs to tilt towards scenario innovation and financial-grade core capabilities, "agents," regarded as the next technological evolution focus, have begun to enter the industry's vision, expected to become a key force in overcoming the dilemma.

03 Hope for Breakthrough: The Unique Value of Financial Agents

At the main industry development forum in July 2024, Jing Xiandong, Chairman and CEO of Ant Group, stated that professional agents can address the key issues of general large models in rigorous industrial applications.

Compared to pure large models, the core breakthrough of financial agents lies in constructing a closed-loop mechanism of "perception-reasoning-planning-execution-evolution." This mechanism drives agents to leap from "passive Q&A" to "active decision-making." Their capabilities can be broken down into three stages:

In the planning stage, agents accurately analyze the business scenario to which a question belongs and dynamically plan the tools and knowledge to be invoked (e.g., customer position data, market indicators) to avoid result drift;

In the execution stage, they call existing systems through tool interfaces to ensure the certainty of index query and knowledge recall;

In the expression stage, they output rigorous, compliant, and illustrated conclusions, strictly adhering to financial professional norms of "viewpoint first, reasonable and evidenced."

Due to these features, agents exhibit unique advantages in addressing many issues faced by large models in financial industry applications.

Regarding insufficient industry adaptability, agents can accurately understand financial business areas and scenarios. For instance, in the planning stage, AntChain's agents analyze scenarios and plan tool and knowledge invocations to avoid result drift. Their core applications stably recall professional knowledge, and the execution stage ensures information certainty through interface calls to existing systems. They can also accurately extract financial entity information, and the expression stage outputs compliant and professional conclusions, overcoming large models' limitations.

Faced with the compliance barrier posed by strict supervision, agents ensure "using the right knowledge" rather than "self-expression" by enhancing retrieval and compliance capabilities, building a precise retrieval system, and combining multiple technologies to make the reasoning process transparent and traceable. Simultaneously, a dynamic optimization mechanism is established to supplement compliance samples and evaluate attribution, ensuring compliant and accurate results.

However, agents' current capabilities are limited. Chen Kani believes the bottleneck lies in agents' base large models, whose capabilities are constrained by compliance costs and iteration speed. There are also challenges in localized deployment. Due to confidentiality requirements, financial institutions can only use open-source models and cannot call them externally. The deployment cost of "full-blooded" models (like deepseek) is high (reaching nearly ten million yuan and requiring continuous maintenance investment), facing the risk of technological iteration and easily falling into an "innovation dilemma."

The "2025 In-depth Application Report on Financial Agents," jointly released by AntChain and 01Think Tank, also noted that currently, agents' financial business scenarios are mainly "financial business RPA (Robotic Process Automation)," automating rule-based tasks in customer service, risk control, and other scenarios. However, due to base model capability and thinking chain technology limitations, they have yet to achieve fully autonomous work.

Despite technological advancements and decreasing costs of computing power, the collaborative deployment of multiple agents in intricate scenarios, such as investment research analysis and quantitative trading, is gradually becoming a reality. These agents are evolving from single-purpose tools into intelligent solutions that cater to the entire business chain.

Several industry leaders have embarked on exploring tailored solutions. AntChain's "Trusted Agent" stands out as a quintessential example, offering valuable insights into the implementation of agent technology.

04 Practical Exploration: AntChain's "Trusted Agent" Solution

In financial services, every numerical figure represents real wealth and trust, making the industry inherently intolerant of errors. Even a deviation of one in ten thousand can result in irreparable losses.

Acknowledging this industry characteristic, Qi Xiang, head of AntChain's large model and agent algorithm team, proposed the "Trusted Agent" framework, built on three pillars: reliable supply, controllable processes, and optimizable results. This framework ensures a robust guarantee system.

In terms of reliable supply, knowledge engineering serves as a cornerstone. Financial knowledge often resides in PDF reports, PPT charts, and obscure API comments. AntChain employs a suite of "intelligent disassembly techniques" to bring this knowledge to life:

150 bank documents are automatically processed into 20,000 precise FAQs, covering 50% of business traffic. Database comments are transformed into user-friendly "instructions," enhancing Chat BI data retrieval accuracy by 4-5 percentage points. Even dormant API interfaces are revived and put into use, with 20% of them reactivated through offline agent debugging.

For controllable processes, AntChain strengthens retrieval and compliance capabilities. Financial decisions must be governed by rules, and every step of an agent's reasoning must adhere to these rules. AntChain has developed an industrial-grade retrieval system that covers over ten links, including question understanding, multi-path recall, and re-ranking, ensuring relevant knowledge recall within one second. In terms of knowledge compliance, reinforcement learning (e.g., preference alignment) encourages agents to prioritize externally retrieved knowledge, combined with SFT (supervised fine-tuning) + DPO (direct preference optimization) to meet financial compliance and rigor standards. SOP collaborative rules convert business processes (like enterprise query steps) into natural language SOPs, embedding them into the agent's reasoning process to boost answer accuracy by over 10 percentage points.

Regarding result optimization, AntChain focuses on quantitative evaluation and iteration. A dynamic optimization mechanism is in place to continuously enhance agent output quality. This includes building a synthetic data generation system to supplement rigorous and compliant training samples in the financial domain and designing agent-specific evaluation metrics, leveraging both manual annotation and automated tools for end-to-end effect evaluation and attribution analysis.

Based on the "Trusted Agent" architecture, AntChain has implemented 100 financial agent application scenarios. The "2025 In-depth Application Report on Financial Agents" (click "Read the original text" at the end of the article to download the full report) reveals that these scenarios span four major sectors: banking, securities, insurance, and general, comprehensively penetrating the entire business chain, including customer service, internal operations, risk management, and more.

For instance, in customer service scenarios, agents address traditional financial service pain points such as limited service hours, slow response times, and homogeneous experiences. They enhance efficiency and accessibility with 7x24-hour online service, reshaping the service experience and bolstering customer loyalty through deep personalization. The full-process assistant agent covers pre-sales consultation, product usage, after-sales operation and maintenance, and other crucial links.

In marketing and sales scenarios, agents leverage their robust data analysis and content generation capabilities to drive financial institutions from traditional scattershot marketing to "precision drip irrigation" intelligent marketing. Their core value lies in deeply understanding customers, achieving personalized outreach, and maximizing marketing ROI. For example, the marketing agent boosts efficiency by 20% in pilot scenarios, while the financial Q&A agent provides C-end users with zero-illusion fund interpretation and wealth allocation advice.

In risk management scenarios, agents process and analyze more complex and voluminous data than traditional rule engines, enabling earlier risk identification and precise early warning. This facilitates the transition from "post-remediation" risk control to "in-process interception + pre-prevention." For example, the risk control modeling agent automates machine learning modeling, improving model discrimination (KS value) while reducing human costs for long-tail needs.

Financial agents offer a path of "controlled innovation" for "rigorous industries," fostering deep technology-business integration within compliance boundaries. In practice, agents have demonstrated the potential to break the "high investment, low penetration" dilemma, yet this journey still faces challenges like regulatory coordination, data governance, and knowledge precipitation.

To unleash their potential and overcome these challenges, addressing the core issue of technical credibility is paramount. Financial scenarios demand significantly higher standards for AI models than general fields, necessitating a comprehensive security system spanning the entire data lifecycle, with transparent and traceable algorithms. Only by transforming technical capabilities into "deployable and trustworthy" financial-grade productivity can we solidify the technological foundation for large-scale AI implementation, establishing a crucial link from "controlled innovation" to "value realization".

-End-