The 7th China Fintech Forum of the China International Fair for Trade in Services was held in Beijing from September 10-11, 2025, with the theme "Technology Empowerment - Digital Transformation and Application of the Financial Industry." Chen Liyu, Senior Vice President of BAIRONG-W, attended and delivered a speech.

The following is the speech transcript:

Distinguished guests, good afternoon. From the perspective of a fintech company, I would like to share with you our practice in banking intelligent transformation. When we engage in banking intelligent transformation, we need to clarify a fundamental question: What are the core capabilities of large language models? Where are their capability boundaries? What scenarios are they suitable for? What scenarios are they not well-suited for?

The industry currently focuses extensively on the relationship between large models and Artificial General Intelligence (AGI). By drawing analogies from the logic of human intelligence formation, we can better understand the core pathways. Humans acquire various information from text, video, and the real world through sensory organs such as vision (eyes) and hearing (ears) - this process is similar to large models' collection and input of multimodal data.

The deepening of human intelligence relies on the progressive transformation of "information - knowledge - wisdom": on one hand, it requires massive information accumulation through "reading thousands of books" to build fundamental data reserves; on the other hand, it needs to combine practical scenarios of "traveling thousands of miles" to refine scattered information into structured knowledge; ultimately, through continuous understanding, learning, and practical verification, knowledge deeply integrates with real-world needs to form wisdom with decision-making value. This closed loop from information input to wisdom output essentially aligns with the intelligent evolution direction pursued by large models and represents one of the core logics that AGI development must follow.

So, what are the specific manifestations of large models' core capabilities? Based on practical observations, they can be summarized into four points:

First, the widely recognized powerful generation capability, which forms the foundation for large models to create value in content creation and solution output scenarios.

Second, generalized language understanding capability. This understanding is not limited to fixed expressions in training data but can break through existing textual constraints to precisely capture core problems and logic behind expressions that have never directly appeared, thus achieving flexible adaptation to complex language scenarios.

Third, efficient few-shot learning and in-context learning capabilities. This capability is particularly evident in actual business operations: traditional Natural Language Processing (NLP) technology often requires over 1,000 sample data points for initial training when completing task decomposition and information extraction; leveraging large model technology, fewer than 100 samples can train NLP models that achieve over 80% accuracy in business scenarios, with continuous optimization of model understanding precision through subsequent data annotation and iteration.

Fourth, knowledge association and reasoning capabilities, but we must objectively recognize the essential differences from human reasoning: human reasoning is premised on logical understanding, forming judgments based on cognitive understanding of intrinsic relationships between things; while large models' "reasoning" is essentially still statistical probability prediction based on massive data - they achieve imitation and presentation of knowledge associations by learning implicit logical patterns in training data, rather than truly possessing human-like logical cognitive capabilities.

Previously, Chain of Thought (CoT) technology received widespread industry attention. From a technical essence perspective, chain of thought belongs to the category of prompt engineering, with its core logic being structured decomposition of complex problems, breaking down reasoning processes into clear intermediate steps, thereby guiding models to more precisely understand problem logic and organize analysis pathways, ultimately improving language models' understanding and processing capabilities for complex tasks.

The value of this technical approach lies in its ability to effectively compensate for models' limitations when directly addressing multi-step reasoning problems - by explicitly presenting reasoning stages, making the model's analysis process more logical and traceable, rather than relying on single inputs to directly generate results. This makes it better suited to actual needs in scenarios requiring deep logical deduction.

What I just discussed were some peripheral topics. Today's theme is how large models can create value in business scenarios. This brings up a crucial point that many guests have emphasized: this type of intelligentization requires substantial cost investment. How do we measure the value of such cost investment? Therefore, we propose that AI, our so-called AI Agents, should also complete KPIs like humans. In my view, banking intelligent transformation should focus on three core processes, building a complete pathway from business upgrading to value realization:

First, promoting AI-enabled business processes. Banking business systems consist of multiple specialized processes, with different processes corresponding to different departmental settings, and each department conducting work around specific job KPIs - essentially, each business process consists of a series of coherent business activities. Taking credit business as an example, from front-end customer acquisition and mid-term anti-fraud verification to credit approval and post-loan collection, each link requires specific personnel to execute. The core of business process AI-enablement is using artificial intelligence technology to replace or assist manual completion of these standardized, repetitive business activities, achieving improved process efficiency.

Second, implementing AI-enabled business process know-how. The business rules, risk judgment logic, customer service experience, and other tacit know-how that banks have accumulated during long-term operations constitute important components of their core competitiveness. Converting this professional experience into structured knowledge that AI models can learn, enabling models to understand business logic and make decision judgments like experienced practitioners, is key to deepening intelligent transformation - this requires not only algorithmic optimization at the technical level but also establishing the complete chain of "business experience extraction - knowledge graph construction - model training iteration."

Third, establishing standardized evaluation systems for AI value delivery. The application effects of AI technology cannot be measured solely by technical parameters; the core consideration is whether it can "deliver value at scale" - that is, whether AI solutions can stably implement in large-scale business scenarios and continuously reduce costs, increase efficiency, and control risks for banks.

For example, we are currently providing AI-driven post-loan collection solutions for multiple financial institutions. The background of this demand is closely related to the "three lows and one high" challenges currently faced by the banking industry: while market interest rates decline, spreads narrow, and overall returns face pressure, non-performing asset risks show certain upward trends influenced by macroeconomic conditions.

From industry leadership perspective, how do we control non-performing assets? Current banking practice involves outsourcing non-performing loans starting from M1 stage to external agents. When non-performing loans increase, do we need to add staff? Adding staff requires additional budget, which is a budget issue. Now that spreads are narrowing and returns declining, decreased returns mean reduced budgets for banks. How do we solve this problem under these circumstances?

We can only use silicon-based workforce. What capabilities must silicon-based workers possess?

First, human-like intelligence to precisely identify human emotions, understand customer intentions, and plan next-step strategies. Collection is extremely difficult, requiring pressure on customers while responding appropriately based on customer emotions. They must record key communication nodes with each customer and provide guidance for the next call, so intelligence requirements are very high.

Second, end-to-end interaction capability. Silicon-based workers cannot sound different from humans - they must have anthropomorphic qualities with latency controlled within 500 milliseconds, while incorporating collection expert capabilities. We know that collection currently involves about 2 million collection agents working nationwide, with annual salaries of approximately 150,000 yuan per person, representing a 300 billion yuan cost market, with significant portions in the M1 stage.

Bairong first achieved AI-enabled business processes by creating 9 major intelligent agents with refined division of labor and collaborative operations.

The front office features collection agents directly facing customers - a labor-intensive position. The middle office houses an intelligent brain providing strategic guidance for next customer communications, including a voiceprint recognition agent primarily for identifying anti-collection black market agents - if one voiceprint matches multiple phone numbers, it's likely a black market agent. The back office quality inspection agent operates like quality inspection specialists.

How are collection agents created?

First, we trained collection-specific large models using approximately 150,000 effective samples from top sales personnel for training. We conducted extensive corpus annotation, including emotional annotation, to rapidly learn entire learning processes. Additionally, we have builders ensuring speed and suppressing hallucinations, designing a one-to-four intelligent agent architecture supporting front-line collection agent operations.

I believe this AI-enabled business know-how promotes democratization of banking business capabilities. It raises the ceiling of banking business capabilities by transforming industry best practices into agent operational processes and specialized capabilities, enabling us to achieve the third goal of delivering value at scale. Taking collection as an example, we can simultaneously solve three major pain points: under unchanged budgets this year, productivity can increase threefold while complaints decrease by 95%. This is our example from a bank credit card's M1 stage (1-30 days overdue) operational practice, which has shown significant cost reduction effects: leveraging this solution, partner institutions successfully reduced 10 manual agents while handling 90 million collection cases. Behind this data, we can further see the scalable potential of AI technology in cost optimization - if case volumes increase to 900 million or even 9 billion cases, how much cost would be saved?

Next, let me briefly introduce credit reporting. Many leaders discussed this product today. What's most difficult about credit reporting? AI-enabling human expert experience. Technology departments don't understand business, while business departments are reluctant to participate in large model intelligent agent production. What should we do? I think cooperation with third-party technology companies is a good choice.

Credit reporting agents replicate human experts' thinking and working logic, knowing what data to examine, where to find it, from which dimensions to analyze, and what template to ultimately form - this is what human experts do. The most difficult aspect of credit reporting isn't producing a report but having depth, associations, and usefulness for subsequent decision-making.

Credit reporting also requires industry analysis, analyzing characteristics of different industries. Technology-intensive enterprises and capital-intensive enterprises have different industry characteristics. We need industry cycle data and understanding of industry indicators, which are crucial for future credit decisions.

For due diligence report generation products, we achieved multi-intelligent agent collaboration, from front-end data OCR input to cross-verification between financial statements, which is most critical for small and micro due diligence reports. This morning, Agricultural Bank leaders mentioned drone data, satellite remote sensing monitoring data, and customer manager actual investigation data - how do we conduct cross-verification between data sources and balance between reports? This requires multiple intelligent agents working together to complete in-depth due diligence reports.

Let me introduce BAIRONG-W: BAIRONG-W is a Hong Kong-listed one-stop AI technology leader serving over 7,000 financial institutions with deep cultivation in the financial industry for over ten years. In May this year, Morgan Stanley published the "China AI: Awakening Sleeping Giant" white paper, focusing on China's AI industry core - China's top 60 AI enterprises. BAIRONG-W is the only company included in Morgan Stanley's "China AI 60-Finance AI" list.

This concludes my sharing today. Thank you all, and I hope to have further exchanges with everyone present.