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 Applications in the Financial Industry." Song Rongxin, Vice President of Qifu Technology, attended and delivered a keynote speech.

The following is the transcript of the speech:

Distinguished leaders and guests, good day. I am honored to have the opportunity to exchange ideas with everyone at this platform during the trade fair. The speeches by the leaders before me have been tremendously insightful.

As a company that has long been committed to empowering the financial industry with the most advanced artificial intelligence technology, Qifu Technology would like to share today our practical experience in AI-Native + Finance over the past year, and discuss the application trends of intelligent agents in financial services.

Let me begin by sharing several sets of data:

From a global perspective, approximately 45% of financial institutions were applying AI in 2022, and this proportion is expected to jump to 85% by 2025. In terms of market investment, the scale of AI market investment is growing very rapidly, with the banking sector being the most active. The market is projected to exceed $27 billion by 2027. In China, the State Council has proposed that the penetration rate of next-generation intelligent agents will exceed 70% by 2027 and surpass 90% by 2030.

According to Gartner data, the implementation rate of generative AI in China is also rapidly increasing, from 8% last year to 43% this year, which aligns very closely with our practical experience. As of today, virtually no financial peer has not been exploring intelligent agents. We can see that policy, capital, and industry are all jointly driving AI toward deep applications in financial scenarios.

To systematically track this trend, we conduct annual research on financial AI products in the market. This year, we also established teams in the UK, Silicon Valley, and other locations to engage in in-depth exchanges with local banks and technology companies. After detailed analysis of over 600 products, we observed a very clear trend.

In 2023, approximately 85% of AI products were mainly concentrated in peripheral applications such as customer service and document processing. By this year, one-third of products have already penetrated core scenarios like risk control and customer acquisition. For example, a company I recently followed called Casca has not only applied large model technology to core processes but also used AI-Native architecture to rebuild the entire bank's credit platform. With a 90% reduction in personnel investment, it achieved a ten-fold efficiency improvement while serving a large number of small and micro enterprises that were previously underserved. Overall, the application of financial large models is indeed transitioning from "supportive tools" to "core business processes."

From Qifu Technology's own development perspective, our evolution on this path has been highly consistent with the industry. Those familiar with us may know that we started working on large models for finance relatively early. We established a large model team of over 100 people in 2023 and released a foundation large model at that time, though we no longer develop foundation large models now. We believe that period of technical accumulation was very important, but from an application perspective, 2023 was mainly focused on relatively peripheral scenarios like customer service assistants and marketing quality inspection. By 2024, we achieved more substantial results with products like Jarvis and ChatBI, gradually entering business efficiency enhancement areas. While these received considerable recognition, business feedback clearly told us that these applications were not core enough. Taking credit as an example, if you don't engage in risk control or customer acquisition, you're essentially not working on the core functions.

There were several reasons we didn't pursue core applications in 2024:

First, from a technical perspective, inference costs were still quite high and not cost-effective.

Second was the hallucination problem. Although we used RAG technology to solve many hallucination issues, in core financial scenarios, we often need to perform multi-step reasoning, which wasn't resolved very well last year.

Third was the product entry point - our thinking wasn't very clear on this aspect.

2025 marked a turning point. On one hand, new models like DeepSeek R1 have indeed driven rapid industry iteration. On the other hand, leveraging our understanding of underlying data, we've gained clearer insights into entry points. As mentioned earlier about creating super employees, we deeply resonate with this concept. Many people might think that large model applications must replace certain core modules within existing systems to achieve excellent results. I now believe that starting with human assistance is an excellent choice. For instance, converting physical labor time in many important positions to mental labor time is also very valuable.

Additionally, to build better assistants, we've updated our AI-Native architecture to fully leverage the advantages of data scenarios and models.

Based on this foundation, the intelligent agents we've developed today - AI Compliance, AI Credit Officer, and AI Approval Officer - can serve as effective assistants for core financial positions, directly entering core financial risk control processes and achieving the transformation of large models from peripheral to core applications.

Some may ask why we can implement AI in core financial business processes. Let me briefly explain:

The first important reason is our commitment to long-term investment. To date, we have invested over nearly 10 billion RMB in R&D funding in the AI field, maintaining a technical team of nearly 1,000 people dedicated to financial scenario innovation. These investments have resulted in substantial accumulations, including over 900 patents, more than 80 copyrights, and frequent publications in international top journals. We also collaborated with the China Academy of Information and Communications Technology to release China's first financial large model standard. More importantly, having such long-term investment determination gives us the confidence to pursue deep exploration of AI applications during critical moments, rather than remaining at the conceptual level.

The second reason is our data and scenario foundation. I won't read all the data here, but you can see that under Qifu Technology's base of 200 million registered users, the daily business volume and scenario interaction volume are enormous. It's precisely this complex and diverse environment that provides intelligent agents with rich "environmental data," enabling continuous learning and evolution.

Honestly, I didn't realize this was so important before, but since developing intelligent agents, I've gained deep appreciation for the importance of this data. For example, through our recent research, we found that many bank relationship managers spend considerable time on customer relationship reviews near the end of their workday, often taking two hours to listen to customer call recordings or make notes in the system about customer loan needs and conversation content, mainly to avoid missing customer needs or sales opportunities and to prepare for future marketing.

I'll share an intelligent agent shortly that we're developing to provide AI assistance to relationship managers, significantly reducing time investment in this area. But the main point here is accuracy - according to our data review, if customer relationship reviews rely solely on human effort, accuracy reaches about 60%. Using general large models, accuracy is at most 70%, but after infusing daily environmental data into large models, accuracy can exceed 85% and continues to improve.

With rich scenarios and data, we also need an AI architecture that can maximize the potential of scenarios and data. Large model training capabilities and data processing capabilities are certainly essential basics. Currently, we see many intelligent agent applications are still rule-based or logic-driven. After model training completion, extensive data extraction and annotation work is required in advance, making iteration quite slow. Once core financial scenarios become complex, workload increases exponentially, and effects aren't sustainable.

We believe that under AI-Native architecture, intelligent agents themselves become the system's central hub, forming a closed loop of data, models, and business feedback. Simply put, the more you use it, the faster it learns, and the better the results become.

Next, I'd like to share several examples we've already implemented:

The first example is the AI Credit Officer, an intelligent agent that empowers core scenarios in credit customer acquisition and business operations. What is the core scenario for customer acquisition? People might think it's the credit application systems everyone has now, or some marketing tools, but we don't think so. We believe the core scenario for credit customer acquisition lies with relationship managers themselves. For instance, when relationship managers want to proactively acquire customers, where should they go to encounter potential customers? With the AI Credit Officer, they don't need to think about it - it directly recommends locations and detailed information.

For example, regarding the customer contact record summaries I mentioned earlier, how to record with minimal time investment - with the AI Credit Officer, you don't need to record manually. It can automatically tag and may do better than you would yourself. For instance, when wanting to provide meticulous customer care to multiple clients to improve stickiness and retention rates.

With an AI assistant, you can directly assign tasks. You can tell the assistant: "Please help me set up a task to send customized birthday wishes to every customer on their birthday based on their region, and send SMS in the morning." You only need to issue such a task, and the assistant will automatically handle it.

For product recommendations, you can also tell the assistant: "Help me find customers who have houses but no mortgages and, based on all contact records from the past three months, may have loan needs. Send them a recommendation message, and notify me if they show interest." You can give such instructions to the assistant.

The AI Credit Officer is an intelligent assistant we developed for relationship managers, aimed at truly satisfying complex core needs through intelligent agents, enabling everyone to become a sales champion. We will release the official B2B version this month. Current internal pilot results show it can help relationship managers improve per-person conversion rates by 15% and efficiency by 30%, particularly expanding the number of customers each person can manage from hundreds to thousands.

Another achievement is the AI Approval Officer, which enters the most core approval and risk control processes in finance. In traditional system architectures, AI serves more as a tool for material recognition or field extraction, combined with strategies and models for risk identification. This approach has high costs, long implementation cycles, and requires frequent credit iterations when there are many material modules. Under Qifu's AI-Native architecture, the Approval Officer is a true digital employee that can integrate full-scale data including credit, transactions, and behavior to complete cross-modal, end-to-end risk control reasoning.

For example, in review and approval scenarios, it has achieved zero supplementary documentation requirements, 100% automated review, and T+0 approval efficiency.

In risk modeling scenarios, it can serve as an intelligent assistant for risk strategy personnel, functioning like a senior modeling expert. In daily modeling work involving feature selection, attribution analysis, and model evaluation, after setting tasks for these intelligent assistants, they can operate continuously 24/7 until achieving set objectives.

Additionally, we have intelligent agents that can serve as McKinsey consultant roles for financial institutions in enterprise credit.

Finally, I want to emphasize that AI-Native intelligent agent applications in core financial scenarios have just begun. Looking toward the future, China possesses the world's most complex financial scenarios and largest user base, with hundreds of millions of daily interactions and real business cycles, becoming the best soil for the continuous evolution of intelligent agents. We believe the story of AI in the financial industry has entered a completely new stage. Qifu Technology will continue to firmly advance in the AI-Native + Finance track, cultivating deep roots in China and promoting our experience globally, contributing a Chinese solution to the intelligent upgrade of the world's financial industry.

Thank you all!