The 2026 World Artificial Intelligence Conference (WAIC) and the High-Level Meeting on Global AI Governance were recently held in Shanghai.
At the conference, Qifu Technology (QFIN) delivered a keynote presentation titled "From Intelligent Risk Control to Financial Agents: AI Native Reconstructs the Safe Growth Paradigm for Inclusive Credit." The company's Chief AI Scientist, Yang Qing, stated that the next phase of financial AI is not merely about keeping large models at the level of Q&A, writing, or isolated efficiency gains, but about enabling AI to progressively integrate into financial production systems, thereby reconstructing data comprehension, business models, process execution, and decision-making services.
Qifu Technology outlined a six-layer application pathway for large financial models. The first layer focuses on enhancing knowledge work efficiency, establishing a "Prompt + Draft Generation + Expert Review" workflow for roles in operations, compliance, and analysis. The second layer involves structuring complex data, transforming intricate financial materials such as documents, contracts, transaction records, reports, conversations, and videos into interpretable, model-ready, and evaluable business variables.
The third layer is business model enhancement, where large models participate in classification, ranking, and risk representation learning to strengthen risk control and operational models. The fourth layer centers on intelligent interaction and human-machine collaboration, making financial service interactions more controllable, traceable, quality-assured, and iterable.
The fifth layer is the "Skill-ification" of processes, where expert knowledge and business workflows are solidified into reusable, auditable, and iterable capability assets. The sixth and highest layer is Agent-based decision-making services for credit business. Here, a large model acts as a planning hub, invoking models, rules, Skills, and business systems within compliance, risk, and authorization boundaries to generate more complete, consistent, and auditable decision support suggestions.
Regarding the structuring of complex data, Qifu Technology has independently developed and open-sourced the FCMBench evaluation system to systematically assess multimodal material comprehension capabilities in financial credit scenarios. For business model enhancement, the company is exploring a credit investigation large model approach, allowing it to participate in classification, ranking, and risk representation learning to form verifiable, model-ready risk sub-scores.
In the area of process Skill-ification, Qifu Technology recently launched and open-sourced the automatic modeling framework ModelEvo. This framework attempts to solidify processes such as task clarification, sample construction, feature mining, model development, performance evaluation, and report review by modeling experts into a Skill Pipeline. Compared to traditional AutoML, ModelEvo places greater emphasis on the continuous accumulation of business knowledge, process experience, and outcome feedback, enabling financial professional processes to iteratively self-improve through execution, review, and updates.
At the higher level of credit business Agents, Qifu Technology cites the "AI Approval Officer" as a typical practical example. Operating within compliance and risk boundaries, the AI Approval Officer comprehends application materials and business status, invokes risk control models, approval rules, and compliance verification tools, and generates approval suggestions, key evidence, and manual review points to assist human approvers in making more complete, consistent, and auditable judgments.
Comments