How ICBC Revolutionized Credit‑Card Risk Management with AI‑Driven Data Architecture

1. Current Situation and Challenges

The credit‑card business is rapidly moving toward scenario‑based and intelligent depth, and the risk‑control management model for credit‑card personnel faces new challenges such as cross‑business monitoring, agile perception of risk trends, and the need to improve operational efficiency. There is an urgent need to strengthen data support for business decisions, enhance AI‑driven risk perception, and achieve precise capture, rapid response, and comprehensive coverage of risk signals across the entire business chain.

ICBC’s Software Development Center credit‑card shared data service team focuses on risk‑management scenarios for core credit‑card staff. Following a "proactive prevention, intelligent control, comprehensive management" approach, they propose a digital risk‑control solution of "unified management, efficient prevention, intelligent control". Based on the risk characteristics of credit‑card business and profile tags of core staff, the solution builds a multi‑dimensional, high‑timeliness risk‑monitoring indicator system and leverages an intelligent decision‑engine tool to achieve precise identification, early warning, and rapid handling of risks in key links such as credit approval and core staff, forming a full‑process, intelligent risk‑control closed‑loop that supports a virtuous interaction between business development and risk control.

2. Solution – Data Layering and Standard Data Service Support

Using an enterprise‑level data‑mid‑platform model, the team processes credit‑card data into thematic, metric, tag, and model forms, constructing a three‑layer data architecture (source layer, aggregation layer, extraction layer) to respond to rapid business changes, enhance data value, and promote data‑asset development.

Source Layer Data Storage : Following the principle of "compliant and complete", upstream applications ingest structured credit‑card data (e.g., loan transactions), unstructured data (e.g., images, log data), and external data (e.g., government data) into a data lake for centralized storage and bank‑wide sharing, providing the foundation for downstream aggregation and extraction.

Aggregation Layer : Based on the data lake and unified design standards, data are standardized, stitched, and dimensionally reduced for business themes such as card issuance, limit adjustment, and installment, forming unified business‑view aggregates that lower integration difficulty and data‑use barriers.

Extraction Layer – Metric Processing : Targeting business needs, the team builds a layered metric model (atomic, composite, scenario‑derived) covering over a hundred indicators across pre‑loan, in‑loan, and post‑loan stages, supporting three core scenarios: personnel behavior anomalies, risk transmission, and continuous risk evaluation. This modular metric system enables flexible rule configuration for diverse credit‑card scenarios.

Leveraging an enterprise‑level decision engine, the solution integrates expert rules and model algorithms to enable dynamic rule configuration and automated alerts. Real‑time monitoring of high‑risk behaviors of core staff triggers automatic grouping and differentiated handling, improving precision and response efficiency.

Flexible Rule Configuration : Templates for issuance, limit adjustment, and installment allow visual adjustment of thresholds and parameters, with automatic iteration based on historical decisions to reduce manual trial‑and‑error.

Real‑Time Monitoring and Alerts : Rules based on marketing, investigation, and approval risk indicators automatically monitor and alert high‑risk personnel daily, supporting automatic or manual account freezing and early intervention.

Decision Support : Expert rules and model outputs feed the engine to support differentiated personnel management, simplifying processes for high‑performers and restricting risky actors, while pushing alerts to approval systems for timely intervention.

Closed‑Loop Optimization : A management dashboard visualizes performance distribution and risk trends, enabling rapid adjustments. Historical risk‑alert data feed back into the system for continuous learning, reducing false alerts by 50% and expanding coverage across the entire credit‑card workflow.

3. Application Effect

The digital transformation solution has markedly improved risk‑identification timeliness (from monthly manual checks to minute‑level automatic handling), expanded monitoring coverage to all core personnel and business types, and enhanced warning quality, cutting ineffective alerts by half and allowing managers to focus on genuine high‑risk events.

4. Future Outlook

As AI and big‑data analytics continue to evolve, the risk‑monitoring system will further improve real‑time precision and intelligence, moving from reactive response to proactive value creation. ICBC plans to deepen multimodal large‑model technology in financial scenarios, build a cross‑modal cognitive hub, strengthen the financial data‑asset library, innovate intelligent risk‑control applications, and optimize compute‑security foundations, achieving comprehensive upgrades in risk‑identification accuracy, compliance control, and innovative capability.