Cohort: 1 

Location: Beijing, China 

Application Title: AI infrastructure across over 20 financial business domains 

Team Members: Yang Longru (executive sponsor), Jin Haimin, Luo Yi, Gao Hongsheng, Liu Chengyan, Bao Xinyue

The challenge 

Banks operate under strict regulatory requirements while managing complex processes that span credit assessment, risk control, customer engagement and capital allocation. Although AI technologies have advanced quickly, their adoption in large financial institutions has remained limited. 

Traditionally, AI in banking has been applied in scenario-specific ways. Examples include customer service chatbots or isolated risk models. These tools can improve individual tasks, but they do not integrate across full business processes. Infrastructure limitations, data sensitivity and the need for auditability make it difficult to deploy large models at scale. As a result, many AI initiatives remain experimental and fail to translate into sustained operational impact. 

For a bank with thousands of branches and hundreds of thousands of employees, this fragmentation limits the ability to use AI consistently across the organisation. 

The solution 

A financial AI platform built as shared infrastructure 

ICBC addressed these constraints by developing AI as a foundational capability rather than a collection of individual applications. 

At the centre of this approach is a large language model trained specifically on financial data and banking workflows. The model is designed to support reasoning across documents, transactions and regulatory rules, rather than generating generic responses. ICBC built a centralised AI platform around this model, allowing specialised AI agents to be deployed rapidly across business units. 

This infrastructure supports more than 20 financial domains, including credit operations, risk management and financial markets. Instead of building separate systems for each function, teams can adapt the shared platform to their needs while operating within a common governance framework. 

How it works 

From model training to daily use 

The financial language model is trained using a combination of supervised learning, fine-tuning and reinforcement techniques, with data drawn from compliant public sources and ICBC’s internal records. Data processing includes automated validation and safety checks to meet regulatory standards. 

A central control layer manages deployment and monitoring. This allows new AI agents to be released to branches within minutes rather than months. Staff interact with the system through existing tools, using AI to assist with analysis, documentation and decision support. Human review remains part of the process in areas where judgment or accountability is required. 

The underlying infrastructure includes large-scale computing resources that support continuous training and updates without interrupting operations. 

How it differs from conventional approaches 

Moving beyond isolated financial AI tools 

Conventional banking AI solutions focus on individual use cases and are often developed independently. This leads to duplicated effort and inconsistent standards across the organisation. 

ICBC’s approach provides a shared platform with domain-specific intelligence embedded at the core. Rather than adapting generic models to financial tasks, the system is designed around banking requirements from the outset. This enables broader adoption while maintaining consistency, security and auditability. 

The same framework can be extended across branches and teams without rebuilding systems each time. 

Impact 

• More than 400,000 users, with around 20,000 staff using the platform daily 
• Approximately 50,000 person-years of workload handled through AI-assisted processes 

Why it matters 

Making AI usable at financial scale 

For large financial institutions, the challenge is not whether AI works in principle, but whether it can be deployed safely and consistently across the organisation. ICBC’s platform shows how AI can be treated as shared infrastructure rather than a collection of experiments. 

By embedding domain-specific models into daily workflows and aligning them with regulatory requirements, the approach supports broader use of AI while preserving control and accountability.