Banks are beginning to treat artificial intelligence less as a customer-facing novelty and more as an engineering capability. CITIC BaiXin Bank has released a white paper, prepared with Baidu as a technology partner, on building an AI-native approach to software development. The document focuses on the work that happens before a banking app reaches a customer: understanding requirements, designing systems, writing and testing code, releasing changes, monitoring operations, and checking security. Sina Finance reported that the bank organized the white paper and that Baidu supplied technical support and examples related to coding standards.
That focus is important because banking is a demanding environment for AI adoption. A model can help summarize a requirement or suggest code, but a financial institution cannot hand responsibility for security, compliance, customer data, or material operational decisions to a tool. The white paper’s central idea is therefore not full automation. It is a structured division of labor in which AI participates throughout the development process while people retain responsibility for judgment and risk.
A Bank Maps Where AI Can Assist Its Engineers
The white paper examined 247 specific work items and grouped the principal development process into seven stages and 42 key links, according to JRJ. It found that roughly 80% of those links have some scope for AI assistance. That does not mean 80% of a bank’s technology work can be safely handed to a model. It means there are many repeatable tasks where AI can help engineers prepare, check, explain, or accelerate work under defined controls.
The report divides tasks by the degree of human involvement they require. Clear, repetitive work can be AI-assisted under supervision. Work requiring professional judgment can use an AI system for analysis and preliminary options, followed by a human decision. Matters involving strategic direction, major risk, cross-departmental coordination, or accountability remain human-led. This framework is useful because it avoids the false choice between banning AI from a regulated environment and treating it as an autonomous engineer.
The approach has a clear commercial logic. Banking software teams are asked to deliver new features quickly, maintain legacy systems, respond to changing rules, and protect highly sensitive data. AI can help with code generation, test creation, defect review, documentation, and operations. But a faster pipeline is only valuable if it also preserves traceability and security.
Pilot Results Need Attribution, Not Hype
The white paper describes a “SPEC × Harness” approach, which the bank presents as a way to make requirements explicit before asking AI to execute tasks. The premise is simple: an AI system needs a clear description of the intended outcome and equally clear boundaries on what it may not do. In a financial setting, that can reduce the risk of an apparently plausible output that conflicts with controls or business rules.
JRJ reports that, in certain internal scenarios, AI-assisted defect review achieved more than 90% accuracy and reduced the time to review a single issue from 15 to 30 minutes to one or two minutes. The same report says pilot work cut labor input for important system modules by 45% and raised test coverage by 80%. These are bank-reported pilot outcomes, not independent performance audits. They should be read as evidence of what one institution says it has achieved, rather than a universal benchmark for the sector.
The distinction matters. AI-assisted software development is highly sensitive to the quality of source code, requirements, testing data, governance, and human review. A productivity gain in one part of a bank may not transfer directly to another, particularly where legacy systems or regulatory obligations differ. Still, the white paper shows that Chinese financial institutions are beginning to quantify AI’s role in internal technology work rather than discussing it only in broad strategic terms.
Banking AI Shifts From Tools to Operating Discipline
The report’s emphasis on responsibility is perhaps its most consequential point. It says AI can assist execution, but final decisions and accountability must stay with people. That principle aligns with the work already underway in Chinese finance, where banks have been building AI loans around chips, compute, and models while regulators and financial institutions consider how AI changes risk management. It also depends on the more rigorous evaluation work described in China’s expanding AI data-services market, because development tools cannot be trusted without reliable testing inputs.
CITIC BaiXin Bank describes several controls, including internal deployment, keeping data within secure areas, automated security checks, peer review, and periodic audits. Those measures are not unique to banking, but the sector’s combination of personal data, transaction systems, and regulatory accountability makes them essential. AI-native development cannot simply mean faster development. It has to mean a development process that produces evidence of what the system did, what a person reviewed, and where responsibility lies.
The white paper also provides a more grounded view of the “AI-native” label. It is not a claim that banks will rebuild every system around a single model. It describes an operating model in which AI is inserted into many stages of the technology pipeline with rules for human oversight. That is a less dramatic story than autonomous banking, but it may be the more durable one.
For Baidu, the partnership offers a route into a high-value enterprise market where customers need models, tools, implementation support, and governance expertise together. For CITIC BaiXin Bank, it offers a chance to formalize lessons from internal experimentation. The larger question is whether this kind of structured approach can become a standard way for Chinese banks to adopt AI without creating a new category of software and compliance risk.