Artificial intelligence has transitioned from experimental applications to the core operational backbone of the banking industry. Digital China Information Service Company's Chief Consulting Expert, Sun Zhongdong, outlined this transformation at the 8th China Fintech Forum, held during the China International Fair for Trade in Services in Beijing on September 9. His address, titled "Global AI Implementation Paradigms and China's Strategic Framework," delved into three critical layers: worldwide AI advancements with a focus on banking practices, the evolving global regulatory landscape, and China's comprehensive approach to tackling implementation hurdles.
The central thesis is that AI has now entered the most error-sensitive business lines in banking. Success is no longer determined by model demonstrations but by deployment capabilities, semantic alignment, and runtime governance. Leading financial institutions are bypassing peripheral use cases and assistant layers, directly embedding AI into mission-critical systems such as credit adjudication and anti-money laundering. This shift is driving modernization across front-office operations like corporate credit, retail lending, and transaction banking, as well as mid-office functions including development and data governance. Institutions leveraging AI in core modernization efforts are seeing capability improvements ranging from 25% to 50% in related areas.
Where the industry stands
Approximately 30% of banks have achieved transactional task replacement, yet most institutions remain primarily in the assistant-type application phase. The industry is pivoting from "creating assistants" to "redesigning core business processes." These scenarios are not peripheral trials but rather the lifeblood of banking operations. For instance, in document examination, intelligent agents must integrate document recognition, awareness of the Uniform Customs and Practice for Documentary Credits (UCP) rules, and compliance clause implementation to achieve transformation within the core business process, not merely layer a chat window on top of existing workflows. The fundamental question has shifted from "whether a large model can be connected" to "how to redesign the core operational chain using an agent-based approach."
Global regulatory frameworks
The regulatory landscape has evolved rapidly, with key documents emerging from various jurisdictions. The European Union's AI Act, Singapore's AI Agent Governance Framework followed by the Runtime Safety Assurance Framework for Financial Agents (currently the world's most advanced management framework for financial agents), and the National Financial Regulatory Administration's "Document No. 8" issued in June 2026 collectively provide critical guidance for moving AI from demonstration to core business governance. These frameworks can be distilled into five fundamental principles: first, the principle of user accountability, where responsibility cannot be shifted to models or partners. Second, risk-based tiering, where governance intensity escalates with business risk levels. Third, mandatory human checkpoints, particularly in high-risk scenarios, ensuring critical gates in human-machine collaboration remain under human control.
Fourth, end-to-end traceability is essential. Given the inherent opacity of large models, banks must establish auditable and reviewable evidence chains when deploying them in core operations. This requirement underpins the emphasis on ontology in subsequent discussions. Fifth, governance must shift from post-hoc to runtime controls. When AI fails within critical workflows, the cost of after-the-fact remediation is prohibitive, necessitating forward-moving control points. International major banks are rapidly building these governance capabilities.
Key gaps and the Chinese approach
Despite the proliferation of AI pilots globally, only about 5% produce measurable performance impact, leaving 95% failing to justify their investment. Most pilots stagnate at demonstration stage, failing upon real-world deployment. The bottleneck is typically not model capability but deployment infrastructure, revealing two critical gaps. The first gap concerns Forward Deployed Engineer (FDE) capabilities, a model systemized by Palantir. FDEs must simultaneously understand business operations, AI technology, data structures, and institutional semantics. This composite talent is exceptionally scarce, with a regional bank often identifying only a handful of qualified individuals across the entire organization. Around June 2026, major players including Anthropic, OpenAI, Accenture, AWS, and Google began intensively investing in FDE capabilities, though current supply primarily serves general industries. When compounded with financial business and compliance complexity, this talent becomes even harder to secure.
The second gap lies in context and semantic layers. The challenge extends beyond traditional data governance to enabling large models to genuinely comprehend banking operations. Data and platform providers such as Databricks, Microsoft, Snowflake, and Palantir are all thickening their semantic layers so models can understand fields, rules, and business meanings, rather than simply processing tables. In response to these gaps, China's strategic framework relies on three pillars: customer journey reshaping, intelligent agents, and ontology. The first pillar addresses the FDE shortage, while the latter two respectively generate productivity and governance constraints.
Customer journey reshaping as a talent bridge
Customer journey reshaping differs fundamentally from typical process optimization. It has been an effective method for enhancing bank capabilities during the digital era. With AI's arrival, the key lies in identifying pain points and finding entry points where models and agents can genuinely add value. Global banking is adopting similar methodologies. The current challenge: business personnel lack complete understanding of technology and AI, while technology teams lack comprehensive business knowledge. This round of implementation is primarily a business problem before it is a technical one. Customer journey reshaping provides a shared language and methodology, enabling business and technology teams to form joint task forces that clearly define "where to apply, how to apply, and to what depth," partially offsetting the shortage of forward-deployed talent.
Intelligent agents: defining relationships before engineering
Before deploying agents, two elements require clarification. The first is VRF, or the positioning of the human-AI relationship. Without clear relationship definitions, subsequent governance complexity and applicable boundaries become unmanageable. The second is ADC, which transforms agents into digital employee job descriptions, specifying responsibilities, authorities, inputs, outputs, and collaboration methods. Two strengthening paths must proceed in parallel: governance must satisfy regulatory requirements such as Document No. 8, and agent engineering should keep pace with emerging tracks like environment engineering (Harness). While model and foundational engineering capabilities primarily originate from major technology companies, the application institution's critical task is using correctly and managing properly, not building alternative underlying capabilities.
Ontology: containing hallucinations and ensuring traceability
Ontology has gained momentum across multiple industries recently. Rather than simply benchmarking against specific companies, we should examine what principles can solve. For banks, ontology serves as an internal business world model for machine use, making operations interpretable and traceable while constraining hallucinations when large models function. It does not directly eliminate the internal hallucination mechanism of models but instead uses axioms, definitions, and rules within core business chains to constrain outputs within acceptable boundaries. Semantic alignment means: banks contain numerous internal definitions and rules that large models do not inherently know. If large models are to be used in core operations such as loan adjudication, mechanisms must enable them to understand the institution's specific practices, not just general financial knowledge from public corpora.
Consider an example: a 17-year-old entrepreneur applies for a 5 million yuan loan with over 100 million yuan in transaction flow. Based solely on surface materials, an agent might lean toward approval. However, the bank has an internal axiom: applicants must be at least 18 years old to receive such loans. The application is therefore intercepted. The age threshold represents just one among countless axioms; the value of ontology lies in organizing these axioms into a machine-executable, human-auditable business world, simultaneously responding to regulatory traceability requirements and reducing semantic drift in core scenarios.
Synthesis and path forward
Customer journey reshaping, intelligent agents, and ontology together form a tripartite methodology for AI implementation in core banking operations: journey methods bridge the composite talent gap, agents generate productivity, and ontology satisfies regulatory demands for interpretability, traceability, and runtime constraints while mitigating hallucination risks from semantic misalignment. Global competition should not focus solely on model capabilities. Whether banks can truly implement AI depends more critically on whether deployment, semantics, and governance can be simultaneously achieved.
Comments