China has moved from broad, cross-sector AI rules to a dedicated, finance-specific framework that sets out concrete obligations for banks and insurers deploying artificial intelligence. On 18 June 2026, the National Financial Regulatory Administration (NFRA) issued its first comprehensive AI rulebook for the sector—Guiding Opinions on the Safe Development and Application of Artificial Intelligence in the Banking and Insurance Industry (Jin Fa 2026 No. 8)—marking a shift from principles to operational governance.
A new, sector-specific AI rulebook for finance
The NFRA framework comprises 32 guiding principles across seven pillars: governance; development and deployment; data governance; computing infrastructure; risk management; capability building; and supervisory support. It binds banks and insurers (including China operations of foreign institutions) and extends to AI vendors and computing suppliers through outsourcing and supply‑chain provisions.
Operationally, the rules require: full‑lifecycle model management from business requirements through retirement; graded risk classification of AI applications; mandatory human oversight and emergency shutdown for high‑risk uses; admission controls for generative‑AI models; and mandatory reporting when generative AI is used in public‑facing or high‑risk scenarios. A standout provision is the explicit prohibition on using personally identifiable information—names, ID numbers, mobile numbers and bank account numbers—to train or fine‑tune generative AI models, forcing institutions to desensitize datasets and prove exclusion via data lineage checks.
What counts as “high‑risk” and why it matters
The guidelines formally classify “high‑risk AI applications,” including fund transfers, asset valuation, credit approval, underwriting, claims assessment, risk management and any AI directly influencing financial contract decisions. These applications must be approved by institutional risk committees, incorporate human oversight at critical decision points, and trigger regulatory reporting when generative AI is deployed in customer‑facing or high‑risk use cases. The approach pushes AI governance from a technology initiative to a board‑level issue, aligning China’s banking regulation with emerging global standards.
Broader mainland controls: filings, content labelling, and continuous supervision
Beyond NFRA’s finance‑specific rules, AI deployments sit within a wider PRC architecture that includes the Cyberspace Administration of China (CAC), People’s Bank of China (PBOC), and China Securities Regulatory Commission (CSRC). The Interim Measures for Generative AI (effective August 2023) and related algorithm governance rules impose security assessments, filing obligations and content controls that are especially stringent in financial services.
Two practical compliance themes have intensified: content labelling/traceability and continuous supervision. The Measures for Labelling AI‑Generated Synthetic Content require both explicit user‑visible labels and implicit metadata tags to enable traceability across the AI content value chain. Meanwhile, regulators have shifted from one‑off approvals to lifecycle oversight: where use cases, model functionality, data sources or user reach change, institutions may need supplementary security assessments, updated filings or proactive engagement with regulators. Enforcement campaigns—such as CAC’s 2025 “Qinglang: Rectifying the Abuse of AI Technology”—signal that non‑compliance can trigger rectification orders, public criticism and administrative penalties.
Data governance and model oversight in financial institutions
Data and model governance expectations are tightening in parallel. The PBOC’s Administrative Measures for Data Security in its business areas (effective 1 May 2025) require financial institutions to implement data classification and grading, maintain data inventories, identify personal/sensitive/important data, allocate responsibilities and apply full‑lifecycle data security controls. In December 2025, NFRA’s Implementation Plan for High Quality Development of Digital Finance encouraged banks and insurers to build enterprise‑level AI and model management platforms to centralize development, deployment and monitoring.
National standards are also taking shape, covering machine‑learning security assessment, synthetic content labelling and training data security, and are expected to act as regulatory yardsticks for “appropriate” safeguards in supervisory practice.
Where AI is being used—and how far it can go
China’s banking sector is rapidly scaling AI across customer acquisition, decision‑making, operations and risk control. Industry leaders report hundreds of AI‑powered scenarios: ICBC has implemented over 200 use cases and logs more than 1 billion AI calls annually, spanning algorithmic credit advisors, forex trading assistants and intelligent risk detection. Bank of China has deployed the DeepSeek R1 model for automated coding, internal memos and business intelligence. Listed banks have added AI wealth assistants, intelligent investment research tools and automated document audit workflows.global.
Regulators, however, emphasize that AI remains primarily supportive and cannot replace human decision‑making. At the 2025 Bund Summit, NFRA vice‑minister Xiao Yuanqi noted AI applications in finance focus on process optimization and external services, with three key areas: intelligentization of mid‑ and back‑office operations, customer engagement, and financial product delivery.
Practical steps for compliance teams and boards
For financial institutions operating in China, the immediate agenda is operational readiness:
- Inventory AI applications and tag high‑risk/public‑facing use cases.
- Map human‑oversight and emergency shutdown paths for each high‑risk scenario.
- Run data lineage checks to prove exclusion of names, IDs, phone numbers and bank‑card numbers from generative‑AI training/tuning sets.
- Align model governance with lifecycle controls: validation, monitoring, drift detection, periodic review and decommissioning.
- Prepare for algorithm/model filings and security assessments where required, particularly for customer‑facing assistants and wealth‑management support tools.
- Implement dual labelling (explicit and implicit) for AI‑generated content where applicable, and preserve labels across dissemination channels.
- Strengthen third‑party risk management: vendor diligence, audit rights, change‑management controls, business continuity and substitution planning.