A credit model approves a loan in milliseconds with no human in the loop. A chatbot explains a policy to a retiree and quietly invents a clause that does not exist. As banks and insurers race to embed large models (大模型), China's regulators have decided that some decisions are too consequential to leave to the machine alone.
From encouragement to guardrails
The shift has been staged over roughly two years. At its March 2025 technology work meeting, the People's Bank of China (PBOC, 人民银行) told the sector to safely and orderly advance the application of AI large models (人工智能大模型) in finance. Later that year, at a December 2025 industry forum, PBOC's technology department signaled the next step: it would study and issue a policy document for the safe application of AI in finance and draft a safety specification for large-model technology in financial use, with risk classified and graded management (分类分级管理) and high-risk application access control (高风险应用准入管理).
That intent hardened into text on 18 June 2026, when the National Financial Regulatory Administration (NFRA, 金融监管总局) released the Guiding Opinions on the Safe Development and Application of Artificial Intelligence in Banking and Insurance (《关于银行业保险业人工智能安全开发应用的指导意见》, document No. 金发〔2026〕8号). It carries 32 articles spanning governance, development, data, compute, risk, capability, and supervision.
The core principle: whoever uses it owns it
The document opens with four operating principles:
- Whoever uses it is responsible (谁使用谁负责) — the financial institution, as the service provider and AI user, carries the liability.
- Self-reliant and controllable (自主可控) — raise the autonomy of key platforms, software, and hardware.
- Pragmatic and efficient, and
- Safe development — obey cybersecurity and data-security law.
In practice this means the regulator does not certify AI as "safe"; it makes the bank or insurer prove it can govern the model end to end.
The high-risk red lines
The most concrete part is the risk tiering. Applications touching money movement, asset valuation, credit approval, underwriting and claims, risk management, or anything directly facing customers are treated as high-risk. For those:
- deployment must be approved by the institution's risk-management committee;
- human supervision and intervention are required at key steps, with explicit emergency-stop and model-exit conditions;
- a backup process or human alternative must exist if the model fails.
Where a model's reasoning is not explainable, the guideline says it may serve only as an auxiliary tool — a human makes the final call on any decision with a real financial or customer impact, and the raw data, inference path, and trigger records must be preserved for traceability.
Data and model hygiene
Two rules bite directly into how models are built:
- No personal data in training. Names, ID numbers, phone numbers, and bank-card numbers must not be used to train or optimize generative-AI models, closing a common leakage path.
- External models must be filed. Generative-AI models brought in from outside must have passed the Cyberspace Administration of China's (CAC, 网信办) filing (备案) process — linking financial AI to the broader generative-AI service registry.
On outputs, AI-generated content must be clearly labeled and actively explained to consumers, and institutions using generative AI in public-facing or high-risk scenarios must report to NFRA.
Building the rails, not just the brakes
The guideline is not only restrictive. It also pushes capability:
- build enterprise-grade Model-as-a-Service (MaaS) platforms so models are shared and reused rather than re-built per team;
- let large institutions export compute services to smaller ones, and explore shared infrastructure;
- stand up high-quality datasets and knowledge engineering per business scenario;
- run adversarial testing, prompt-injection drills, and supply-chain audits, including for open-source components.
It even tasks NFRA with an annual assessment of its own AI policy and effect, a rare self-review clause in Chinese financial rulemaking.
The longer runway
This sits on earlier scaffolding: a 2021 PBOC Evaluation Specification for AI Algorithms in Financial Applications, a 2023 financial-industry Disclosure Guide for AI Algorithms in Financial Applications, and NFRA's December 2025 Guiding Opinions on High-Quality Digital-Finance Development in Banking and Insurance. The June 2026 document is the first to pull those threads into a single, 32-point operating manual for AI itself.
Honest limitations
This article is based on NFRA's published guideline, PBOC meeting reports, and China Financial News commentary; it does not measure how banks and insurers have operationalized the 32 articles, which will vary by institution size. The document sets principles (e.g., "high-risk," "self-reliant") whose exact thresholds depend on forthcoming implementation specs we cannot confirm here. No monetary penalty amounts or compliance budgets were published in the source, so cost impact is not quantifiable. The guideline governs banking and insurance; securities and fund regimes may follow separately and are out of scope.
What readers can do now
- If you sell AI to banks or insurers, map your product to the high-risk tiers now — committee approval and human-override hooks will be procurement prerequisites.
- If you run model risk, build the filing and labeling evidence trail early, since external models must already carry CAC registration.
- If you are a compliance lead, draft the annual self-assessment the guideline implies before regulators ask for it.