A compliance officer at a mid-sized AI vendor gets a questionnaire from a client: prove you manage model risk, hallucination, bias and privacy across the lifecycle. Until now, the honest answer was "we have a policy." After 1 November 2026, that answer stops being enough.
Key takeaways
- YD/T 7173-2026, "Artificial Intelligence — Safety and Governance — Capability Requirements for System Risk Management," was issued under MIIT Announcement 2026 No. 18, published 15 July 2026 and effective 1 November 2026.
- It is an industry/recommended standard (行业标准), not a mandatory national standard (强制性国标) — so compliance arrives through procurement and certification, not an immediate legal penalty.
- It defines three capability systems: organisational governance, process management (identify–assess–respond–improve), and risk protection across eight properties.
- Drawn up by the China Academy of Information and Communications Technology (CAICT, 中国信通院) with more than 20 companies, universities and research institutes.
- As of 29 October 2026 the standard is not yet in force; it takes effect 1 November, so firms have days, not months, to align.
What the standard actually requires
The document is narrow and operational. It specifies the management and technical capabilities that AI research and application entities must have when they run risk management, rather than banning any particular model or use. Its backbone is three capability systems.
The first is organisational governance (组织建设能力): a clear three-tier chain of accountability — leadership decides, management approves, operators execute — plus training, dedicated resources, full-lifecycle documentation, and a mechanism to report major risks to regulators, suppliers and users.
The second is process management (流程管理能力): a closed loop of risk identification, assessment, response and continuous improvement. Unacceptable risks are defined as hazards and graded major, moderate or general; responses are framed as four options — avoid, reduce, accept or transfer — with assessment and response reports retained.
The third is risk protection (风险防护能力): the system should demonstrate eight properties — reliability, controllability, safety, fairness, transparency, explainability, privacy protection and accountability.
Why an industry standard, not a law
This matters for how the rule lands. A mandatory national standard (GB) can carry direct legal force; an industry/recommended standard (YD/T) signals the expected baseline and is enforced indirectly — through customer procurement requirements, certifications such as ISO/IEC 42001, and pilot assessments. CAICT says it has already used the framework to help domestic manufacturers pursue ISO/IEC 42001 certification, and that more than 10 AI companies have built risk-management capability on it.
For foreign readers, the distinction is the headline: China is governing AI through technical standards that move faster than legislation and travel through supply chains. A vendor that cannot show the three capability systems may still legally operate, but it will struggle to sell to cautious enterprise and government buyers.
The timing is deliberate. By publishing in July 2026 and setting a November effective date, regulators gave larger vendors a short runway and smaller ones a template to copy, rather than a long transition that lets compliance slip. Because the document is technical rather than prohibitory, it is also easier for overseas suppliers to adopt than a ban would be — which is part of why CAICT frames it as input to international norms rather than a wall. A standard a foreign firm can read and match is more useful to Chinese buyers than a rule they cannot satisfy.
Where it goes next
CAICT plans three follow-on steps: training and interpretation sessions for technology providers and users; pilot applications in finance and healthcare to produce reusable playbooks; and a 2026 AI risk-management capability-maturity assessment. A companion effort on agent-service risk management is already underway, focused on tool-supply-chain security, memory safety, reasoning execution and platform operations.
The standard also positions China in a global conversation. CAICT explicitly situates it alongside the EU AI Act, the US NIST AI Risk Management Framework and Singapore's 2026 AI safety consensus, casting the document as both domestic rule and an input to international norms.
Honest limitations
The standard is "recommended" (行业标准), so I have not overstated enforcement; pressure arrives via procurement and certification, not a statute on day one. The drafting-organisation count differs slightly between CAICT's "22" and the national standards platform's longer list, so I used "more than 20" to stay safe. I did not verify whether any province has already made the standard mandatory in local procurement; CAICT mentions pilots in finance and healthcare but not binding rules. The effective date (1 November 2026) is from the national standards platform, where the status reads "即将实施" (about to take effect), confirming it is not yet active on the article date.
What readers can do now
- If you sell AI into Chinese enterprises or government, map your risk-management documents to the three capability systems before 1 November 2026.
- If you are a Chinese AI vendor, use the standard's "one table, three lists" toolkit to prepare for the 2026 capability-maturity assessment CAICT plans.
- If you are a foreign compliance team, read YD/T 7173-2026 alongside ISO/IEC 42001; CAICT explicitly links the two, and alignment may ease cross-border certification.
According to Li An, chief scientist at BrainNet (脑机网), the value of the standard is less about what it compels than about giving smaller AI firms a concrete checklist they can actually staff and audit.