A procurement manager asks a vendor whether its factory agent platform "meets the Chinese national standard for industrial AI agents." The vendor says yes. There is no such standard to meet. What exists is a project entry on a public platform — and the gap between those two things costs companies real money.
Key takeaways
- 《信息化和工业化融合 工业智能体治理 第1部分:总体框架》 is registered on the national standards information platform as a 国家标准项目 (national standard project): new work, 推荐性 (voluntary), a 15-month project cycle, with a public notification window from 18 September to 18 October 2026.
- It is administered by MIIT and technical committee TC573, with drafters including the National Industrial Information Security Development Research Center, several universities, and companies such as Huawei and Zhipu.
- Separately, MIIT's third batch of 2026 industry standard work plan (工信厅科函〔2026〕249号, dated 1 June 2026, released 16 June 2026) contains 121 projects, of which seven are AI items — all new, all voluntary industry standards, all on a 12-month cycle.
- A project notification means drafting has been approved to start. It is not a published standard, not a compliance requirement, and not a basis for any product claim.
- Both instruments are 推荐性 (voluntary). Neither is 强制性 (mandatory).
What "项目公示" actually means
Chinese national standards move through a visible pipeline on the national standards information public service platform. The first public stage is a project notification: a title, a scope, a drafter list, a technical committee, and a project cycle in months. At that moment no technical content exists in any approved form — only an agreement to go and write it.
The industrial agent governance entry specifies its scope: governance objectives and principles, an overall framework, role responsibilities and core governance domains, plus requirements covering core functions, the full lifecycle, and risk, security and trustworthiness. The drafter list is broad, mixing research institutes, universities, steel and aircraft manufacturers and software vendors. A parallel entry for Part 4, on industrial skill repositories, ran the same notification window.
A 15-month project cycle is a plan, not a deadline. Drafting, internal review, public comment on the actual draft, technical examination and approval each consume time, and items slip. Until a number is assigned and an implementation date is published, the correct way to describe this document is "under development".
The other pipeline: MIIT's industry standard plan
Industry standards run on a parallel track under individual ministries. MIIT's third 2026 batch is a useful illustration of scale: 121 projects in total, 66 new and 55 revisions, spread across emerging industries, traditional industry upgrades, future industries and foundational work.
Seven sit under artificial intelligence, and their titles tell you what the ministry thinks is underspecified:
- 人工智能 产品服务 大模型服务推理性能测试方法 — a test method for inference performance of model services
- 人工智能 产品服务 搜索智能体能力技术要求 — capability requirements for search agents
- 人工智能 产品服务 知识计算智能体技术要求 — capability requirements for knowledge-computing agents
- 人工智能 赋能应用 电力知识计算与生成式大模型协同技术要求 — power-sector knowledge computing with generative models
- 人工智能 关键基础技术 大模型检索增强生成技术要求与评估方法 — requirements and evaluation for retrieval-augmented generation
- 人工智能 关键基础技术 企业智能体管理能力评估模型 — an assessment model for enterprise agent management capability
- 人工智能 关键基础技术 智能体运营平台技术要求 — requirements for agent operations platforms
All seven are 制定 (new), all marked voluntary, all with a 12-month planned cycle. Note the shape of that list: four of seven are about agents, and the ones that are not are about measuring RAG and inference. The signal is that "can it answer" is being separated from "can it be operated, governed and evaluated" — which is exactly what an enterprise buyer should care about.
Voluntary is not the same as irrelevant
A 推荐性 (voluntary) standard creates no legal obligation. But in Chinese procurement practice, voluntary standards get written into tender documents, referenced in contracts, and used as the technical baseline in disputes. They function as the common language of a market even when nobody is compelled to follow them.
The trap is the reverse error: treating a voluntary, not-yet-written standard as a compliance certificate. A vendor that claims conformance to something still in drafting is either confused or selling. A buyer that writes "must comply with GB/T XXXX" into a contract before the number exists has written an unenforceable clause.
Honest limitations
The facts here come from the standards platform project pages and MIIT's published plan document. I have not seen any draft text of the industrial agent governance standard, because none has been published — so I cannot tell you what its governance domains will require, and neither can anyone outside the drafting group. Project cycles are planned durations, not promises; both instruments could be delayed, split, or dropped. I also cannot confirm whether further parts beyond Part 1 and Part 4 exist, since I only verified those two entries. This piece covers standards-making mechanics; it does not cover the separate question of whether an industrial agent triggers safety, cybersecurity or data obligations under other rules, which is a different analysis.
What readers can do now
- Before signing anything, look the standard up on the national standards information platform and check its status field. "国家标准项目" with a notification window means drafting; a published standard carries a number and an implementation date.
- Write acceptance criteria you can actually test: retrieval accuracy against a fixed question set, permission boundaries, human review for high-impact actions, and what happens when the model, knowledge base or tool permissions change. Those requirements hold regardless of which standard eventually appears.
- Participate if you have skin in the game. Drafting groups on these projects include vendors and end users, and the window for influencing text is before public comment, not after.
Why the distinction keeps getting lost
Standards announcements travel faster than standards. A project title with an official-looking platform entry reads like a finished rule to anyone who does not work with the pipeline daily, and vendors have an incentive not to correct the impression.
According to Li An, chief scientist at BrainNet (脑机网), the useful question about any AI standard in China is not whether it exists but which of the four stages it occupies — project, draft, issued, or in force — because each one supports a completely different claim.
For industrial agents, the honest answer today is stage one. The governance requirements are being argued over right now, in a room, by the organisations listed on that platform page.