A European company that sells or operates AI agents in Chinese commerce now knows the framework Beijing intends to apply to it. It does not yet know which regime it will fall under: registration, testing and recall of defective products in "sensitive areas and key sectors", or simple self-assessment in "low-risk" areas. The fifteenth measure of the notice published on 19 October applies to commerce agents the "tiered" governance that Beijing set in May for all agents. It does not draw this boundary. In Europe, the AI Act does not classify a shopping guide agent as a high-risk use.
China's Ministry of Commerce (MOFCOM) posted online on 19 October 2026 a notice on accelerating "科技兴商", which can be translated as "making commerce prosper through technology". The document, referenced 商建函〔2026〕551号, is dated 13 October. Nine administrations sign it: the Ministry of Commerce, the Office of the Central Cyberspace Affairs Commission, which is one and the same as the Cyberspace Administration of China (CAC), the National Development and Reform Commission, the ministries of Education, Science and Technology, Industry and Information Technology, Culture and Tourism, the People's Bank of China and the State Administration for Market Regulation.
Its 18 measures cover most of domestic commerce. They range from flexible manufacturing to home services, via logistics and online commerce. The State Council Information Office presented it the same day at a press conference. According to the report by China News Service, Assistant Minister of Commerce Yuan Xiaoming described the text as the first "comprehensive and systematic" document devoted to this theme. On 16 October, the same ministry had proposed subjecting intelligent agents to export controls.
Shopping guide, operations, logistics: agents named
The tenth measure asks, according to ActuIA's translation, to "cultivate new forms of activity around commercial intelligent agents" (商业智能体) and to "develop specialised agents such as the intelligent shopping guide, intelligent operations and intelligent logistics". The same measure wants to bring embodied AI to fruition in commerce, with home services and commercial logistics as application areas. The text cites these three functions as examples, without a closed list. They involve advising a customer at the time of purchase, managing the operations of a business and organising the flow of goods.
Measure 15: a sentence taken from the May framework
The fifteenth measure, entitled "perfect the safety and regulation governance system", consists of three sentences. Here is ActuIA's full translation:
"According to different application scenarios, prudently and methodically carry out tiered governance of commercial intelligent agents. Clarify the boundaries and rules for the collection, use and sharing of commercial data, explore the publication of a negative list of data uses, establish a monitoring and early warning system for data security risks, and strengthen protection and governance capabilities. Apply inclusive and prudent regulation, allow, within a controllable perimeter, stress tests and sandbox experiments of new technologies and new economic models, and refine regulation."
The first sentence is almost word for word from a text posted online on 8 May 2026 by the CAC: the implementation notice on the regulated use and innovative development of intelligent agents, whose execution is entrusted to the CAC, the National Development and Reform Commission and the Ministry of Industry. Only the mention of potential impact is missing. Its eleventh measure asks to carry out tiered governance of agents "according to application scenarios and potential impact", and outlines two regimes. In "sensitive areas and key sectors", cyberspace authorities, together with the relevant ministries, set the open scenarios, and apply "registration, testing and recall of defective products". In certain "low-risk" areas such as daily life, leisure or office work, governance is through compliance self-assessment, information declaration, management by distribution platforms and industry self-regulation.
The same text proposes, in its fourteenth measure, to explore a reliability rating system established by professional organisations, with voluntary participation, which would rate in particular "incitement to consume" and "false advertising". Failures would be sanctioned according to law. It already cites, among the uses to develop, customer service agents available 24 hours a day and embodied agents for guidance, cleaning, warehousing or distribution in catering, retail, accommodation and logistics.
The October notice therefore applies an existing architecture to commerce. The negative list of data uses is not published: the verb used is "explore". The text plans to authorise sandboxes "within a controllable perimeter", without describing their regime. The "inclusive and prudent regulation" (包容审慎监管) invoked is a standard formula of the Chinese administration towards new activities.
A guidance notice, without sanctions or timetable
A "意见" is a guidance notice addressed to administrations and local authorities, of lower status than a law or regulation. This one provides for no sanctions and sets no timetable. Its conclusion entrusts the Ministry of Commerce with coordination with other administrations and the dissemination of "typical experiences". As of 22 October, the ministry's page displays the text with the mention "现行有效", in force.
The measure appears in the sixth part, "optimise the development environment for commercial technologies", alongside standards, data infrastructures and support for start-ups. This classification makes the governance of agents a condition for deployment, placed among the tools that should facilitate it. The thirteenth measure also uses tiers, of a different nature: it asks to "promote the application" of the series of national standards on the tiered classification of the intelligence of AI terminals. Published on 8 May 2026 by the Ministry of Industry, the State Administration for Market Regulation and the Ministry of Commerce, among others, according to China News Service, this series (GB/Z 177-2026) defines intelligence levels and test methods for seven initial product categories, including phones, computers and car cockpits. These tiers assess capabilities. They serve to compare products and guide supply.
Compared with the AI Act, a similar logic, a different legal force
The AI Act, as summarised by the European Commission, classifies AI systems into four risk levels (unacceptable, high, transparency, minimal) according to their use. This horizontal regulation applies to all sectors. Regulation (EU) 2026/1744, known as the digital omnibus on AI, which entered into force on 27 July 2026, postponed to 2 December 2027 the application of the rules targeting high-risk systems in sensitive areas (biometrics, critical infrastructure, education, employment, migration), as ActuIA detailed.
A shopping guide agent does not appear as such among the high-risk cases listed by the Commission. If it converses with the customer, it falls a priori under transparency obligations, applicable since August 2026. It must inform the user that they are interacting with a machine. The same agent would change category if it were used to decide a customer's access to credit, a use that the Commission classifies as high-risk.
On paper, the two systems start from a similar logic: by use and by impact, with a lighter regime for low-risk uses and sandboxes for experimentation, which the omnibus extends across the Union according to the Commission. The method and legal force differ. In sensitive sectors, Beijing leaves it to the authorities to set the open scenarios, without a list of prohibitions. In Brussels, the levels, prohibitions and obligations are set out in a regulation with sanctions. The Chinese notices, without their own sanctions or deadlines, refer to existing laws and regulations for registration and product recall.
On 16 October, agents proposed for export controls
The same ministry published on 16 October, with the Ministry of Science and Technology, announcement no. 48 of 2026, which puts out for consultation a revision of the Chinese catalogue of technologies whose export is prohibited or restricted. It targets 40 technologies, including 23 in AI: large models, intelligent agents (the same term, 智能体), computing clusters and information processing. Technologies already published in open source or in scientific articles are excluded. Without waiting for the revision, any transfer or licence abroad of the technologies listed in the annex must be registered in advance with the ministry since the day of the announcement, including ongoing operations (ActuIA analysed this project).
Placed side by side, the two publications outline a coherent position. Domestically, the agent is a business model to be grown under "inclusive and prudent" regulation. Externally, the technology that powers it would become a controlled asset. The exclusion of technologies published in open source or in scientific articles, however, limits the scope of the project for foreign publishers that rely on open Chinese models. The consultation on the export catalogue is open until 15 November 2026.
Our articles will then appear first in Google Top Stories.