NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Beijing Wrote Rules for AI Agents While Enterprises Were Running Pilots

China's agent rules, a finance security standard and a software action plan landed within months. The bottleneck for enterprise AI is now permission, not models.

2026-11-16 · 1192 words · NeuroAI
Beijing Wrote Rules for AI Agents While Enterprises Were Running Pilots

On 8 May 2026, three Chinese agencies published a joint document about agents. Written in the flat register of Chinese industrial policy, it was one of the first national frameworks aimed at agents (智能体) as a product category rather than at large models (大模型). The Cyberspace Administration of China, the National Development and Reform Commission and the Ministry of Industry and Information Technology signed it together.

Three and a half months later, a finance-industry standard told banks and phone makers who is allowed to click "pay". Two weeks after that, the industry ministry set numerical targets for agent software inside the country's software industry.

Read together, the three documents describe something more interesting than regulation: a market where the hard part of selling an agent is no longer whether the model is good enough, but whether the agent is allowed to touch the system it is meant to operate.

A policy written for agents, not models

The May document, the Implementation Opinions on Standardized Application and Innovative Development of Agents (智能体规范应用与创新发展实施意见), does three things that matter to anyone building or buying agents.

First, it names the use cases. Agent deployment is organised around five directions — scientific research, industrial development, consumption, livelihood, and social governance — and 19 specific scenarios. That is a procurement map as much as a policy statement.

Second, it sorts agents into tiers. Policy analysts describe a positive-list and negative-list approach: some applications are encouraged, some are restricted, and the classification turns on capability and risk rather than on which model the agent runs.

Third, it separates who decides. An October interview published by Legal Weekly (法治周末) describes three decision modes in the framework: the user decides, the user authorises the agent to decide, and the agent decides on its own. The first two map onto consent practice that already exists; the third is new territory.

For a product team, that is the operational point: the permission model is not an implementation detail, it decides whether the product can be sold.

The permission problem became a product problem

The clearest example came from banking. In late 2025, a handset launched with an on-device assistant that could operate other apps for the user. Within days, several banking apps would not log in or complete payments, and the assistant's financial functions were pulled within a week.

The industry's answer arrived on 27 August 2026: a group standard on agent security for financial applications, issued by the Beijing Fintech Industry Alliance (北京金融科技产业联盟). It requires what the industry calls dual authorisation — an agent operating a financial app needs permission from both the user and the institution. Screen reading and simulated clicks are out unless the app itself opens the door.

The drafting list includes Postal Savings Bank (邮储银行), ICBC (工商银行), China UnionPay (中国银联), Bank of China (中国银行) and Bank of Communications (交通银行), alongside Huawei, Ant Group, Volcano Engine and Tencent Cloud.

This is why China's agent race looks different from America's. US agent products mostly assume the open web; Chinese ones assume app-by-app integration, where the target app's developer issues the permission. Tencent, Alibaba and ByteDance each own super-apps, which makes their agents easier to build and harder to open.

Money is arriving more slowly than the rules

The forecasts are enormous: an IDC report with Inspur, released in September 2026, projects that active agents worldwide will grow from 79.4 million in 2026 to 2.216 billion by 2030, with China's agent count compounding at 151.2% a year. The same report estimates China's AI computing market at US$51.5 billion in 2026, up 37% year on year, on the way to US$150.1 billion by 2030.

Software revenue is arriving more slowly than those numbers imply. China's software industry booked RMB 15.48 trillion (≈US$2.15 trillion / HK$16.8 trillion) in 2025, a base so large that a new category takes years to appear in the total. Daily token (词元) calls passed 500 trillion by June 2026 — usage is real, but usage is not willingness to pay.

Baidu's own numbers make the gap visible. Its AI Applications segment, which includes its office agent products, grew 3% year on year in the second quarter of 2026, even as its cloud infrastructure line grew 50%. Baidu launched an enterprise edition of its DuMate agent in June, and says a self-evolving agent it released this year has drawn trial applications from about 3,000 companies. Applications and trials are not revenue yet.

The procurement shortlist is being written now

On 11 September 2026, MIIT published its action plan for "AI plus software". The targets are unusually concrete: by 2028, reach 20,000 above-scale software companies (规模以上软件企业), run 100 intelligent retrofit projects, build 100 benchmark agent software applications in key industries, and incubate at least five open-source projects. By 2030, the ministry wants key software "fully intelligent".

Two details matter more than the targets: an instruction to build agent software app stores and a skills library, and a push for model-as-a-service and agent-as-a-service delivery, with software shifting from product delivery to value delivery.

If MIIT publishes that list of 100, it becomes the closest thing state-owned and large enterprise buyers have to a pre-approved shortlist. Vendors who treat the plan as a press release will be reading a procurement document two years late.

What this means outside China

China is one of the first major markets to write agent-specific rules nationally, and the choices matter even for people who will never sell there.

The most portable idea is the split between what the user decides and what the agent decides. Most agent products elsewhere still treat authorisation as a single consent screen. Dual authorisation for financial apps effectively says that the user's consent is not enough when a third party's system is the one being operated — a position banks elsewhere are likely to reach on their own, probably after an incident.

The least portable part is the delivery model. App-by-app permission requires leverage over app developers, which in China means the platform companies. Elsewhere, the equivalent leverage sits with Apple and Google, or with nobody at all.

Honest limitations

This article draws on policy documents, a group standard, company filings and one IDC report. It does not assess enforcement: we found no evidence of penalties under the May framework, and the tiering lists have not been published in full. Scenarios and decision modes are described as policy analysts present them, not from a line-by-line reading. Vendor adoption figures are self-reported, and agent-count forecasts have a poor record in every market. No regulator or executive was interviewed.

What readers can do now

  • If you sell agents into China, map your permission model to the three decision modes before your next product review, and be ready to show audit logs for every action an agent takes on a customer's system.
  • If you buy agents, require dual authorisation wherever money moves, and insist on a human confirmation step that cannot be configured away by an administrator.
  • If you build agents for other markets, borrow the design question rather than the rule: which actions should require consent from the user, and which should require consent from the system being acted on?

Related coverage

Cite this article

NeuroAI (脑机网), "Beijing Wrote Rules for AI Agents While Enterprises Were Running Pilots," 16 November 2026, https://neuroai.site/a/na-corp-china-agent-rules-enterprise

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Companies & Stack” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories