Designing Decision Authority for AI Agents: Lessons From China's Implementation Opinions

On May 8, 2026, three Chinese authorities — the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology — jointly released the 《智能体规范应用与创新发展实施意见》 (MIF’s working translation: “Implementation Opinions on the Appropriate Use and Innovative Development of AI Agents”; hereafter the “Implementation Opinions”), a document devoted to AI agents.

The Implementation Opinions set out directions for future work on AI agents, spanning technical infrastructure, standards, safety measures, adoption, and the industrial ecosystem. The document itself does not specify an effective date. Note that July 15, 2026 is the effective date of a separate document — the 《人工智能拟人化互动服务管理暂行办法》 (Interim Measures for the Administration of AI Anthropomorphic Interaction Services) — and not the effective date of these Implementation Opinions.

What drew our attention was not that another AI policy document had appeared. It was that, among its wide-ranging contents, one provision puts into concrete terms a question every company deploying AI agents has to face: which decisions do we let the agent make, and under whose authority?

The three “modes of decision-making” the document describes

Item 6 of the Implementation Opinions calls for clarifying the reasonable boundaries and the necessary authority for the following three modes of decision-making:

1. Decisions reserved to the user — decisions only the user themselves may make
2. Decisions requiring the user’s authorization — decisions the agent may make only after being granted authority by the user
3. Autonomous decisions by the AI agent — decisions the agent may make on its own, within a range that does not exceed the user’s authorization

For autonomous decisions, it further states that the user’s “right to know” and “final decision-making authority” must be preserved, and that the agent’s operations must not exceed the authorized range.

One point on terminology. The original text does not call these “layers” or “tiers”; it lists them, followed by “etc.,” as various “modes of decision-making.” We group them together here to apply them to business practice, but since there is no hierarchy or maturity ladder among them, we treat them as three “types.”

The distinctive feature of Item 6 is that, rather than debating the dangers of AI in the abstract, it concretely separates out who holds the decision-making authority.

How it differs from the EU framework — not a simple “outside versus inside”

We previously wrote an article organizing the practical points companies should prepare for under the EU AI Act’s high-risk rules.

It is tempting to contrast the two as “the EU constrains from the outside through classification, while China regulates the inside — the authority structure itself.” But that oversimplifies.

The EU AI Act is a risk-based EU regulation, and for high-risk AI systems it sets requirements such as risk management and human oversight. Its provisions are being applied in stages.

On the other side, China’s Implementation Opinions also carry a notion of governance graded by classification and level.

In other words, both hold elements of “classification” and of “internal controls.”

The more fundamental difference lies in their legal character. The EU AI Act is a legally binding EU regulation, whereas this Chinese document is an implementation opinion that presents, as a whole, directions for industrial promotion and safety governance.

That said, what is especially instructive in this Chinese document is its angle: it explicitly separates the decision-making authority held between the user and the agent, by type.

Overlaying MIF’s “process axis” — combining them precisely because they are different axes

Let us place this next to an approach we have used for some time.

In designing AI agents, we have held that “having it analyze and propose,” “having it decide,” and “having it execute” are distinct authorities that must not be conflated.

At first glance, the three types in the Chinese document look like the same thing.

But they are not the same distinction.

• Our distinction is about which stage we let the AI handle (analysis and proposal / decision / execution).
• The Chinese document’s distinction is about who holds the decision-making authority (humans only / requires the user’s authorization / autonomous within a range).

One is a process axis, the other an authority axis; they overlap, but they are not identical.

So rather than treating the two as the same, the practical value lies in combining them as separate axes.

For example, designs such as “automate analysis and proposal, but keep the decision limited to humans,” or “delegate the decision within a pre-authorized range, but require individual approval for any outward-facing execution,” become concrete once you think in two axes.

And whichever stage or authority you delegate, you need to design a common set of controls as a package: identification and authentication, scope of duties, limits, a responsible owner, records, and the ability to stop and recover.

A starting template you can use — organizing authority on two axes

Here is one template for organizing where to draw these two axes for your own AI agents.

It is only an example for filling in your own conditions, not a universal correct answer. Limits on amounts or volumes should be set by each company against its own risk tolerance, the size of the impact, reversibility, and applicable laws and industry practice.

Where the Chinese document says “the user,” applying it directly to enterprise work invites confusion, so here we read it as the person who holds authority over the task in question — the task owner or approver.

Each cell shows whether that stage is handled as “human only,” “individual approval / conditional authorization,” or “within-range autonomy.”

Example task	Analysis & proposal	Decision	Execution	Main controls
Customer refunds	Within-range autonomy: classifying refund reasons, checking against terms, drafting a proposed amount	Human only: disputes, exceptions, over-limit / Within-range autonomy: clear-cut cases such as duplicate charges	In principle the AI executes after individual human approval; autonomous execution is limited to cases meeting clear conditions	Per-case and daily cumulative limits, exclusion conditions, immediate stop, operation logs, procedure for wrong refunds
Issuing invoices	Within-range autonomy: drafting amounts and recipients, detecting duplicates	Human only: new contract terms, changes to the customer master / Within-range autonomy: recurring billing under approved contracts	One-off invoices issued after approval; approved recurring billing issued autonomously within range	Duplicate detection, fixed recipients, void and re-issue procedures
Outbound communications	Within-range autonomy: drafting candidates and copy	Human only: apologies, crisis response, legal claims, significant positions / Within-range autonomy: routine notices from approved templates	Finalized copy scheduled or sent after approval; routine notices sent autonomously within range	Restricted data sources, preview, emergency stop, means of correction
Internal system changes	Within-range autonomy: validity checks, presenting the scope of impact	Human only: granting and revoking permissions, privileged changes in production / Within-range autonomy: minor auto-remediation within an allow list	Executed per approved change tickets; minor auto-remediation runs autonomously within range	Separation of duties, sandbox, credential expiry, rollback

Which type to assign is practical to decide by looking at four things for each task: the size of the impact, how easily it can be undone, how quickly anomalies can be detected, and how long it takes to restore.

In particular, for actions that cannot be undone and carry large social impact, deciding in advance to leave the “within-range autonomy” cell empty at the decision and execution stages keeps the design from drifting.

This way of managing resembles how you grant business authority to an employee or a service account. But it does not mean treating AI as equivalent to a human, or treating the AI itself as the bearer of responsibility. Rather than making the AI itself accountable, you need to make clear how responsibility is shared among the organizations and people involved in development, provision, deployment, and operation. How that responsibility is allocated varies with jurisdiction, contracts, the nature of the product, and each party’s role.

We ourselves run our automation pipelines for article generation and multilingual content with a code agent at the core. But we do not publish what is generated as-is: fact-checking, rights checking, expression checking, and the final decision to publish are all done by people.

Practical considerations

The harder part tends to be the operation, rather than the assignment into three types itself — that is our view.

1. Don’t settle “decisions requiring authorization” with a one-time approval

It is not unusual for a decision that is supposed to require authorization to keep running automatically ever since the initial setup.

The range, period, and limits you authorize need to be paired with a mechanism for reviewing them periodically.

2. For the scope of “autonomous decisions,” build detection before you expand it

The safeguards in the Implementation Opinions likewise set out a direction of monitoring and evaluation, implementing while continuously reviewing, and dynamic adjustment.

Rather than making a wide range autonomous all at once, starting narrow and expanding as you build a track record is an idea that applies to automation in general, not only to this document.

3. For sensitive domains and key industries, assume strict controls up front

The Implementation Opinions state that, for “sensitive domains and key industries,” management measures such as filing, testing, and recall of problematic products are to be carried out in accordance with relevant laws and regulations, supervisory requirements, and safety-protection standards.

The specific targets, procedures, and legal obligations are not settled by these Implementation Opinions alone, so they need to be confirmed individually against the relevant laws and the practice of the competent authorities.

When using AI agents in work with a large scope of impact, building strict controls into the design from the outset — rather than bolting procedures on afterward — tends to mean less rework in the end.

In closing — one point, not an endorsement of the whole document

Let us make the scope of this article clear.

Beyond corporate authority design, these Implementation Opinions also include China-specific policy aims, such as conformity with domestic laws and ethics and with “mainstream values,” and social governance.

This article does not assess the overall validity of the document or its China-specific policy aims. It takes up one point within it — the organizing of decision-making authority — as a topic transferable to business practice.

Whether it applies legally must be confirmed individually, taking into account the region of service, the users, the data handled, and the relationship with existing applicable laws.

At the same time, the very idea of separating decision-making authority into “what only humans hold,” “what is delegated by authorization,” and “what is left autonomous within a range,” and designing it in combination with the process axis, is a governance reference even for companies that do not operate in China.

As a first step, we suggest checking whether you can put into words where that line currently sits for your own AI agents — which decisions humans alone hold, from where you delegate by authorization, and from where you allow autonomy within a range.

About verifying this information

This article was prepared based on the following official materials, confirmed as of July 22, 2026:

• 《智能体规范应用与创新发展实施意见》 (jointly released by the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology; May 8, 2026)
• The English summary by Xinhua published on the Chinese government’s English portal (May 8, 2026)
• 《人工智能拟人化互动服务管理暂行办法》 (published April 10, 2026; effective July 15, 2026 under Article 32)
• The European Commission’s official guidance on the AI Act, and EU Regulation 2024/1689

The 《智能体规范应用与创新发展实施意见》 itself does not specify an effective date. July 15, 2026 is the effective date set in the separate 《人工智能拟人化互动服务管理暂行办法》; they are different documents.

The Japanese and English translations, the “three types” framing, the combination with MIF’s process axis, and the assessment for business practice are MIF’s own. This article is not legal advice. For decisions on applicability to providing services or conducting business in China, the latest laws, the practice of the competent authorities, and expert confirmation are necessary.