On 10 July, Alibaba's Qwen disabled its humanlike and user-created agent functions, with its broader agent services following on 15 July. ByteDance's Doubao goes the same way on 15 July, redirecting affected users to a separate companion app called Maoxiang, and Tencent's Yuanbao removed a comparable feature back in June. The South China Morning Post's reporting ties all three moves directly to China's Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services, issued by the Cyberspace Administration of China alongside four other regulators in April and taking effect on 15 July. Millions of users lose access to accumulated chat histories and custom agent settings in the process, with Qwen confirming permanent deletion and no migration path.
What the rule is actually aimed at
The measures apply to services that, in the regulator's own language, mimic a real person's personality, thinking and way of speaking to sustain an ongoing emotional relationship with the user. That's a narrower target than "AI agents" as a category, and CRN Asia's analysis is clear that the rule explicitly carves out customer service bots, knowledge Q&A tools, workplace assistants, education software and internal copilots, provided they avoid sustained emotional engagement. Services that clear a scale threshold, more than a million registered users or 100,000 monthly active users, face additional obligations regardless: security assessments across eight defined areas, algorithm filings with provincial regulators, and app store verification. Enforcement is already visible on the ground. Shanghai's internet regulator reported on 26 June that it had stripped more than 14,000 non-compliant agents from platforms, citing bots impersonating officials, inappropriate role-play content, and unauthorised harvesting of personal data.
Why "we don't sell companion apps" isn't the same as being exempt
It's tempting to read this as a China-specific consumer-app story with no bearing on enterprise AI deployments, and for most organisations reading this, the letter of the rule bears that out. But the exemption is defined by what a product does, not by what category its vendor puts it in, and the design trend in enterprise assistants over the past two years has been pushing steadily in the direction the rule is meant to catch. Brand-voiced chatbots with a name, a persona and a consistent conversational tone, wellness or coaching assistants that check in with employees or customers over time, internal AI companions positioned as always-available colleagues rather than tools: these are increasingly common product choices, made for good engagement reasons, and several of them sit closer to "sustained emotional interaction" than most product teams would assume when they chose the design. An assistant built to remember a user's mood across sessions and respond with warmth and continuity is doing, functionally, exactly what the rule describes, regardless of whether anyone on the product team ever used the word "companion."
A third axis of AI regulation to track, not a duplicate of the ones you already know
Enterprises running AI deployments across multiple jurisdictions are already tracking the EU AI Act's risk-tiered obligations and the shifting landscape of US export controls and deregulation. This rule doesn't fold neatly into either. It isn't about risk classification by use case, and it isn't about where a model is trained or hosted. It's a behavioural design rule: does this product simulate an ongoing emotional relationship, yes or no. That makes it a genuinely separate compliance axis for any organisation with a Chinese subsidiary, a Chinese user base, or product teams building assistants intended for the Chinese market, and it's a useful prompt for every other organisation too, since "does our assistant's design cross into sustained emotional engagement" is a reasonable question to ask under most jurisdictions' emerging AI rules, not just China's.
The practical, unglamorous urgency
Set the design-philosophy question aside for a moment, because there's a harder deadline underneath it. If any part of your organisation, a regional team, a contractor, a business unit operating in China, has built internal tooling, workflows or integrations on top of Doubao or Qwen's agent features specifically, those features stop working on 15 July, and in Qwen's case the underlying data is being permanently deleted with no migration path offered. Doubao is marginally more forgiving, with read-only access preserved until 15 October before deletion follows its standard privacy policy. Either way, this is the kind of shadow dependency that rarely shows up on an architecture diagram until the day it breaks, and it's worth a direct, dated question to any team that might have quietly wired a consumer agent platform into something the business now relies on.
- Ask directly whether any team, subsidiary or contractor has built internal tooling on Doubao's or Qwen's agent features, and treat "probably not" as an answer that needs verifying, not accepting.
- If dependencies exist, export or migrate anything on Qwen before 15 July and anything on Doubao before its 15 October deletion date.
- Review your own customer- or employee-facing assistants against the "sustained emotional engagement" test, independent of whether you operate in China, since the design pattern it describes is spreading globally.
- Track this rule as a distinct compliance axis alongside the EU AI Act and US export policy, rather than assuming your existing AI governance framework already covers it.
- Revisit persona-driven or companion-style AI product designs with legal input before launch in any jurisdiction with anthropomorphic-AI rules in development, since China is unlikely to be the last regulator to draw this line.
The headline version of this story is that China banned AI companions and enterprise agents dodged the bullet. The more useful version is that a major regulator has now written down, in specific and testable terms, exactly what separates a tool from a companion, and that definition is worth measuring your own AI products against even if you never expected a Chinese law to be the reason you did it. Want help mapping your AI product portfolio against emerging anthropomorphic and agentic AI rules across jurisdictions? Email sales@halfteck.com.