The two most-used consumer AI apps in China are deleting the agents their users built. ByteDance’s Doubao and Alibaba’s Qwen, the country’s first- and second-largest AI apps by monthly active users (QuestMobile, first quarter 2026), will disable user-created agents between July 10 and July 15 and render the underlying data unrecoverable after October 15 (TechNode). The proximate cause is a new Chinese rule on humanlike AI, but the rule is only the occasion. The response to it is the signal, and it points at a cost that Western retail has decided does not apply to its own march into agentic commerce.

The rule targets companion software, not shopping carts. China’s Cyberspace Administration and four other agencies published the Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services in April, effective July 15 (The Next Web), governing services that “simulate human personality traits, thinking patterns and communication styles to provide sustained emotional interaction” (SCMP). Customer-service bots, knowledge Q&A and workplace assistants are carved out, so long as they avoid that sustained emotional bond (The Next Web).

What the deletion says

What the two platforms did with that rule is the tell. Neither rebuilt its agents to comply; both deleted the category. Doubao is retiring its entire agent feature, citing “product function adjustments,” while Qwen is pulling its humanlike personas, its user-created agent functions and its broader agent services alike, with no migration path offered (SCMP). A rule aimed at emotional companions could have been met by stripping out the emotional part. Instead the market’s two largest consumer platforms removed the user-agent layer wholesale, which is a statement about what that layer was worth to them once a cost was attached to keeping it. Pan Helin, a member of the MIIT expert committee, put the moment plainly: current agents, he said, “are not yet mature” (SCMP, as reported by The Next Web).

No filing calls these agents unprofitable — but a company that deletes a feature rather than rebuild it has told you what the feature was worth without printing the number.

The same machine

Western retail is wiring its checkout to the same class of consumer-facing autonomous agent, on the working assumption that the cost China just refused to pay stays external. The commitment is not rhetorical. Google, Shopify, Target, Walmart and Etsy have signed a shared protocol to let agents transact (Google), and Target, Sephora, Nordstrom and Wayfair have wired their catalogues into OpenAI’s (OpenAI). The bet is that an agent acting on the shopper’s behalf is a channel; the unexamined clause is that it stays a cheap one.

The Chinese rule does not travel; the cost structure it exposed does. An agent that acts for a user, persistently and at scale, builds a compliance and liability surface that grows with its autonomy, and someone eventually prices it. In China the price took the form of anti-addiction systems and identity checks for minors (The Next Web). The Western version is arriving through different doors: access law and product liability. A federal court found, in a preliminary injunction, that a shopping agent needs the platform’s own authorization to transact, and the shopper’s consent does not suffice (Search Engine Journal); YouGov found that 74% of US adults blame the company, not the model, when a chatbot gets something wrong (YouGov). The externality differs; the mechanism does not.

Where the objection runs out

The strongest objection to all of this is that the analogy breaks at the root. The Chinese rule governs companionship and spares the task bots retail deploys, so a coat-finding agent sits outside it and no regulator has yet moved against one. For the thesis to fail, that wall has to hold: the cost of autonomous consumer agents would need to stay penned inside the emotional category and never cross into the transactional one. It is not holding. The Chinese platforms did not excise the companionship and keep the commerce; they deleted the whole user-agent layer, because judging which agent is “emotional enough” for each user costs more than deleting the category outright. And the Western cost is already crossing over, through the access and liability doors a shopping agent has to walk through anyway.

The invoice

The real question under the agentic-commerce build-out was never whether agents can sell. It is whether the cost of running them stays external, and courts and regulators decide that, not retailers. China’s answer took one rule and a fortnight to produce, and the two companies with the most to gain from consumer agents chose to walk away rather than pay for them. If a Western court or regulator makes a comparable call on access, liability or data retention, and one court has already made it on access, the retailers wired deepest into agentic checkout will find out whether they built a channel or a liability. A platform can delete a feature and move on; a brand that has routed its checkout through an agent cannot delete the shopper’s memory of who sold her the wrong coat. That memory is the cost being budgeted at zero, and a regulator on the other side of the world has already written the first invoice.