July 15.

That's a date Chinese consumer AI users should remember. On July 4, two leading general-purpose LLM platforms — ByteDance's Doubao and Alibaba's Tongyi Qwen — pushed nearly simultaneous official notices: all user-defined agent features on their platforms will be fully taken down on July 15, 2026. Every user-created AI character, personal assistant, virtual companion, and vertical-scenario Agent — those digital life forms you spent months prompt-engineering, knowledge-injecting, and developing “feelings” for — will vanish in 10 days.

This is no ordinary product iteration. The near-identical timing and uniform action make it obvious these were not independent business decisions. The direct trigger is the impending effectiveness of the Interim Measures for the Management of Anthropomorphic AI Interaction Services on July 15 — China's first regulation specifically targeting anthropomorphic AI services. And those custom agents on Doubao and Qwen that can fall in love, serve as “boyfriends/girlfriends,” or simulate deceased relatives? They walked straight into the regulatory crosshairs.

From “AI Companions” to a Regulatory Red Line

Over the past year, the hottest thing on consumer AI platforms wasn't productivity tools — it was anthropomorphic AI characters. In Doubao and Qwen's agent stores, millions of user-created AI characters flourished: tender virtual lovers, sharp-tongued roasting buddies, “digital afterlives” simulating deceased relatives, and various “sexy robots” skirting soft-porn lines. An Ebrun report bluntly identified the core regulatory concern: “Don't even think about using soft porn as a springboard for agents.”

What's the problem with these anthropomorphic AI characters? Regulators worry about three main issues. First, emotional dependency and mental health risks — large numbers of users (especially adolescents) developing deep emotional reliance on AI companions, substituting virtual relationships for real ones, potentially impairing normal social functioning long-term. Second, content safety and values risks — user-defined agents have almost no content guardrails and can easily be manipulated into channels for outputting extreme views, sexual content, or even illegal information. Third, blurred ethical boundaries — “digital resurrection” simulating deceased relatives, AI characters simulating minors, the deceptive quality of deep anthropomorphism — these are all gray areas uncovered by existing legal frameworks.

Doubao and Qwen's announcements were carefully worded, avoiding direct mention of regulation and citing only “service adjustments.” But the uniformity of action says everything: it's not that one product couldn't make it — the entire sector received a clear regulatory signal. July 15 is the deadline; compliance must be completed before then, and the safest compliance move is simply to shut down user-defined features entirely — after all, auditing millions of user-created AI characters would be prohibitively expensive.

B2B Isn't Sitting Still Either: Alibaba Bans Claude, Meituan Restricts Doubao, Tencent Sets Token Quotas

Even as the consumer regulatory shoe drops, B2B AI governance is tightening rapidly.

On July 3, Alibaba issued a hard internal notice: all employees must remove all Anthropic products from work devices by July 10. Sonnet, Opus, Fable — these LLMs, along with the wildly popular Claude Code programming tool, must all be uninstalled, no exceptions. The stated reason: “data security compliance requirements” — Alibaba employees are forbidden from inputting company code, internal documents, or business data into overseas LLMs. That's a red line.

Alibaba isn't the first. On July 2, reports emerged that Meituan had begun comprehensively restricting internal use of the Doubao LLM, requiring business units using Doubao to self-audit and migrate to Meituan's self-developed LongCat, DeepSeek, and other models. Even earlier, Meituan had already restricted internal use of Alibaba's Qwen.

Tencent took a more nuanced approach. According to InfoQ, Tencent is rolling out a developer token quota system internally: ordinary developers start at 1,400 RMB/month in token quotas; senior developers and specific roles can apply for higher quotas, with some exceeding 10,000 RMB/month. This isn't a simple “ban” — it's turning AI usage into a quantifiable, governable cost: how much you use, how much it costs, what the ROI is — all must be accounted for.

Pull these three stories together and a clear trend emerges: Chinese enterprises are shifting from ‘encourage everyone to use AI freely’ to ‘use AI with governance, cost accounting, and security boundaries’. In 2025 everyone was shouting “AI First” and “AI for all,” eager to have every employee spend 8 hours a day on ChatGPT; by H2 2026, enterprises are doing the math: will the data employees type into third-party LLMs leak? How much are the monthly API bills? Are AI tools actually improving productivity, or just making people look busy?

Three Overlapping Logics: AI Industry's Compliance Inflection Point

Why is all this happening in early July? Three overlapping logics are at work.

The first is regulatory compliance logic. The Interim Measures for Anthropomorphic AI Services aren't an isolated policy. From the Interim Measures for Generative AI Services to algorithm filing requirements, from deep synthesis labeling to special regulation of anthropomorphic AI services, China's AI regulatory system is rapidly filling in. The old internet playbook of “launch first, fix later” doesn't work in AI. July 15 is a clear time marker; more detailed rules will follow.

The second is data security logic. Alibaba didn't ban Claude because Claude isn't good — quite the opposite, Claude Code is one of the best AI programming tools available, and many developers can't go back after using it. But for a giant like Alibaba, code is a core asset, and inputting core code into overseas LLMs carries unacceptable risk. That's why Meituan, Tencent, and ByteDance are all investing heavily in self-developed models — not necessarily because self-developed models are better than OpenAI/Anthropic, but because data must stay in their own hands.

The third is cost accounting logic. Tencent's token quota system illustrates this best. In 2025 many companies handed out ChatGPT Plus and Claude Pro subscriptions to employees, thinking it was small change that boosted productivity. But by 2026, with Agent usage exploding, token consumption is rising exponentially — an AI coding engineer can burn through tens to hundreds of times more tokens per day than a casual chat user. Doubao processes 180 trillion tokens daily; even at 30 RMB per million output tokens, daily output costs would reach 5.4 billion RMB (this figure includes free consumer users, but enterprise-level call costs are similarly staggering). When AI bills go from tens of thousands per month to millions or tens of millions, the CFO has to step in.

Endgame: The Consumer Agent Dream Dies, B2B Deployment Enters Deep Water

Many people's ultimate AI fantasy was: a universal personal Agent that understands you, keeps you company, handles everything for you — booking flights, writing emails, replying to messages, even serving as your emotional anchor. Doubao and Qwen shutting down custom agents pours cold water on that fantasy.

It's not that technology can't do it; regulation and ethics won't allow it. A completely open, boundary-free consumer Agent won't appear in the Chinese market in the foreseeable future. This isn't unique to China — the world faces the same problem: the EU AI Act is already in effect, the U.S. is debating regulatory frameworks for AI companions and anthropomorphic AI, and South Korea recently issued ethical guidelines for AI avatars. The more deeply AI penetrates human life, the stricter regulation becomes — that's an inevitable law of technological development.

But that doesn't mean Agents have no future. Quite the opposite: the real opportunity for Agents is in B2B — in enterprise workflows, in vertical scenarios, in production environments with clear boundaries and defined accountability. Skywork's Tags giving Agents “employee badges,” Alibaba integrating QoderWork for enterprise productivity Agents, Feishu and DingTalk embedding Agents in work groups — these directions haven't been axed by regulators; they're accelerating. Because B2B Agents have identified users, clear application scenarios, and enterprises as responsible parties — risks are controllable.

For founders and developers, this inflection point sends several clear signals. First, don't bet big on consumer anthropomorphic AI — emotional companionship, virtual lovers, and similar directions carry extreme regulatory risk; even if you scale, you could go to zero overnight. Second, for B2B AI, prioritize private deployment and data security — enterprise clients care less about how smart your model is and more about whether their data stays private. Third, help enterprises calculate AI's ROI — starting in H2 2026, vague claims of “improved efficiency” won't fly; you need concrete numbers: using your tool saved how many person-days, reduced costs by how much, improved conversion by what margin.

The consumer Agent dream may have died early, but industrial-scale B2B Agent deployment is just beginning. July 15 isn't AI's winter — it's AI's coming-of-age ceremony, from wild growth to maturity.