When AI Wants to Be Your New Portal
The Battle for User Entry Points in China's Tech World
On December 1, 2025, something remarkable happened in China’s tech ecosystem. ByteDance released its Doubao Phone Assistant—not a physical smartphone, but system-level AI software that could take over any Android device. The first batch of 30,000 engineering units, priced at 3,499 yuan and bundled with ZTE’s nubia M153, sold out within a single day. On secondhand platforms, scalpers pushed prices to 7,999 yuan. Some enterprising individuals even started renting these phones at 600 yuan per day. For a moment, it seemed like the future of human-computer interaction had arrived.
Then, just 24 hours later, the walls closed in. On the evening of December 2, users began reporting that their WeChat accounts had been suspended for “abnormal login environment.” By December 3, multiple banking apps and Alibaba’s suite of applications blocked access from Doubao-enabled devices. On December 5, ByteDance was forced to issue an official statement, announcing the suspension of all financial operations, gaming functions, and point-accumulation features. The most ambitious AI assistant in China had run straight into the battlements of the country’s super apps.
Zhou Hongyi, the outspoken founder of 360 Security, captured the essence of what had happened: “ByteDance isn’t making phones—they’re trying to steal the souls of every smartphone out there.” His observation cuts to the heart of a question that goes far beyond any single product: Why would an AI assistant trigger such an aggressive response from China’s most powerful digital platforms? The answer lies not in technology, but in a war over user entry points—a battle that determines who controls how billions of people access the digital world.
Two Paths to the Same Goal
This contest is playing out through two dramatically different strategies. If ByteDance’s Doubao chose to scale the walls of China’s app ecosystem, another player—Alibaba’s DingTalk—has been quietly building roads.
Doubao’s approach relies on what’s known as a GUI Agent: the system uses a visual AI model called UI-TARS to literally “see” what’s on the screen, then simulates human taps and swipes to operate any application. This requires no cooperation from app developers, no API integration, no permission whatsoever. The AI can compare prices across shopping platforms, automatically apply coupons, and complete purchases—all by mimicking what a human user would do. It’s technologically elegant, and to the platform owners who depend on controlling user attention, existentially threatening.
DingTalk’s strategy couldn’t be more different. On December 23, 2025, the company unveiled what it calls Agent OS—billed as “the world’s first intelligent operating system designed for AI in the workplace.” Rather than bypassing other applications, DingTalk invites them in. Over 100,000 developers, 140,000 enterprise applications, and more than 3 million companies now connect through authorized APIs. The company’s founder Chen Hang put it in characteristically modest terms: the system’s codename is “Mulan,” representing “a bud about to bloom.”
The contrast between these approaches reveals a deeper truth about how AI is reshaping digital ecosystems:
WeChat’s response to Doubao was hardly surprising. As Zhou Hongyi observed, platforms like Taobao and Meituan monetize through user attention and advertising. When an AI assistant completes tasks directly—skipping the browsing, the recommendations, the sponsored content—the entire business model collapses. WeChat remains, in Zhou’s words, “the most conservative and stubborn defender of the entry-point logic.” Since April 2024, even Huawei’s and Xiaomi’s built-in AI assistants have been unable to access WeChat’s core functions.
The struggle for user entry points is nothing new. The browser wars of the 1990s, Microsoft’s antitrust battles, Google Chrome’s eventual dominance—each platform shift has triggered fierce competition for the gateway through which users access everything else. In China, the mobile era produced a distinctive pattern: the rise of super apps. WeChat and Alipay, with their mini-program ecosystems offering some 200 public services, created a duopoly that compressed potential competitors before they could emerge. Now, AI agents threaten to redraw these boundaries entirely. Google DeepMind CEO Demis Hassabis has declared 2025 “the true start of the agent-based era.” According to research from early 2025, 63% of websites have already received traffic from AI agents.
The AI Gets a Body
If DingTalk’s softer approach has avoided the confrontations that stymied Doubao, it’s not merely a matter of diplomatic strategy. The company has made a genuine technical bet on what it means for AI to operate in the physical world.
Agent OS represents a fundamental reconception of the AI’s role in enterprise work. Traditional AI assistants function as “copilots”—helpful supplements to human decision-making. DingTalk’s agents are designed as what Chen Hang calls “silicon employees”: autonomous entities with their own employee badges, specific permissions, and 24/7 availability. The system manages not hardware resources, as a traditional operating system would, but “intelligent resources” and “enterprise data.”
The most striking manifestation of this philosophy is DingTalk Real—what the company describes as the world’s first enterprise-grade AI hardware built specifically for agents. This isn’t a consumer gadget but an industrial device designed for deployment in manufacturing plants, energy facilities, and other complex environments. It features multimodal sensing capabilities including cameras and microphone arrays, enabling it to perform visual inspections, equipment monitoring, and real-time data collection.
The practical applications are already emerging across Chinese enterprises. At Ningde Times, the world’s largest battery manufacturer, AI agents now handle visual inspection tasks around the clock. At Mengniu’s dairy farms, agents monitor livestock health and feeding schedules. DingTalk’s “Wukong” agent can interpret natural language requests like “schedule a meeting with the person I couldn’t reach last week,” then autonomously coordinate with relevant parties to find suitable times. A recruitment agent can collect resumes from across the internet, screen candidates, and schedule interviews—completing in 20 minutes what would take a human recruiter hours.
Managing the Risks
As AI agents gain the ability to act autonomously in the physical world, China has been developing governance mechanisms to match. The approach reflects a consistent pattern: pilot programs first, legislation to follow.
On December 15, 2025, just days before DingTalk’s announcement, the Ministry of Industry and Information Technology issued China’s first national product approval for L3 autonomous vehicles—cars that can legally take over driving responsibility from humans under specific conditions. Two models received permits: Changan’s Shenlan SL03 for roads in Chongqing, and BAIC’s Arcfox Alpha S for designated highways in Beijing. The approval followed two years of city-level pilots that allowed regulators to observe real-world performance before setting national standards. The shift in liability—from driver to system—required careful legal groundwork that the pilots made possible.
This “pilot first, legislate later” approach extends to AI more broadly. In March 2025, four Chinese government agencies jointly issued the “Measures for Labeling AI-Generated Synthetic Content,” requiring clear identification of AI-generated text, images, and video across the production and distribution chain. The regulation took effect in September, addressing concerns about deepfakes and misinformation with specific technical requirements rather than blanket prohibitions.
DingTalk Real embeds this governance philosophy directly into hardware. Most intriguing is what the company calls the “physical safety valve”—a kill switch that allows operators to cut power to the AI system instantly. Chen Hang was characteristically direct about its purpose: “If things ever get out of control, DingTalk provides an emergency function: pull the plug.” The device also processes data locally—”data doesn’t leave the factory”—addressing corporate concerns about sensitive information flowing to cloud servers. These design choices acknowledge real anxieties about AI autonomy while offering tangible solutions.
The pattern across these developments—autonomous vehicles, content labeling, hardware kill switches—suggests a pragmatic framework: enable innovation in controlled environments, observe what happens, then codify the lessons into regulation. Whether this approach scales beyond China’s particular institutional context remains to be seen. But it represents one answer to the question of how to govern AI systems that can act on their own.
What Comes Next
The Doubao saga ended, for now, with a concession. ByteDance pulled back from its most aggressive features, and the super apps held their ground. But Li Liang, ByteDance’s vice president, framed the outcome philosophically: “Whether this particular attempt succeeds or not, AI is definitely the future.”
He’s probably right. The technical capability to build AI agents that can operate any application—with or without permission—now exists. The question is not whether someone will try again, but under what terms. DingTalk’s enterprise approach offers one model: authorized access, clear boundaries, physical safety valves. The consumer space remains contested territory.
For now, the super apps have won a battle. Whether they’ve won the war is another matter entirely.