Who is betting on Personal AI, and who is afraid of Personal AI?

Over the past month or so, the domestic Personal Agent sector has suddenly become crowded.

Alibaba is advancing Qwen Go, Tencent is exploring an independent personal Agent, ByteDance continues to extend toward system-level Agents on phones, StepFun has launched an agent phone, Ant has upgraded Lingguang, and Today AI has secured more than $200 million in funding. Internet giants, model companies, and startups have all turned their attention to personal AI assistants almost simultaneously.

But interestingly, just as various companies were racing to launch their own Personal Agents, on October 14, the list of third-party apps refusing automated operations disclosed by Doubao Phone Assistant included a large number of high-frequency internet applications such as WeChat, Taobao, and Meituan.

This precisely reveals a contradiction in China's Personal Agent industry: every company wants AI to become the user's new entry point, but they are not necessarily willing to hand over their existing entry points to someone else's AI.

Therefore, what is more worth pondering than whether Personal Agent will become the next hit application is this: is it actually creating a new AI market, or is it redistributing the users, traffic, and transactions that already existed in China's internet?

And this may be the true starting point for understanding the current competition in China's Personal AI.

From Chatbot to Personal AI, the collective pivot of the big tech companies

Over the past few years, the consumer-facing competition among China's large model vendors has gone through three rounds of goals.

First, competing on model capability: from 2023 to 2024, companies chased each other on answer quality, reasoning, search, and multimodality, and leaderboard scores were almost the entire topic of conversation.

Then, competing on user scale: from the price war in 2024 to the free-access shock brought by DeepSeek in early 2025, large model apps opened up for free one after another, and together with large-scale advertising, they expanded the pie.

In the third round, competing on task delivery: in 2025, Manus popularized the Agent concept, and AI began connecting tools and invoking services, attempting to move from giving answers to delivering results; by 2026, Qwen completed 120 million AI orders during the six days of Spring Festival—having AI complete transactions on people's behalf is no longer an engineering challenge.

This round of change is not a feature upgrade of Chatbot, but a change in the goal of competition: in the past, companies fought over a user's single question or single search, and once the answer was over, the value was realized; now what they are fighting for is the opportunity to continuously deliver tasks and needs for users. The more tasks someone handles, and the longer they handle them, the deeper their understanding of the user, and the harder this relationship is to replace.

What new value does Personal AI actually create?

It can be broken down into two kinds of demand.

One is incremental demand: things that no one could satisfy in the past because they were complex, trivial, and required long-term memory. Flight delay compensation is a typical example—to file a claim, one had to queue for customer service and haggle over the plan, and many people simply gave up after calculating the cost. When the Agent itself checked the order, contacted customer service, refused the voucher and insisted on cash, and returned 70 yuan to the user's account, the abandoned demand was reactivated; this kind of demand did not fail to exist in the past, but the cost of handling it was higher than the benefit. The Agent pushed the cost close to zero, and service calls and transactions that would not have happened originally did happen.

The other is substitution of existing demand: booking tickets, shopping, price comparison, and reservations were previously completed by users themselves in apps, and now they are handed to the Agent to execute. This only changes the path and does not create new transactions.

The proportion of the two determines the direction of the industry. If the incremental part is large enough, this is a new market; if it is mainly migration of existing demand, then it is redistributing the traffic and user relationships of existing platforms. The answer is still unclear: Sensor Tower data shows that Meta's Muse surpassed 2.5 million downloads 13 days after launch, with initial daily active users close to three times that of ChatGPT in the same period; but TickerTrends tracking shows that its average daily net new users subsequently fell from a September peak of 198,000 to 74,000, a decline of 62.4%. Curiosity has been proven, but long-term high-frequency use has not.

The competitive logic of entry points has changed

Once Personal AI takes hold, the competitive logic of entry points will change accordingly.

In the past, the core logic of platforms was to attract users into their own apps and complete search, selection, and transactions within the platform; Personal AI attempts to split this step apart: users express their needs to AI, and AI decides which services to invoke and which platform to use.

The object of competition has changed from "which app users stay in" to "who understands needs, organizes services, and decides the execution path." In other words, what is truly noteworthy about Personal AI is not that there is one more AI application, but that it may separate the demand entry point from the service delivery platform—and this separation opens up space for the next round of internet competition.

Different approaches in the same competition

Once entry points may be separated, all parties must make their moves. Looking back at the actions of the past month or so in the context of each company's own foundations, there are five different paths: what each wants to gain and what it needs to protect determine its approach.

Alibaba's answer is Qwen Go: unveiled on October 13, embedded in the Qwen App rather than being a standalone application. The first batch of demonstration scenarios—flight delay compensation, Taobao price protection, and car maintenance inquiry and booking—have long chains and many variables, and all fall within Alibaba's own ecosystem: compensation funds go into Alipay, and price protection is completed on Taobao.

At the Yunqi Conference in September, Qwen announced that it would build a Personal Agent based on the Agentic capabilities of Qwen 3.8 to serve 300 million users; previously it was already able to handle tasks in dozens of areas such as shopping, travel, and logistics, and Qwen Go expands the radius to long tasks that require repeated checking, communication, and comparison.

What is special about Alibaba is that it simultaneously holds personal Agent and trading platforms, covering the complete chain from demand to order. Qwen Go is more like an integrator of the old ecosystem than a new entry point—putting Taobao, Alipay, Amap, and Fliggy into a single conversational entry point. The question is, as search, recommendation, and transactions are increasingly completed by AI, how will Taobao's bid ranking be priced?

Morgan Stanley predicts that China's monetizable consumer AI market will be about 300 billion yuan by 2030, with revenue still mainly from advertising and commissions—the pie may get bigger, but the people collecting the money may change. Keeping transactions within its own system is the key to Alibaba's answer to this question. There are 23 Alibaba-affiliated apps on the rejection list: its own Agent can touch Taobao, but others' cannot.

Unlike Alibaba's bet on transactions, Tencent's trump card is connection. With WeChat in hand, Tencent has still opened up multiple Personal Agent fronts outside WeChat.

At the end of September, the cloud-based personal assistant LightVela launched, connected through WeChat and online all day; in October, it was reported that Tencent was developing a standalone app "Yuanbao Assistant," giving it running space and permissions independent of the main Yuanbao app, to carry long-cycle tasks and long-term memory; the PC-side Marvis also upgraded its positioning in late September from "computer repair" to "AI butler," turning toward managing personal digital assets. A Citi report forecast that Tencent will also update the products and strategies of Hunyuan Hy4, WorkBuddy, and Xiaowei at the end of the month.

Within WeChat, the moves are restrained. In June, WeChat released guidelines for mini-programs to connect to the AI ecosystem, and JD.com, Meituan, Ctrip, and Didi officially announced integration; the native assistant "Xiaowei" began gray testing the same month, able to send messages and invoke mini-programs to place orders, with payment still confirmed by the user, and in September it began internal testing of "Xiaowei AI Social." These moves point in the same direction: services go through mini-programs, payments go through WeChat Pay, and identity and relationships stay in WeChat.

WeCom is another entry point: on October 14, WorkBuddy entered the WeCom contacts as a bot, invoking ten major capabilities including messages, documents, meetings, and cloud drives within the scope of enterprise authorization; QuestMobile September data shows it is already the top office Agent by monthly active users.

The two lines run in parallel because WeChat is both Tencent's greatest advantage and its constraint: 1.4 billion monthly active users, relationship chains, and the service loop are all in WeChat, but this also means that every mistake an Agent makes within WeChat will directly transmit to Tencent's core business.

So among the five, Tencent's posture is the most contradictory—it is closest to the core scenarios of Personal AI, and its trial-and-error cost is also the highest, so it places more aggressive product forms outside WeChat, while WeChat itself only maintains small-scale testing within the gray-release scope.

Tencent's difficulty is that its entry point is too precious, while ByteDance's difficulty is that its entry point is not deep enough—its problem is not users; Doubao has over 100 million monthly active users and ranks among the top two domestic AI applications alongside DeepSeek; what it lacks is permissions.

Doubao Phone Assistant debuted last December with ZTE Nubia's engineering phone M153, and this September the consumer version was released, with the NaviX Ultra carrying it going on sale starting at 5,999 yuan, with nearly 400,000 reservations; internally, the personal assistant "Xiaodou" is still in internal testing. The phone assistant takes the GUI route: visual recognition to read the screen and simulated clicks, not relying on app vendors to open interfaces, but its operations bypass the apps' original interaction and distribution systems and easily run into risk controls—when the technical preview launched, there were incidents of WeChat accounts being forcibly logged out and Taobao login being blocked.

On October 14, Doubao Phone Assistant announced the results of the 30-day public notice of the SAEP protocol, and also disclosed the rejection list: 155 apps, covering almost all high-frequency scenarios such as communication, social networking, e-commerce, travel, and payment, including 66 Tencent-affiliated and 23 Alibaba-affiliated; AI applications such as Qwen, Yuanbao, WorkBuddy, DeepSeek, and Kimi also reject each other.

ByteDance bets on both apps and phone systems because standalone apps cannot obtain execution permissions; only system-level entry points can read and write the screen and schedule other applications. But this also makes it the party most dependent on others' openness: cross-app task execution is its core selling point, and this premise is precisely in the hands of the 155 apps that reject it.

The value of the path it bets on is built on the openness of other ecosystems, and the rejection list shows that others do not intend to open up.

ByteDance is moving from apps toward systems, while StepFun bundled models, systems, and terminals together from the start.

On July 13, it released the AI terminal brand STEPX, the agent operating system Step AOS, and the personal agent Amoo; on October 13, the first agent phone STEPX Neo was officially released, priced at 3,999 yuan.

Contrary to Doubao's GUI route, StepFun takes the API cooperation route: calling service interfaces opened by partner vendors, with the first batch connecting Ctrip, Alipay, Didi, Meituan, and others—when a user raises a travel need, Amoo can book itineraries, call a car, and make payments, stringing together a complete task chain. This route does not touch the core interests of app vendors and is easier to turn into long-term cooperation, but the pressure of ecosystem building all falls on StepFun: whether WeChat, Taobao, and Douyin are willing to open up directly determines how many tasks Amoo can complete.

Why does a model company build phones? The large-model API business has been dragged into a price war; StepFun has no super app and no traffic like ByteDance, so combining software and hardware is one of the few feasible paths for it to obtain a direct user entry point. It does not rely on hardware profits, does not do pre-installed ads, and sets no hard shipment targets;

Chairman Yin Qi said: "Hardware is just a container; in the end, everyone still uses that Agent service." The cost is that the agent phone category has not yet been validated, and the premise for this bet to hold remains the openness of top apps.

A big company either has an ecosystem it can integrate or a user base large enough to distribute traffic to. Today AI has neither; the only thing it has is an identity that does not depend on any platform. It was disclosed on October 21 that the company has raised a cumulative $200 million in funding, with investors including IDG Capital, Tencent, 5Y Capital, Ant Group, Sequoia China, and Transsion Holdings. Founder Qi Junyuan was previously Vice President of Product at Feishu and head of Doubao's PC client, and earlier founded Teambition.

Today AI has nothing except one identity: it does not depend on any platform. It was disclosed on October 21 that the company has raised a cumulative $200 million in funding, with investors including IDG Capital, Tencent, 5Y Capital, Ant Group, Sequoia China, and Transsion Holdings. Founder Qi Junyuan was previously Vice President of Product at Feishu and head of Doubao's PC client, and earlier founded Teambition.

The domestic version of Today AI launched on September 25, focused on continuous observation and long-term memory: it proactively records itineraries, project summaries, and health conditions, and connects to applications such as WeCom, WeRead, Baidu Netdisk, Kingsoft Docs, and Tencent Meeting through Connectors.

When big companies build a Personal Agent, they all have to answer "what happens if we eat into our own app's pie"; Today does not need to protect any existing business model and can turn itself into an assistant that does not depend on a single platform. But its cross-platform capability relies on Connectors, and the Connectors are all held by the platforms.

What is more subtle is the capital structure: Tencent and Ant are both advancing their own Agents while both being its shareholders—this can be understood as hedging, and it also means Today AI must keep its distance between the respective ecosystem interests of its two major shareholders.

Five paths, pointing to the same question

The five companies' paths are not mutually exclusive, nor are they necessarily final strategies, but together they show that competition in Personal AI has shifted from comparing model capabilities to comprehensive competition across models, user relationships, service connections, operating permissions, and existing ecosystem resources.

Alibaba bets on transactions and fulfillment, Tencent guards high-frequency connections, ByteDance bets on distribution and terminal permissions, StepFun moves from models to operating systems and devices, and Today AI bets on cross-platform context. Existing advantages are the starting line, and also the constraints on the next stage of transformation.

Personal AI reshapes the internet, and the biggest variable may be China's super app ecosystem

In the end, all five approaches run into the same external condition: China's super app ecosystem. To see how unusual this wall is, we can first look at an overseas comparison.

The wall overseas, and the wall in China

On September 8, Meta released the personal agent Muse. It runs in an independent virtual machine in the cloud; users assign tasks through a standalone app or WhatsApp, and it can browse the web, fill out forms, place orders, and book trips on its own; at launch it connected partners such as Spotify, OpenTable, Ticketmaster, and Plaid, and for services without partnerships it can still operate the browser like a human.

Overseas internet services are relatively fragmented, and APIs and browsers leave a foundation for Personal AI to connect. Even so, conflicts still occur: on September 20, Amazon banned Muse's ability to place orders on users' behalf, citing unauthorized access, suspected storage of credentials, and failure to identify itself as AI—behind this is $68.6 billion in annual advertising revenue; when an Agent bypasses the homepage and search to place orders directly, ad slots lose their buyers.

China's situation is more closed. Overseas, it is scattered conflicts on fragmented services; in China, it is a super app system: WeChat and WeChat International have more than 1.4 billion monthly active users, Douyin and Taobao are both 900-million-level entry points, and services, transactions, payments, and user relationships are aggregated inside a few platforms, with data isolated from one another. The rejection list of 155 apps is the most direct presentation of this barrier.

So in China, ecosystem integration capability beyond model capability may be the more critical element.

A contradictory competition

Platforms face a dilemma. If third-party Agents help users compare prices and place orders, consumers no longer need to go through the e-commerce platform's full search, recommendation, and advertising path—the platform may gain new transactions, but may also lose traffic distribution and commercial revenue; if third-party Agents are restricted, the platform may miss the opportunity to be invoked by the next-generation entry point.

What is more complicated is that big companies' own Agents are also disrupting their own apps: Qwen Go bypasses Taobao's search and recommendations, and the future Yuanbao assistant will bypass service distribution within WeChat. So what Personal AI brings may not be "big companies versus startups" competition, but rather a self-transformation that the giants actively promote and must also handle cautiously.

The law has not provided a ruling either. In August, the U.S. Court of Appeals for the Ninth Circuit vacated Amazon's preliminary injunction against Perplexity, on the grounds that it is the user themself who accesses the server, and AI is merely a tool—the conflict between Agents and platforms will ultimately have to be resolved at the commercial level. Some platforms are also looking for a third path: WeChat does not choose openness or closure, but acts as its own gateway, with services going through mini programs and payments going through WeChat Pay. Meituan CEO Wang Xing said that in addition to serving consumers and merchants, serving AI Agents is becoming increasingly important. If this path holds, future competition will not be "open AI versus closed apps," but "who becomes the gateway that must be passed through."

The value chain may be cut apart again

Future industry roles can be divided into three layers: the user relationship layer, who continuously understands users and takes on personal needs and long-term goals; the intelligent orchestration layer, who decides which model, which tool, and which service provider to invoke; and the service fulfillment layer, who ultimately provides real services such as goods, content, payments, and transportation.

In the past, super apps Integrating these three layers within a single platform is precisely the foundation of its business model; Personal AI may gradually pull them apart. The real question becomes: will the user relationship, the power to assign tasks, and the transaction revenue belong to different companies in the future? The rejection list of those 155 apps is the first round of answers from all sides.

Four questions that still have no answers

First, cross-platform capability: can Agents continuously and reliably connect to a sufficiently broad range of services? SAEP and API partnerships are only the beginning.

Third, business sustainability: Morgan Stanley estimates that Meta's average annual revenue per user is about $70, while the annual service cost corresponding to a user consuming only 5% of the free quota reaches $124, exceeding the revenue contributed—whether the costs can be covered is something no one has yet proven.

Fourth, distribution of ecosystem interests: the Agent is responsible for selection, the platform is responsible for fulfillment—who gets the traffic, data, revenue, and user relationships? This is the hardest one to resolve, because it is a question of interests, and questions of interests are precisely the kind that China's super app ecosystem is best at creating and least willing to concede on.

Back to the question at the beginning: who is betting on Personal AI, and who is afraid of Personal AI? The answer may be the same group of people. The bettors see the possibility of a restructuring of the entry point, while the fearful worry about the depreciation of the existing entry point; these two identities now overlap in almost all mainstream Chinese internet companies.

The real barrier in this competition may not lie in how much information about users AI remembers, but in whether it can reliably, stably, and at low cost transform a person's personalized needs into services completed across platforms. In China, this is both a technical problem and, even more, an ecosystem and business problem.

This article comes from the WeChat public account "AI Large Model Works," author: Qima, published by 36Kr with authorization.