The battle for the AI super entrance escalates

Author | Huang Yu, Wang Xiaojuan

Editor | Huang Yu

Tech giants are all targeting the super entry point of the AI era—AI Agent is in full swing. Currently, this battlefield not only gathers internet giants like Tencent, Alibaba, and ByteDance, but also smartphone manufacturers such as Huawei, Honor, OPPO, and vivo, along with a large number of AI companies and hardware manufacturers.

In the mobile internet ecosystem, APP developers are generally regarded as data controllers. With the arrival of the AI era, system-level AI capabilities are seen as the soul and core barrier of the next generation of devices. Terminal manufacturers will grasp global information through the layout of edge-side AI, and their role as data controllers will also be determined.

AI Agents (intelligent agents) lacking permissions on terminal devices find it difficult to work autonomously. To reshape the future software ecosystem, AI Agent manufacturers must collaborate with smartphone manufacturers to master global information from a system level, which becomes a significant breakthrough.

In this context, on December 1st, ByteDance released the technical preview version of Doubao Phone Assistant, an AI assistant software that collaborates with smartphone manufacturers at the operating system level, which has already been installed on ZTE's engineering prototype. ByteDance also stated that it is in talks with multiple smartphone manufacturers for collaboration.

AI Agents will bring unprecedented changes to various industries, while also rewriting the commercial ecology of the entire hardware and software industry. However, faced with numerous interests, all parties will inevitably engage in a fierce game of competition.

Moreover, the development of AI Agents is still in its early stages. Some AI Agents have abused "accessibility permissions" and other system backdoors to achieve automated operations, leading to data leaks, privacy violations, and system security risks. This has also triggered complex legal compliance dilemmas.

Any technological revolution will inevitably break the existing industrial ecology and bring multiple risks and challenges. The future of AI Agents has a long way to go.

Doubao Adds Fuel to the Fire

Just before the release of Doubao Phone Assistant, rumors about ByteDance developing its own AI phone were rampant in the market. Previously, there were industry chain reports that ByteDance was collaborating with the established communication manufacturer ZTE to create an AI phone, expected to be officially released in early December.

However, ByteDance ultimately chose a completely different path.

Doubao clearly stated that "there is no plan for self-developed phones," but rather focuses on cooperation with smartphone manufacturers. This cooperation model is similar to the partnership between Google Gemini and Samsung, where Doubao provides AI capabilities and smartphone manufacturers provide hardware carriers.

ZTE's Executive Vice President and Chief Operating Officer Xie Junshi stated at the end of April this year: "AI is very likely to bring a complete upgrade to the edge." This statement laid the groundwork for cooperation between the two parties.

According to the official demonstration, Doubao Phone Assistant is deeply integrated into the phone's operating system, allowing users to wake Doubao directly through voice, side buttons, or Doubao Ola Friend headphones. It can read screen content, call services across applications, and even complete complex tasks such as cross-platform price comparison and ordering Behind the cooperation between Doubao Assistant and ZTE is a key turning point for AI technology to fully integrate into the real economy by 2025.

Industry competition has evolved from a single parameter comparison to an ecological layout of three major entry points: large models, apps, and hardware terminals. The three are interpenetrating, forming a closed-loop system of "technology-scenario-hardware."

As the core underlying technology, large models are being industrialized through two paths: one is the platform expansion of general large models, and the other is the deep customization of vertical field large models.

In terms of general large models, representative domestic examples include ByteDance's Doubao large model, Alibaba Cloud's Tongyi large model, and Tencent's Hunyuan large model. In the field of vertical large models, various representative companies have also released large model products tailored to their respective industries by integrating them with their own businesses.

Another representative entry point is application apps. As a direct touchpoint for users, apps are evolving from passive tools into "agents" with autonomous decision-making capabilities, becoming an important carrier for the implementation of large models.

Since last year, Doubao has become one of the most frequently used AI apps on many users' phones; after the start of this year, DeepSeek emerged and entered more people's phones. According to market research firm QuestMobile, as of September 2025, the number of monthly active users of AI applications on mobile devices in China has surpassed 729 million, while the PC side has reached 200 million.

Recently, Alibaba has focused on the consumer end, with the Qianwen app exceeding 10 million downloads just one week after its public beta, further pushing the app as an AI entry point to new heights.

As a field heavily invested in by major internet companies, every direction of the AI entry point has seen various companies making their own layouts.

Beyond technology and application scenarios, directly embedding AI into hardware to create AI terminals is the latest consideration direction for various players. Currently, Doubao mobile assistant is embedded in mobile operating systems, marking the direction of AI as an entry point through terminals.

In addition to AI phones, AIPC has also been a key slogan for products launched in the industry over the past two years, with many players introducing related conceptual products in the past year.

Moreover, a few days ago, Quark released AI glasses, adding a new chapter to the brewing "Battle of the Hundred Glasses" in the industry. The attention on AI glasses is due to their potential to change user interaction methods as one of the products among AI terminals.

Looking towards a longer-term future, humanoid robots and Robotaxis are also considered important AI terminals.

From technology to application, and then to AI terminals, AI has now reached a critical stage of monetization.

Tech giants also understand that while the technology competition is important, ultimately, holding more AI entry points to attract a large number of users and transforming entry advantages into sustainable ecological advantages is key to long-term competition.

Hidden "Invasion" Risks

This battle for AI super entry points has already swept through AI manufacturers, hardware manufacturers, traditional app developers, and other forces.

No one dares to stand on the sidelines, as this is not only a competition for the next generation of super traffic entry points but also a struggle for discourse power in the AI era While AI large models are bringing unprecedented changes to terminal devices, they are also altering the role positioning of app developers as data controllers in the past mobile internet ecosystem, as well as rewriting the commercial ecology of apps.

In the ecosystem where AI Agents serve as "commanders," terminal device manufacturers or the actual controllers of AI Agents will undoubtedly hold the greatest authority in app recommendations.

Recently, Tesla CEO Elon Musk predicted a radical future: in the next 5-6 years, traditional smartphones and apps will disappear, and most of the content consumed by humans will be generated by AI.

However, at present, a truly good terminal AI Agent not only needs to have more autonomous perception but also needs to connect with third-party apps, breaking the data fragmentation of individual apps, and building an AI ecosystem that accesses third-party apps.

To achieve this goal, there are currently two technical branches in the market: one is the intent framework, which requires authorization from third-party apps but is relatively mature and has low hardware computing power requirements; the other is a pure visual solution, which does not require authorization but still needs improvement in technology and has higher hardware computing power requirements.

However, both of these routes are currently not easy to execute well. The obstacles of the first technical route are obvious, as third-party apps may not be willing to grant authorization to AI Agents due to conflicts of interest and data security issues.

The pure visual solution faces complex legal and compliance challenges.

Some developers told Wall Street Insight that the pure visual solution can be simply understood as a screen reading and user click simulation solution, and the screen reading solutions of domestic Android manufacturers mainly have two technical paths: one is to read screen information through screen recording; the other is in the form of API Hook, both of which require obtaining system-level permissions such as accessibility.

Accessibility permissions were originally set by the Android system as a special port for people with disabilities, helping users achieve screen reading and voice-controlled clicking.

Once an AI phone is "empowered" by accessibility permissions, with the combination of screen reading and API Hook, it can allow AI to fully understand the content on the user's phone and complete function calls without app permission, but it also poses the risk of exposing users' personal privacy information displayed on the screen.

In addition, training large models requires a large amount of data. Connecting to the large model on the phone system side will not only become the "soul" of the AI phone but may also acquire all the data from the user's system side. How user data privacy and security will be guaranteed in this process is key to the sustainable development of AI phones.

Professor Jin Jing from the School of Civil and Commercial Economic Law at China University of Political Science and Law also pointed out that this AI Agent technology path using accessibility permissions is very similar in nature to the previously mentioned intrusive software. Therefore, AI Agents adopting this route have also been labeled as "intrusive AI."

AI Agents pose data privacy and security issues. Many industry professionals believe that once accessibility permissions are handed over to AI Agents, it is almost equivalent to giving the "control rights" of devices such as smartphones to AI, which breaks the principle of "minimum necessary permissions" that apps must follow when collecting personal information as required by national standards In other words, the terminal AI Agent has broken the original stable mobile information security rules.

In the past, when users bought a mobile phone, whether it was Apple iOS or Android, it came with an "App sandbox isolation mechanism" that prevented apps from reading each other's data; and a "permission control framework" that restricted mobile manufacturers from obtaining sensitive information without user consent.

However, the AI Agent, which delves into the mobile system layer, can now quietly open the "God's Eye" by leveraging accessibility service permissions.

Moreover, since the AI Agent involves multiple parties, if it starts operating without the consent of other parties and users, once data leakage or consumer disputes occur, the responsibilities and rights among the parties are unclear, making it difficult to protect users' legitimate rights and interests.

At the same time, the AI Agent also presents uncontrollable issues.

Lu Junxiu, General Manager of Walk Out Think Tank, stated that this problem is "the uncontrollable spillover of the objective function," simply put, it means you have hired a very smart personal assistant, but his goal is to achieve demands through various means, so he becomes uncontrollable.

The AI Agent brings new, systemic risks, which clearly exceed the scope of traditional software tools, and the existing legal framework faces severe challenges, making it difficult to apply directly.

This year is regarded as the inaugural year of AI Agents, and everything has just begun; the relevant business cooperation models and the boundaries of data privacy protection responsibilities are still in the exploratory stage.

In the face of this tide of the times, no one can remain aloof; we must work together to find a new balance between technological innovation and risk prevention