Summary of Key Points
Recently, smartphone manufacturers such as Nubia and Samsung have successively launched "agent phones" equipped with on-device AI models, featuring AI that can perform practical tasks for users, such as booking flights and comparing prices (rather than simply providing answers or image editing services). However, AI phones currently face a critical bottleneck: despite their many functions, they have not become a compelling reason for users to upgrade their phones. The main obstacles are not insufficient computing power, but whether AI can obtain the necessary permissions from app developers to perform these tasks on their behalf, as well as how to balance the commercial interests and responsibilities of all parties involved. The essence of this competition is a renegotiation of the order of how users interact with their phones—whether AI agents can become the primary point of contact for users' needs, replacing traditional apps.
1. AI phones are not lacking in useful features, but in features that make users consider an upgrade essential
AI phones are no longer a novelty; features like image editing, voice transcription, and large-scale model-based queries have become standard on mid-to-high-end devices. According to Counterpoint, nearly half of all phones in 2026 will have generative AI capabilities, with high-end models likely to have them all. The problem is that, while these features are useful, they are not sufficient to motivate users to upgrade their phones.
For example, when you want to book a flight, you still have to open the Ctrip app, enter the dates, and select the flight yourself; when you want to compare prices, you need to use multiple apps. AI can at most provide you with information but cannot complete the entire process for you. The core issue for users remains unaddressed: "I don't want to learn how to use each app separately; I just want to say, 'Book the cheapest flight to Shanghai for tomorrow,' and then wait for the result."
Nubia's NaviX Ultra aims to overcome this by allowing AI to directly perform tasks instead of just providing answers. However, its success will depend on how frequently users use it. If it is only used occasionally, it will not be a sufficient reason for a phone upgrade.
2. On-device intelligence is not as simple as just putting large models on the phone
Many people think that "on-device AI" means compressing cloud-based models and installing them on the phone, focusing on the performance of the NPU (Neural Processing Unit) and the amount of memory. In reality, it is much more complex. For instance, Nubia's DouBao Assistant only stores locally sensitive models like OCR (Optical Character Recognition) and personal data retrieval for quick responses to prevent privacy breaches, while more complex tasks (such as analyzing the best flight options) still rely on cloud processing. This approach, known as "edge-cloud collaboration," is both secure and efficient. The phone acts more like a "permissions manager," determining which data the AI can access (such as the contact list) and which apps it can interact with (such as making payments via WeChat).
A powerful hardware does not necessarily result in a good user experience. For example, even if the local model is large, it is useless if the apps you use regularly do not allow it to access their functions. Systems with more moderate specifications but the ability to stably connect to commonly used services may be more practical.
3. The challenge of obtaining permissions is ten times more difficult than competing on computing power
The ability of AI to simulate user interactions does not mean it has the authority to perform those actions. This was a problem encountered in earlier technology previews. When the Nubia M153 (technical version) was released in late 2025, its AI could automatically perform actions on WeChat and Taobao, but users quickly reported that bank apps considered such actions a security risk and forced the AI to exit, and gaming apps believed AI was cheating and restricted its use. The team had to temporarily disable AI functions for financial apps.
The NaviX Ultra has changed its approach by obtaining direct interfaces from app developers (e.g., signing agreements with Ctrip to use its booking service). However, this raises a series of commercial issues:
- Why would apps allow AI to bypass their own interfaces? (If users book flights directly through AI, the apps lose advertising revenue.)
- How should traffic and commissions be divided? (If AI brings users to the apps, should the apps share the profits?)
- Who is responsible for errors? (If AI makes a mistake in booking a flight, who is at fault—the phone manufacturer, the AI company, or the app?)
These issues are far more complex than technical ones, as no one wants to risk losing their business model.
4. AI agents could redefine the way users interact with their phones
In the mobile internet era, the typical process is to open an app, find the desired function, and then use it. AI agents reverse this order: users state their needs, and the AI selects the appropriate app to complete the task. This shifts the role of apps from being the primary point of interaction to providing background services.
This has different implications for all parties:
- For phone manufacturers (like ZTE), it provides a new differentiator beyond hardware specifications (e.g., "My phone can perform tasks for me.")
- For AI companies (like ByteDance's DouBao), it allows them to gain direct access to users' transaction intentions (e.g., booking flights, which is more relevant to sales than simply showing flight prices).
- For app developers, it could reduce customer acquisition costs (AI can target specific users) but also result in a loss of traffic if users no longer open the apps directly.
Who will win in this competition? It will depend on three key factors:
1. How many commonly used apps are supported (e.g., WeChat, Alipay, Ctrip).
2. The success rate of tasks (e.g., whether flights can be booked without user intervention).
3. User retention rate (how many users continue to use the AI after a month, rather than just for chatting).
Only if all these aspects are successfully addressed can AI phones truly become a compelling reason for users to upgrade their phones.
In conclusion
The competition in AI phones is not just about computing power but about permissions and the ecosystem. The ability of AI to perform tasks on behalf of users depends on how many apps are willing to cooperate. This battle for control over the user experience has just begun.