Author | Zheng Shijing

Source | Insight New Research Society

By 2026, AI mobile phones are transitioning from a premium feature on flagship models to a standard offering in the market.

According to Counterpoint Research, smartphones with generative AI capabilities are expected to account for 45% of global shipments, rising to 52% by 2027. However, the same report projects a 13.9% year-on-year decline in global smartphone shipments to 1.08 billion units in 2026, marking a historic low. While the overall market shrinks, demand for AI phones surges.

This contrast suggests that for manufacturers, AI is no longer an optional feature but a survival imperative—a 'do or be pushed out' scenario. The question remains: What exactly are AI phones competing for?

Over the past two years, the competition has focused on in-phone features like voice assistants and AI photo editing. Now, the industry is moving toward a more sensitive direction: enabling AI to operate other apps on behalf of users.

The Nubia NaviX Ultra, launched on September 16, exemplifies this shift. Powered by the Doubao mobile assistant, it is touted as the 'world's first AI agent smartphone.' Users can simply say, 'Buy me the cheapest power bank,' and the phone handles opening apps, comparing prices, and placing the order.

This sounds seamless. However, when AI excludes certain options during price comparisons or selects a specific product, users may wonder: Why were those options excluded? Is the AI's decision aligned with their preferences?

While AI enhances efficiency, users may surrender more than just the hassle of tapping screens.

01 Relinquishing Control, Losing Judgment

AI's transition from a chat tool to a task-handling tool has occurred faster than anticipated.

A NielsenIQ report from August 2026 highlights that in AI-driven e-commerce scenarios, AI agents act as actual decision-makers, evaluating options across multiple retailers simultaneously.

Traditionally, users searched for products ('human-to-goods'): opening apps, entering keywords, browsing results, comparing prices, and adding items to carts. Now, AI handles the entire process ('AI-assisted selection'): users speak a command, and AI matches products, generates recommendations, and compares prices in the background, requiring only a final confirmation.

For instance, the consumer version of the Doubao mobile assistant features 'deep reasoning to understand complex instructions and autonomously plan execution paths' as its core capability. This represents a significant leap: AI no longer passively responds but understands complex intentions, plans tasks, and executes them independently.

However, AI's ability to act on behalf of users does not guarantee user comprehension of its decisions. A command like 'Book me a hotel' yields a confirmation without alternatives, price comparisons, or explanations. Users see only the outcome, not what AI filtered out or why. Critically, AI's evaluation and selection process—which products it excludes, its ranking criteria, and whether commercial partnerships influence recommendations—often remains hidden.

This opacity is not unique to AI. For example, under short-video recommendation systems, platforms possess far more knowledge about user profiles, ranking algorithms, and commercial rules than users themselves. Each click, dwell time, and redirection may feed into training models. Such opacity has long been tolerated because it typically remains at the 'influence' level, not crossing the threshold of user-executed actions.

The difference with AI agents lies in the shift from recommendations ('what you see') to actions ('what is executed'). You cannot audit a completed script any more than you can audit a finished click sequence.

A 2026 Harvard Business School study on AI-assisted loan approvals found that when participants' bonuses depended on loan repayment, they preferred adopting algorithm recommendations without reviewing explanations. Even when 'why' is provided, users may choose not to look.

This implies that decision-making opacity is inherent to the agent model, not a deliberate concealment by manufacturers. When AI operates on your behalf, it often decides for you. If you accept the outcome, you may forgo scrutinizing the process.

By completing actions for you, AI agents quietly strip you of final interpretive authority over decisions.

02 Super Apps Guard the Gate, Terminal AI Knocks

In the app era, platforms controlled what users saw first, but the 'opt-out' right remained with users. If you disliked a recommendation, you could simply swipe it away.

The turning point lies in display limitations.

Take mobile gaming platforms as an example: in the first half of 2025, the top 100 mobile games by revenue increased ad spending by 86.6%, surpassing 18 billion yuan. This figure reflects soaring costs to influence user choices, sharply reducing incremental gains. E-commerce and local services face similar customer acquisition cost inflation. When all platforms vie for prime homepage slots and keyword ads, this strategy nears its limits.

Competition thus shifts downward. Previously, intermediaries presented options, influencing choices through sorting, ads, and recommendations while preserving users' 'opt-out' rights. Now, intermediaries make decisions for users, requiring only confirmation.

For 'decision-making intermediaries' to work, AI must access all options. A user's request naturally spans multiple apps: price comparison on e-commerce platforms, payment via financial apps, and fulfillment through local services. A single app cannot view other apps' inventories or initiate actions elsewhere.

To truly act on users' behalf, AI agents must operate across apps—a privilege only the operating system layer can grant. A shopping app cannot open a payment app; each platform operates within its walled garden. This pushes AI agent competition from the app layer to the system layer.

This shift is evident in the rush by phone makers and internet firms toward terminal-layer solutions.

In July 2026, China's Cyberspace Administration approved the first batch of on-device generative AI services, including Apple Intelligence, Huawei Xiaoyi, Xiaomi HyperAI, and Nubia Doubao's large model. Honor released MagicOS 11, supporting multi-step tasks, while Huawei embedded its Pangu on-device model into HarmonyOS 7's kernel.

While approaches differ, the goal aligns: positioning AI closest to user decisions.

That position is the mobile phone—the terminal through which all traffic flows.

Distribution rights at the terminal layer belong to Huawei, Xiaomi, OPPO, and Vivo, not individual apps. If terminal-level AI dominates, app-only vendors—no matter how dominant in the application layer—may reduce to content or service suppliers, becoming backend interfaces called by AI rather than direct user destinations.

ByteDance reacted first. The launch of Doubao phones resembles an 'attack-as-defense' strategy in industry competition. Leveraging Nubia's hardware, it pushed the Doubao assistant to consumers.

Doubao phones are not new. In December 2025, the first-generation Nubia M153 sold out 30,000 units at 3,499 yuan. Using a GUI Agent scheme, AI simulated clicks to operate apps. However, within a week, WeChat flagged 'abnormal login environments,' while Taobao, Alipay, and bank apps triggered risk controls, collectively blocking Doubao.

This likely reflects commercial rivalry between 'AI agents taking over user access' and 'super apps defending traffic sovereignty.' WeChat's social graph, Taobao's transaction data, and Meituan's local services are each company's moat. Allowing external AI to bypass these walls reduces platforms from direct user gateways to backend suppliers invoked by AI.

The second-generation product shifted tactics, abandoning simulated clicks for MCP protocol-driven operations, transforming 'forced entry' into 'knocking for permission.' AI can only act if apps consent. Thus, whether AI opens an app or compares prices depends not just on users or phone makers but also on the app's willingness.

Overall, terminal-level AI seeks cross-app scheduling on users' behalf, while super apps defend their entry points with permissions and risk controls. The shift has begun but remains incomplete. Currently, it resembles a battle for scheduling authority between system and application layers.

Securing scheduling rights positions AI to make decisions for users. However, the standards behind those decisions, potential commercial influences, and user recourse remain unresolved.

03 Efficiency Can Be Built, Trust Must Be Earned

AI's decision-making model offers clear efficiency gains. Yet efficiency alone cannot sustain a business.

Domestically, competition for entry points centers on system permissions. Abroad, AI agents have entered transactional phases, exposing challenges that may or may not recur in China but highlight critical hurdles.

When AI agents mediate between merchants and users, their operations require funding. In January 2026, OpenAI chose to charge a 4% commission on Shopify sales processed via ChatGPT's checkout feature, Overlay on Shopify's standard transaction fees above . Merchants face an additional AI channel fee beyond existing platform costs.

While 4% seems modest, its reasonableness depends on whether it drives incremental orders. If users redirect to original platforms after AI recommendations, the AI channel adds no value—merely another toll.

In reality, this occurs frequently. OpenAI launched in-app shopping in September 2025, emphasizing seamless, no-redirect payments. Within six months, the model collapsed. Reports attribute this to ChatGPT's real-time checkout with Walmart achieving just one-third the conversion rate of redirected transactions. Users let AI recommend products and compare prices but ultimately abandoned payments.

Merchant costs and user redirection are two sides of the same coin. The more accurate AI's recommendations, the stronger users' incentive to compare prices elsewhere. AI can help users choose but cannot prevent them from paying through familiar channels.

The 4% fee does not drive incremental orders; it inserts a Charging process into existing transaction chains. Without corresponding value creation for merchants, the structure collapses.

More daunting than user redirection is user distrust in AI recommendations. This stems not only from fear of commercial bias but also from the possibility of 'AI feeding.'

The 2026 CCTV 3·15 Gala exposed a GEO 'gray industry.' Insiders used GEO optimization software to invent a 'Apollo-9' smart bracelet, fabricating features like 'quantum entanglement sensing technology' and publishing dozens of promotional articles online. Within two hours, when reporters asked an AI model about the bracelet, it detailed the fabricated product as 'standard.'

(Source: CCTV Finance)

GEO, an optimization technique, becomes systemic 'AI feeding' when abused. Users struggle to verify whether AI recommendations are commercially influenced. The more authoritative recommendations appear, the more users lower their guard—concealing traces of manipulation behind that 'authority.'

Regulators have acted. In April 2026, China's Cyberspace Administration launched 'Operation Clean Network: AI Application Disorder,' targeting AI data poisoning and unregistered services, disposing of over 14,000 non-compliant AI products in its first phase. The 'Administrative Measures for Identifying AI-Generated Content' also mandate explicit labeling of synthetic text and images.

However, these rules address 'whether content is AI-generated,' not 'whether AI-recommended products are commercially influenced.' Labels resolve source issues but not motivational ones. Knowing content is AI-generated does not reveal if recommendations are paid placements or hidden ads. Regulation has advanced but remains distant from the accountability AI agents require.

When e-commerce platforms emerged, their upside was market expansion and logistics growth; their downside was offline disruption and regulatory lag. AI agents face similar tensions: they reduce user operational costs but impose additional merchant expenses and platform bypass risks. Their viability may hinge not on their upside but on whether their downsides can be contained.

Ultimately, the critical question for AI phones is not which manufacturer achieves first-mover status but whether the 'decision-making-on-behalf-of-users' model can sustain itself. Efficiency can be engineered; trust must be earned through consistent, verifiable results.

At present, all parties are still vying for entry points and securing positions. However, an entry point is merely a starting point, not the destination. In the future, whoever can first make users feel confident enough to take the next step will be the one who can truly secure this entry point.