AI agents are emerging as a new intermediary layer between users, operating systems and apps. The more data and control they are given, the harder it becomes for users to understand the permissions involved, the consequences of errors and providers’ commercial interests. As AI features are increasingly preinstalled and embedded in operating systems and apps, avoiding them altogether becomes ever more difficult.
AI-enabled smartphones are expected to account for 45% of devices shipped in 2026, meaning that nearly one in two new handsets will support AI functions. That does not necessarily mean, however, that each will have a full-fledged AI agent capable of operating across apps.
As AI assumes a larger role, the smartphone itself is evolving from a collection of separate apps for chat, email, calendars, text reading, audio and video playback, web browsing and social media into an integrated system in which a central AI combines information and acts independently.
Four major concerns arise from this close integration with the operating system.
Four Risks of Smartphone AI
First, information is already being combined on the device. Under the traditional architecture, each app remained largely confined to its own data: the calendar handled appointments, the camera app handled pictures, the messaging app handled conversations and the email app handled correspondence. A system-wide assistant can draw on all these sources and combine them into a single context. The danger lies not in storing the individual data sets on the smartphone, but in linking them.
The second concern is the agent’s autonomy. Rather than merely presenting an analysis, it takes action based on the available data. A conventional smartphone search returns results. An agent that replaces it can send messages, change appointments, use apps or complete transactions.
A conventional search still leaves the user to decide how to respond to misinterpretations or manipulated instructions. With an agent, the same problems can have immediate consequences before the user has a chance to intervene.
The third concern is permissions. Managing them is already difficult under today’s smartphone architecture. Which app should be allowed to do what? Until now, the main question has been whether a particular app may access a user’s contacts. Today’s permission settings generally amount to a yes-or-no choice.
In the future, users may have to do more than simply grant or deny access. They may also need to specify which tasks an agent may perform, how long its permissions remain valid and which combinations of data it may use.
The fourth and final concern is the role of the operating system provider. The provider is becoming the central intermediary for numerous digital activities. Whoever controls the assistant and its standard interfaces can influence which services it favors and how it directs users to apps, retailers or sources of information.
Because it manages all the data, the AI assistant is not only a privacy risk but also a new kind of commercial gatekeeper. The rise of agents therefore has an economic dimension as well.
A Glimpse of the AI-Driven Smartphone
The following scenario is already technically possible, although it is not yet available in full on every device or in every region. A smartphone can read an incoming email requesting an appointment, check the calendar for a suitable opening, schedule the meeting, draft a reply and send it.
The user may then receive little more than a polite notification that a new appointment has been added to the calendar. Before the meeting, the AI automatically assembles all the required documents and produces a briefing on the key issues likely to arise.
If all else fails, the smartphone can still be used to make a phone call. AI-assisted features can already answer the phone, ask what the call is about and filter suspected spam. These capabilities vary by country and device. And this is only the beginning.
In the future, users will be able to ask an AI to make an outgoing call. It could then discuss the terms of a project with a business partner’s AI, create the relevant folder in the cloud and handle all the preparations, just as a good assistant would, leaving only the final decision to be made. A brave new world?
Users Still Have Some Control
Many services allow users to choose which apps may be linked to the assistant. These permissions are separate from controls covering conversation storage, personalization and the use of data to improve or train AI models. The precise options available vary considerably between Apple and Google and from one app to another.
The settings users need to manage are becoming increasingly complex. The agent acts as an intermediary between the operating system, the apps and the user. On Android devices and iPhones alike, AI is not built directly into the kernel, the core of the operating system, as some claim or fear. Instead, it occupies an adjacent layer.
Much of this sounds like science fiction, yet the technology already exists. The temptation is to use these capabilities without thinking through the consequences. In the future, for example, a user could explicitly authorize an agent to execute stock-market trades independently through an interface provided by a broker.
An experienced investor who knows how to configure the AI to make the same decisions he would can avoid a great deal of tedious work. An inexperienced user, by contrast, risks losing an entire fortune. This is only one of many possibilities.
Opting Out of AI
Anyone unwilling to accept the risks of an AI-controlled smartphone can choose a device without AI instead. Such models are still available, but probably not for much longer. Independent providers offer alternatives as well.
CalyxOS, an independent Android operating system, supports Pixel, Fairphone, Motorola and Shift devices. These are viable alternatives, although some require advanced technical knowledge to set up.
Several other privacy-focused Android operating systems exist. /e/OS describes itself as an open-source Android system that is largely free of Google services. Many experts consider GrapheneOS the most secure option from a privacy standpoint.
No alternative operating system is available for Apple devices, leaving users tied to iOS. On an iPhone, Siri’s AI features can, of course, be turned off, while Gemini can be largely disabled on Android. Users can also prevent the system from learning from individual apps and limit the storage of conversations.
Ultimately, AI will become standard on smartphones. Even so, there are many good reasons to choose a device without an AI agent.