The new joint white paper envisions a future where AI agents follow users across every device — and where the smartphone finally steps aside as the sole center of our digital lives.
If a phone maker today advertised its latest handset as a “5G phone,” consumers would probably be confused. Isn’t 5G just standard now?
The same thing is starting to happen with “AI phones” and “AI PCs.” Features like AI photo editing, AI summarization, and AI search have become so common that the label itself is losing meaning. Every device can be AI-powered. Every app can be AI-enhanced. But for users, all this AI saturation has created more confusion than clarity.
A new white paper from IDC and Qualcomm, titled From AI Devices to Personal AI: Agent-Driven Terminal Evolution and Industry Restructuring, argues that we are in a transitional phase. And it offers a clear roadmap for where personal AI is heading next.
AI agents have gone mainstream — but users aren’t satisfied yet
According to IDC’s survey data, 73% of Chinese consumers are now aware of AI agents in 2026, up from 51% in 2025. More notably, usage has jumped from 33% to 59% over the same period. Nearly six in ten consumers have actually used an AI assistant or agent.
That means AI has crossed the early-adopter phase and entered the mainstream. People are using it to write emails, summarize meetings, search for information, edit photos, and even plan travel.
But rapid adoption doesn’t mean satisfaction.
The white paper highlights a glaring gap: 65% of consumers say AI cannot accurately understand their instructions — the number one pain point. Another 59% have experienced context loss during conversations, 54% say AI outputs don’t meet expectations, and 48% find generated code or documents unusable without significant revision.
The more people use AI, the more they bump into its limits. Ask it to revise an article, and it might forget the original tone. Ask it to continue a previous task, and it may demand the background all over again. Ask it to organize a complex project, and what you get is often a logically correct but practically useless template.
“The next phase of competition,” the white paper argues, “has shifted from ‘having AI’ to ‘can it be used long-term'”.
What users really want: an AI that remembers them
The survey asked consumers what they most want from AI improvements. The top answer — at 57% — was personalized interaction. Second was memory — the ability for AI to actually remember who you are and what you’ve told it before.
For years, phone makers have treated AI as a feature checklist: write articles, generate images, summarize recordings, translate calls, remove strangers from photos. But a longer list doesn’t mean more value.
“If every time users have to re-explain their profession, writing style, family members, budget range, and preferred formats,” the white paper notes, “then no matter how many features there are, they’re just isolated tools in a toolbox”.
The key is long-term memory. A personal AI should know what you do for work, how you like to express yourself, what projects you’re working on, what’s on your calendar next week — and which actions require your explicit approval.
The paper proposes a three-layer memory architecture: short-term memory for ongoing tasks, long-term memory for accumulated preferences and personal knowledge, enabled by techniques like context caching, personal knowledge graphs, and on-device vector retrieval.
Today, we write prompts for every single task. Tomorrow, we might only need to add what’s new — because everything else is already in the AI’s memory of us.
From answering questions to actually getting things done
The white paper draws a clear line between two eras. The past few years have been the “responsive phase” — AI that answers questions, summarizes, and generates content, but still waits for instructions and leaves the real work to the user.
The next phase is the “action phase.” AI agents need to understand user goals, break down tasks, plan steps, call tools, and adjust based on results.
Take travel planning. Traditional AI generates a three-day Shanghai itinerary — but you still have to check your calendar, compare flights, book hotels, and share the plan with colleagues. An agentic AI, by contrast, would read your calendar, avoid existing meetings, filter options based on company travel policy, generate a draft, wait for your confirmation, complete the bookings, and sync everything to your calendar.
The user provides intent, not step-by-step instructions.
The paper describes this as a fundamental shift in human-computer interaction: from “operating devices” to “expressing intent”. Agents could become the next major interface paradigm after graphical user interfaces and multi-touch.
The hard part: actually finishing complex tasks
Here’s the catch. The white paper openly acknowledges that most current AI agents still can’t reliably complete complex, multi-step tasks across different apps.
They can handle single-step, single-app operations. But throw in authentication, missing information, changing page layouts, or unexpected errors, and the user has to take over again.
This is why many agent demos look amazing but feel like a letdown in real use. A wrong answer is one thing — you can correct it. But a wrong action? That can have real consequences. An agent that mistakenly sends an email, deletes a file, makes a purchase, or changes a calendar entry could cause real damage.
The paper argues that agents need to know which tasks can run autonomously and which need user confirmation. They need to log their actions so users can review, pause, or undo. And they need to ask questions when unsure — rather than guessing blindly.
“It’s very similar to assisted driving,” the paper says. “You need to eliminate as many bad cases and corner cases as possible”.
The industry roadmap laid out in the white paper sees 2026–2027 as the period for achieving single-scenario, multi-step task decomposition, long-context memory, and more stable execution. Truly crossing device boundaries and proactively anticipating user needs? That’s further out.
AI that follows you, not your devices
IDC projects that the average person will own 7.6 smart devices by 2026, up from 4.5 in 2023. Phones, PCs, tablets, smartwatches, earbuds, cars, home devices — they form a daily network of connected gadgets.
But right now, the AI on each of these devices largely operates in isolation. Start a conversation with AI on your phone, and your PC has no idea what you were talking about. Your smartwatch knows you just finished a workout, but your phone’s AI doesn’t. Your car knows you’re heading home, but your smart home still waits for you to manually turn on the AC.
The white paper identifies two emerging user needs. One is consistency — users want the same AI persona, memory, and capabilities across all devices, without having to re-adapt on every screen. The other is continuity — task progress, context, and service state should carry over seamlessly, without repeated instructions or manual data sync.
“In the future,” the paper argues, “the AI的主体 may no longer belong to any single device, but to the user themselves”. If you break your phone or upgrade your PC, your AI shouldn’t have to get to know you all over again.
The smartphone stays central — but smart glasses are coming
Since 2023, plenty of startups have tried to dethrone the smartphone as the center of personal computing. They’ve mostly failed.
The white paper says smartphones will remain the most important hub for the foreseeable future — the entry point for compute, data, storage, and the AI ecosystem. They’ll handle daily AI tasks and orchestrate resources across on-device, edge, and cloud.
But the smartphone won’t do everything alone. PCs are better for creative work, development, and heavy compute. Smart glasses and earbuds become natural interaction portals. Wearables like watches and rings continuously sense your body and environment. Cars and smart home devices extend the AI experience into mobility and home life.
This is a distributed design for AI. Different devices contribute based on their hardware strengths — sensing, interaction, computation, execution. What the user experiences is a unified intelligent agent, powered by multiple devices working together behind the scenes.
And among all these secondary devices, smart glasses stand out.
IDC projects global smart glasses shipments will reach 19.89 million units in 2026, up 34.6% year-over-year. From 2026 to 2030, the compound annual growth rate is expected to be around 25.5%, pushing annual shipments toward 50 million units by 2030.
In the personal AI ecosystem, the paper argues, smart glasses matter less for their display capabilities and more for what they can see. They offer a first-person view, continuously capturing environmental information from the user’s perspective. Combined with voice, spatial awareness, and multimodal models, AI can directly understand what the user is looking at — without requiring the user to type or describe it.
See an unfamiliar device? Just ask the AI how to operate it. Shopping for groceries? The AI can suggest options based on your dietary preferences. Fixing something? The AI can provide step-by-step visual guidance.
Smart glasses also solve a fundamental friction point of the smartphone as an AI入口: using the phone requires you to actively pull it out, open an app, and translate the real world into text, voice, or photos. Glasses could become an always-on感知 portal.
Of course, this raises sensitive privacy questions. When a device can continuously see everything you see, what gets recorded — and who can access that data — becomes a far more urgent concern than with any consumer device before.
More AI will run locally — and not just for privacy
The white paper repeatedly emphasizes a “local-first, device-cloud collaboration” approach.
Future AI tasks will be distributed based on latency, privacy sensitivity, and compute complexity: frequent, lightweight, and sensitive tasks stay on-device; medium-complexity tasks go to edge computing; truly complex jobs call the cloud.
For ordinary users, this brings several real benefits. Speed — many daily operations don’t need to wait for data to upload and return from the cloud. Availability — basic AI functions still work when networks are unstable or unavailable. Privacy — a personal AI that truly understands you needs access to photos, contacts, calendar, health data, location, and behavioral history. Asking users to upload all of that to the cloud makes trust difficult to build.
And then there’s cost — often overlooked. Agents need to run continuously, frequently processing perception data, memory, and task states. If everything relied on cloud models, token and inference costs would skyrocket as user numbers grow. Having devices handle more lightweight tasks reduces cloud pressure and makes long-running personal AI commercially sustainable.
By 2030, “AI device” might sound as dated as “5G phone”
IDC forecasts that by 2030, GenAI phone shipments will reach about 771 million units — 63.8% of the phone market. GenAI PCs will hit 134 million units, a 47.1% share. GenAI tablets will approach 100 million units, or 68.9%. If basic AI PCs are included, overall AI PC penetration could exceed 82.9% by 2030 — near total普及.
Which means “AI phone” and “AI PC” are probably just transitional marketing concepts. Just as no one today emphasizes that a phone can go online, take photos, or run apps, in a few years, on-device models, NPUs, personal memory, and agent orchestration will simply be baseline capabilities.
Users won’t buy a device because it “supports AI.” They’ll buy it because the AI understands them well enough, works consistently across devices, completes tasks reliably — and because they trust the company with their personal data.
The differences between devices will shift from hardware specs to system capabilities. Screens, cameras, processors, and battery life will still matter — but as components of the AI experience. Lower-power sensors enable continuous perception. More memory holds personal models and memories. More efficient NPUs support always-on background agents. More reliable connections keep tasks flowing across devices.
The digital world has long been organized around devices and apps. Different accounts, different data, different ways of doing things on phones, PCs, and cars. Even the same service, once you cross devices, often feels like starting over.
Personal AI wants to build a different logic: the user as the single center, with devices as capability endpoints distributed across different scenarios. Phones orchestrate. PCs produce. Glasses perceive. Watches monitor continuously. Cars and homes execute. AI services flow between these devices, and users no longer need to care where exactly the computation happens or which app completed the task.
Whether this vision becomes reality depends on whether the industry can solve the hard problems — memory, context, stability, privacy, and trust — before users lose patience with the AI that’s supposed to make their lives easier.