AI assistants on phones are changing in three ways at once. They’re becoming proactive, offering help before you ask by reading context across your apps. They’re becoming agentic, carrying out multi-step tasks inside apps rather than just answering. And more of the work is moving onto the phone, with the cloud handling what the phone can’t. All three make assistants more useful, and all three require deeper access to your personal data, which is the trade-off worth understanding before you switch everything on.
Here’s where things stand in September 2026 and what to watch next.
What do phone assistants do today? #
| Company | What’s shipping | Where it runs |
|---|---|---|
| Pixel 11 (announced August 12, 2026) with “Gemini Intelligence”: agentic tools, suggestions and cues across Android and select apps, including Proactive Assistance, the successor to Magic Cue | Cloud Gemini models plus on-device Gemma models | |
| Apple | Apple Intelligence on the iPhone 15 Pro and later, including the iPhone 17 family, iPhone Air and iPhone 18 Pro | An on-device model, with Private Cloud Compute for larger requests |
| Samsung | Galaxy AI; basic features free, with possible paid “enhanced features” later | Your choice of on-device or cloud processing |
| Amazon | Alexa+, launched February 2025, $19.99 a month or included with Prime | Amazon’s cloud |
Each of these runs a hybrid: a smaller model on the device for quick, personal tasks, and a big model in a data center for the rest.
Five trends shaping the next few years #
1. Assistants that anticipate #
Instead of waiting for a question, proactive features watch what you’re doing and surface information: the booking confirmation when a friend asks for details, the address when you’re about to call a restaurant. Google’s Magic Cue on the Pixel 10 and Proactive Assistance on the Pixel 11 are the clearest examples. We look at what that feature reads in is Magic Cue private.
2. Assistants that act #
“Agentic” is the industry’s word for AI that takes steps on your behalf: filling in a form, moving information between apps, or completing a task across several screens. These features are arriving gradually and usually ask for confirmation before doing anything significant. Expect them to work best inside each company’s own apps first.
3. More AI on the device #
Phone chips keep getting better at running models, and companies prefer local processing where they can: it’s faster, works without a signal, and costs them nothing per request. Google’s system now uses on-device Gemma models alongside cloud Gemini, Apple routes what it can to its on-device model, and Samsung lets you choose on-device-only processing. The limit is memory: the largest, most capable models still live in the cloud.
4. More ways to pay #
Free assistants are expensive to run, and companies are testing ways to cover the cost. OpenAI began testing ads in ChatGPT’s free version for logged-in adult US users in 2026. Samsung says future Galaxy AI features may be paid. Amazon charges for Alexa+ outside Prime. Expect “free” to increasingly mean “free for basic features,” with subscriptions or ads for the rest. We weigh one of those subscriptions in is ChatGPT Plus worth it.
5. Open models you can run yourself #
Alongside the built-in assistants, open-weight models from Google (Gemma), Alibaba (Qwen), Mistral (Ministral) and others keep getting better at small sizes. They’re what make private, offline AI apps possible on ordinary phones, including phones that never get the manufacturer’s AI features.
What this means for your privacy #
The more an assistant anticipates and acts, the more it needs to see: your messages, email, calendar, location and screen. On-device processing helps a lot, because the scanning happens on the phone. It doesn’t make the underlying data local if it already lives in a cloud account, and hybrid systems still send some requests to servers.
Before turning on a proactive or agentic feature, ask:
- What can it read? Which apps and data sources?
- Where does it process them? On the phone, in the cloud, or both?
- Can you limit it to certain apps, or turn it off without losing other features?
- Does it ask before acting?
- What does it keep, and can you delete it?
The switches for each platform are collected in phone AI privacy settings, and how the main assistants compare is in the most private AI assistant.
Where a private, local assistant fits #
Built-in assistants are designed to know a lot about you. If you want the opposite, an assistant that only knows what you tell it, a local chat app is the other end of the spectrum.
Personal LLM runs open models such as Qwen 3.5, Gemma 4, Ministral 3 and GLM 4.6V Flash on your phone. It doesn’t connect to your email, messages or calendar, doesn’t run in the background, and has no account or server. You type your question, attach a photo or a document if you like, and it can read the answer aloud with on-device text-to-speech. After a one-time model download it works with no connection, and it runs on phones as old as Android 7 and iOS 15.1. What it won’t do is act inside your other apps or suggest things unprompted. That’s the point of it.
Many people end up using both: the built-in assistant for timers, directions and quick system tasks, and a local model for anything personal.
What won’t change soon #
- AI will still make mistakes. Bigger models are better, but every assistant can state wrong things confidently. See why does AI make things up.
- Offline models won’t know the news. Anything that runs without a connection only knows what it learned in training.
- The biggest models will stay in the cloud. Phones will keep closing the gap, but hard reasoning and long research will still lean on data centers.
Frequently asked questions #
Will AI assistants replace apps? #
Not soon. Assistants are starting to carry out tasks inside apps, but they rely on those apps to do the work. For now, expect assistants to handle more of the steps between apps rather than replace them.
Are AI assistants moving to on-device processing? #
Partly. Apple, Google and Samsung all run some AI on the phone and send bigger requests to the cloud. On-device processing is growing because it’s faster and more private, but the most capable models still need servers.
Will phone AI features cost money? #
Some already do. Basic built-in features are generally free, but companies are adding paid tiers and ads: Samsung says future Galaxy AI features may be paid, Alexa+ costs money outside Prime, and ChatGPT’s free tier now tests ads in the US.
Can I use an AI assistant that doesn’t read my data? #
Yes. A local chat app that runs a model on your phone only sees what you type or attach, and nothing is uploaded. It can’t anticipate your needs or act in other apps, which is the trade for that privacy.
Which phone assistant is the most private? #
For built-in assistants, those that process more on the device and let you limit what they read score best. The most private option is an assistant that runs entirely on your phone with no account, because there’s no company copy of your conversations at all.