Phones are personal, small and impatient. That makes them the hardest place to add AI well — and the most rewarding when it works. The goal isn't to put a chatbot in the corner of your app. It's to make the moments people already have a little faster, a little easier, a little less annoying.
This is a practical guide to doing that: where to look for opportunities, where the model should run, and the design details that decide whether people trust the feature or turn it off.
Start from a moment, not a model
The weakest AI features start with “we should use AI” and go looking for a place to put it. The strongest start with friction you can already see in your analytics or support inbox. Look for the moments where users:
- Type a lot on a small keyboard — long descriptions, addresses, repetitive notes.
- Search, scroll and give up because they don't know the exact words.
- Copy information from one place to another — a receipt, a label, a screenshot.
- Read something long to find the one line that matters to them.
A field pre-filled from a photo, a long thread summarised in three lines, a search that understands what someone meant rather than what they typed — those are AI features people notice in a good way, often without thinking of them as AI at all.
On the device, or on the server?
Modern phones can run useful models locally. Recognising text in a photo, detecting objects, simple classification and some language tasks run on the device itself — fast, private and available offline. Larger language tasks usually run on a server you control, calling a hosted model.
- On-device — instant, private, works offline, no per-request cost; limited model size and older phones may struggle.
- Server-side — far more capable models, easy to update without an app release; needs a connection, adds latency and an ongoing cost per request.
- Hybrid — a quick on-device step (read the text from the photo) followed by a server call only when needed (interpret it). Often the best of both.
One firm rule: never put a model provider's API key inside the app. Anything shipped to a phone can be extracted. Route requests through your own backend, where you can authenticate users, apply rate limits and keep costs under control.
Latency is the user experience
A two-second wait in a chat window is tolerable; the same wait after tapping a button feels broken. The design has to absorb the time the model needs.
- Stream text as it's generated, so people see progress immediately.
- Show skeletons that match the shape of the result, not a generic spinner.
- Never block the rest of the app — people should be able to keep working while the model thinks.
- If a task genuinely takes a while, make it background work and notify people when it's ready.
- Cache results that don't change, so the second visit is instant.
On a phone, a feature that is slow is a feature that is off.
Privacy and permissions
Be explicit about what leaves the device and why. Ask for camera, microphone or photo permissions at the moment they're needed, explain the benefit in one sentence, and give people a way to use the app without the AI feature. Check what your model provider does with the data you send, and say so in your privacy policy in plain words. Trust lost on a phone is very hard to win back.
Fail gracefully
Models are sometimes wrong, networks drop and providers have outages. A good AI feature is designed so that none of those moments leaves the user stuck.
- Always show the source or the original next to an AI-generated summary.
- Make every suggestion editable before it's saved.
- When the network or the model fails, fall back to the normal, manual flow — silently if possible.
- Avoid confident wording for uncertain results; “Looks like…” beats “This is…”.
Measure whether it helps
Before launch, collect a small set of real examples with known good answers and test against them whenever the prompt or model changes. After launch, watch behaviour, not just usage: how often people accept a suggestion unchanged, edit it, or undo it. A feature people constantly correct is costing them time, however impressive it looks in a demo.