AI features are cheap to add and expensive to run. An app that keeps a model resident, or polls a server every few seconds, will show up in your battery stats within a day. The good ones invoke AI on request and then let the device sleep.
What separates a good implementation
- On demand only: the model loads when you press the button, not when the app opens.
- Local where possible: transcription, translation and summarising are small enough to run on the NPU.
- Visible queue: long jobs run as a notification, not as a hidden background service.
- Honest offline behaviour: features that need a server say so before you tap.
Ten categories worth installing
Rather than a list of brands that ages badly, these are the categories where AI work is actually useful on a phone:
- Offline dictation that writes into any text field.
- Call recording with on-device transcription.
- Photo culling that scores sharpness and duplicate frames locally.
- Document scanning with deskew and text extraction.
- Subtitle translation for video played on the device.
- Notes with local semantic search, no account required.
- PDF summarisation that keeps the file on the phone.
- Habit trackers that detect patterns without uploading journals.
- Language practice with on-device speech scoring.
- Accessibility readers that describe the screen in real time.
How to audit any app in two minutes
Install it, use the AI feature once, then close it and check Settings, Battery, since last full charge. A well-behaved app shows seconds of background use. If it shows minutes per hour with nothing running, the model is resident and your battery is paying for it.
The trade-off to accept
On-device work is private and instant but limited in quality. Cloud work is better but sends your content away. The best apps let you choose per task, and remember the choice.
Bottom line: judge an AI app by what it does when you are not looking. Battery stats do not flatter, and they are the fastest way to separate a useful tool from a demo.
