On-device AI runs models directly on a phone, tablet or computer instead of in the cloud. Apple Intelligence on recent iPhones, iPads and Macs, and Google’s Gemini Nano on many Android phones and in Chrome, are the best-known examples. For marketers and app teams, the common mistakes are ignoring how these features reshape messages and content, assuming every user has them, and overpromising privacy or capability.
What is on-device AI, and why does it matter to brands?
On-device models are small enough to run on a phone’s processor, so they respond quickly, work offline for some tasks and keep data on the device. Apple Intelligence uses on-device models for many features and sends larger requests to Apple’s Private Cloud Compute servers. Gemini Nano powers features such as smart replies, summaries and scam detection on supported Android devices, and Google has exposed it to developers through Android APIs and to web developers through experimental built-in AI APIs in Chrome.
For brands, the impact is indirect but real. These systems summarise your emails and notifications, rewrite text users type, and help people find content in your app. That changes how your messages are seen before anyone opens them.
Mistake 1: ignoring notification and email summaries
Apple Intelligence can summarise stacks of notifications and emails, and mail apps increasingly show AI summaries or priority labels. A promotional email whose key offer is buried in the third paragraph may be reduced to a vague summary. Apple paused news notification summaries for a period in 2025 after inaccurate summaries of headlines drew complaints, which shows how much the summary layer can distort a message.
Fixes:
- Put the main point in the subject line and first sentence of emails.
- Write push notifications as complete, plain statements, not teasers.
- Avoid stacking several unrelated messages in short succession.
- Test how your messages appear on devices with summaries switched on.
Mistake 2: assuming every user has these features
Support depends on device model, operating system version, language and region. Apple Intelligence requires recent hardware and has rolled out language by language. Gemini Nano is limited to certain Android devices. Building a campaign or app feature that depends on on-device AI without a fallback leaves a large part of your audience with a broken or empty experience.
Check your own analytics for device and OS distribution before planning features, and always design a path that works without the AI layer.
Mistake 3: overpromising privacy
“Runs on your device” is a strong privacy message, but it is not always the full picture. Some requests go to the cloud, whether to Apple’s Private Cloud Compute or, with user permission, to third-party models such as ChatGPT through Apple’s integration. If your app adds AI features, describe exactly what happens to data in each case. Vague claims can mislead users and attract regulator attention, particularly under GDPR in Europe.
Mistake 4: expecting frontier-model quality on a phone
Small on-device models are fast and private, but they know less and handle complex tasks less well than large cloud models. App teams sometimes promise long-form writing, deep analysis or detailed product advice from an on-device model and disappoint users. Match the task to the model:
- On-device suits short summaries, classification, smart replies, simple rewriting and offline features.
- Cloud models suit long content, multi-step reasoning and anything needing broad knowledge.
- A hybrid approach sends simple tasks to the device and harder ones to the cloud, with user consent where data leaves the device.
Mistake 5: not making app content readable to system AI
Apple’s App Intents framework lets apps expose actions and content to Siri, Spotlight and Apple Intelligence features. Android offers comparable integration through its own app actions and AI APIs. Apps that do not adopt these frameworks may become less visible as people rely more on assistants to complete tasks across apps. Exposing key actions, such as “reorder my last purchase” or “show my booking”, keeps your app part of the flow.
Quick audit questions for app teams
- Which three actions do users repeat most, and are they exposed to the system assistant?
- Do notifications read clearly when summarised?
- What happens on devices without the AI features?
- Is the data flow for each AI feature written down in plain language?
How Alien Road applies on-device AI in client work
In a typical engagement, discovery reviews the client’s email, push and in-app messaging alongside device and OS data from analytics. We check how messages appear on devices with summaries enabled and where key information gets lost.
Setup is practical: rewriting subject lines, preheaders and notifications so the core message survives summarisation, and working with the client’s app developers on which actions to expose through system frameworks. Privacy wording for any AI features is reviewed so claims match the actual data flow.
Measurement compares open, tap and conversion rates on rewritten messages with earlier versions, segmented by device type where possible. Reporting to the client summarises changes made, results by segment and open questions for the development team.
Why does on-device AI matter for marketing reach?
Every summary, priority label and smart reply sits between your message and the person reading it. If the device decides an email is low priority or reduces a notification to a bland line, your reach drops without any change in your sending numbers. Open rates may slip and nobody can say why.
Clear, specific writing protects you here. Messages that state the benefit, the deadline and the action in plain language survive summarisation better than clever wordplay. That was good practice before on-device AI; now it has a measurable cost when ignored.
How should developers test on-device features?
Testing needs real devices across supported and unsupported models, not only simulators. Check the feature with AI enabled, with it disabled in settings, and on hardware that does not support it at all. Record response times, since small models on older chips can be slower than expected, and log fallback rates so the team sees how often the cloud or non-AI path is used.
Where to start this week
Take your last five marketing emails and five push notifications. Read only the subject line and first sentence of each, and ask whether a summary built from them would still carry the offer.
Rewrite the ones that fail. Then check your analytics for the share of users on devices that support Apple Intelligence or Gemini Nano, so your team knows how much of the audience these features already reach.
Keep reading: AI Optimization · AI Optimization