On-device AI runs language and vision models directly on a phone or laptop instead of sending every request to a remote server. Apple Intelligence on recent iPhones, iPads and Macs, and Gemini Nano on supported Android devices and in Chrome, are the best-known examples. They suit short, private, everyday tasks; cloud models remain stronger for long documents, deep research and heavy creative work. Choosing well means matching the task to where it should run.
What do Apple Intelligence and Gemini Nano actually do?
Apple Intelligence is Apple’s set of AI features built into its operating systems. It includes writing tools that rewrite, proofread and summarise text across apps, notification and email summaries, image tools, and a more capable Siri. Apple runs smaller models on the device and sends larger requests to its Private Cloud Compute servers, which it says are designed so data is not stored. Some requests can also be passed to ChatGPT with user permission.
Gemini Nano is Google’s smallest Gemini model, designed to run on device. On Pixel and some Samsung phones it powers features such as recorder summaries, smart replies and scam call detection. Google has also brought Nano into Chrome on desktop, where developers can call built-in AI APIs for tasks like summarising or translating text without a server round trip.
Feature lists differ by device, region and language, and they change with each software update. Check current support for your devices at the time of writing.
How does on-device AI compare with cloud tools?
- Privacy: on-device processing keeps the data on the hardware; cloud tools send it to a provider under their data terms.
- Speed: local models respond with no network delay for small tasks; cloud models can be slower to start but handle more work per request.
- Capability: small local models handle summaries, rewrites and classification; large cloud models handle long context, complex reasoning and detailed generation.
- Cost: on-device features come with the hardware; cloud tools usually charge per seat or per token.
- Offline use: local models can work without a connection; cloud tools cannot.
- Control: cloud APIs give you model choice, logs and settings; built-in device features offer few controls.
When should a marketing team rely on on-device AI?
On-device tools are a good fit for personal productivity with sensitive material. A sales manager summarising a long client email on an iPhone, or an account lead proofreading a contract summary on a Mac, keeps that text off third-party servers. For agencies handling client data, that matters.
They also fit quick, repeated micro-tasks: rewriting a social caption in a friendlier tone, cleaning up meeting notes, or translating a short message. These jobs do not need a large model, and the result is fast enough to stay in the flow of work.
When are cloud tools still the better choice?
Cloud models win when the job is big. Analysing a full quarter of campaign data, drafting a long article, comparing ten competitor pages, or generating many ad variants needs more memory and reasoning than a phone can provide. Shared workflows also favour the cloud: a team needs prompts, history and outputs in one place, not scattered across personal devices.
Another reason is consistency. A cloud API returns similar output for the same prompt and settings, which makes it easier to review and audit. On-device features vary with each user’s device, operating system version and language settings.
What does on-device AI mean for marketers’ audiences?
The bigger story may be on the receiving end. When phones summarise notifications and emails, your carefully written subject line may reach the user as a short machine summary. Apple’s mail summaries and priority features already reshape how messages appear in some inboxes.
- Put the main point in the first sentence of emails and push notifications.
- Avoid clever but vague subject lines that a summary may flatten.
- Use clear sender names and consistent formatting so priority sorting reads your message correctly.
- Test campaigns on devices with these features turned on, and look at what the summary says.
Browser-based models raise similar questions for websites. If Chrome can summarise a page locally, clear headings and a direct opening paragraph help the summary say what you intended.
What are the limits and risks?
Small models make mistakes, and notification summaries have produced well-publicised errors, which led Apple to adjust and label some summary features. Treat any on-device summary as a convenience, not a record. For client work, keep the original text as the source of truth.
Device management is the other issue. Companies need to decide whether AI features are allowed on managed devices and which external extensions, such as ChatGPT passthrough, are permitted. Apple and Google provide controls for administrators through mobile device management.
How Alien Road applies on-device AI
In a typical engagement Alien Road reviews how a client’s team works across devices and data types. We mark which tasks involve sensitive client or customer data and could stay on device, and which need shared cloud workflows with logs. The result is a short policy, not a long document: which tool for which job.
On the marketing side, we test email and push campaigns on devices with summaries enabled and record what the summary shows. If the summary misses the offer, we rewrite the opening lines and test again. For web pages, we check whether the opening paragraph and headings give a browser summary the right message. Reporting compares open and click rates before and after these rewrites.
Where to start this week
List five AI tasks your team does most often and note what data each one touches. Move the short, sensitive ones to on-device tools where your hardware supports them, and keep the large, shared ones in the cloud. Then send your next email campaign to a test iPhone with summaries turned on, read what it says, and adjust your first sentence if needed. Keep a screenshot of each summary next to the campaign results, because over a few months that file will show which opening styles survive machine summaries best.
Keep reading: AI Optimization · AI Optimization