On-device AI, such as Apple Intelligence on recent iPhones and Macs or Gemini Nano on supported Android phones, runs language models directly on the user’s hardware instead of a remote server. That lowers latency and keeps much data local, but it also creates risks that marketing teams rarely plan for: inconsistent output across devices, summaries that distort brand messages, and staff using personal-device features with client data. Governance should start before adoption, not after.
What does on-device AI actually do?
Apple Intelligence bundles a set of features into iOS, iPadOS and macOS: writing tools that rewrite or proofread text, notification and email summaries, image generation through Image Playground, and a more capable Siri. Smaller tasks run on a local model; heavier requests can go to Apple’s Private Cloud Compute, which Apple describes as servers built on Apple silicon that do not retain user data. Some requests can also be handed to ChatGPT, but only after the user agrees.
Gemini Nano is Google’s smallest Gemini model. It ships inside Android through a system service called AICore and powers features such as smart replies in Gboard, recorder summaries and scam call detection on supported Pixel and Samsung devices. Developers can reach it through Google’s on-device APIs, and Chrome has been testing built-in AI APIs that run a Nano-class model in the browser.
The common thread is simple. A small model, often a few billion parameters or less, does useful work without a network round trip. That design is the source of both the benefits and the limits.
Why the privacy story is only half the picture
Vendors present on-device AI mainly as a privacy feature, and for many tasks that is fair. A sentence rewritten locally never leaves the phone. But “the data stays on the device” does not mean “the data is governed.” The device may belong to an employee, the output may be pasted into a client deliverable, and nobody may have logged that a model touched it.
There is also the hand-off problem. When a request is too big for the local model, it may route to a cloud service. On Apple devices the system tells the user when ChatGPT is involved, but staff click through prompts quickly. On Android, apps that use Gemini Nano may also call cloud Gemini models for harder requests. Your policy has to cover the whole chain, not just the local part.
Which limits matter most for marketing teams?
Small models make different trade-offs than large cloud models. For a marketing department, five limits come up again and again:
- Shorter context: on-device models handle short inputs well and long documents poorly, so a 40-page brief summarized on a phone may drop key sections.
- Weaker factual recall: small models know less about the world and are more likely to state something wrong with confidence.
- Device fragmentation: features depend on chip, RAM, region and language settings, so two colleagues may get different results from the same command.
- Limited language coverage: support for languages such as Turkish has arrived later than English, and quality varies by language.
- No audit trail: local features usually leave no central log of what was generated, edited or summarized.
The last point is the one compliance teams notice first. If a summary of a client email loses a condition (“only if approved by legal”), there is no record showing that the AI produced the summary rather than the employee.
The risk that hits brands from the outside
Most governance documents focus on what staff do with AI. On-device AI adds an external risk: your customers now read your messages through a model. Notification summaries on iPhones condense emails, push notifications and messages into one or two lines. Early versions produced summaries of news alerts that were plainly wrong, and Apple paused summaries for news apps for a time to fix the problem.
For a brand, this means a carefully written promotional push can arrive as a flattened, sometimes misleading summary. Clear first lines help. So does putting the offer, the deadline and any condition in the opening sentence rather than at the end. Treat the first 80 to 100 characters of every email and push notification as the part a model is most likely to keep.
What governance rules should you set first?
You do not need a 30-page AI policy to start. You need a handful of rules that people can remember and that fit how on-device features are actually used. A practical first version looks like this:
- Define which devices may be used for client work, and whether personal phones with AI features enabled are allowed to open client files.
- List the tasks that are fine on-device (proofreading internal notes, summarizing your own meeting recordings) and the tasks that are not (summarizing contracts, drafting legal or medical claims).
- Require human review for any AI-assisted text that goes to a client or into a published asset, with the reviewer named in the project tool.
- Decide whether cloud hand-offs such as the ChatGPT integration are allowed, and if so, which account type is used.
- Set a rule for images: anything made with Image Playground or similar tools must be labelled internally and must not depict real people without consent.
- Name an owner who reviews the policy when vendors ship major operating system updates, since features change with each release.
Keep it on one page. A rule that staff cannot recall during a busy afternoon will not be followed.
Device management settings
If your company uses mobile device management, check what your platform exposes. Apple has added management controls that let administrators restrict certain Apple Intelligence features on supervised devices, and Android Enterprise offers its own policy settings. Availability changes by release, so verify in your MDM console rather than relying on a blog post, including this one.
How Alien Road applies on-device AI governance
In a typical engagement we start with discovery: a short survey of which devices the client team uses, which AI features are switched on and where client or customer data passes through them. Most teams are surprised by how many features are already active by default after an operating system update.
Next comes setup. We draft the one-page rule set above, adjusted for the client’s sector, and map it to the MDM or browser settings they already have. For a retail brand, for example, we also rewrite push and email templates so that the key message survives automatic summarization.
Measurement follows. We test how major templates appear as notification summaries on current iOS and Android devices and record the results in a shared sheet. Reporting is quarterly: what changed in the vendors’ features, which rules were updated, and which templates needed another pass.
Where to start this week
Pick one hour and do three things. First, ask each team member which AI features they have used in the last month on their phone or laptop; write the list down. Second, send your five most important marketing emails and push notifications to a test iPhone with summaries enabled and screenshot what appears. Third, write the first draft of your one-page rule set using the six rules above.
That is enough to move from “we have not thought about it” to a documented position. You can refine the details after the next major OS release, when the feature list will have changed again anyway.
Keep reading: AI Optimization · AI Optimization