On-device AI versus a cloud coach

A coach that reads your habits is reading a diary. If that diary is packed into a prompt and sent to a model in someone else's data centre, you have hired a very smart stranger. If the same summary runs on the phone, using Apple Intelligence, you have hired silicon you already carry. Those are different products. Store screenshots cannot tell them apart. Airplane mode can.

This is not a holy war against every cloud. Weather, maps, and multiplayer games need a network. A nightly paragraph about whether you skipped stretching does not. Apple's own guidance on Foundation Models is the boring, correct default for health, journaling, and personal patterns: keep them on-device when you can. A habit coach is personal patterns.

Privacy in habit apps is the wider checklist. This post is only the coach.

What "coach" usually means in copy

Three jobs get branded as AI. One: turn a sentence into a habit ("read 20 pages on weekdays"). Two: summarise a week ("you skip Thursday, you hit the walk"). Three: generate encouragement. All three can run locally on a capable iPhone with Apple Intelligence. All three can also run as an HTTPS call. The encouragement is the least valuable and the most tempting to outsource, because a cloud model is fluent. Fluency is not the same as a private architecture.

Wendy Wood's habits do not require a language model. They require a stable cue. Phillippa Lally's automaticity curves do not require a pep talk. They require repetition. Gollwitzer's if-then plans are sentences you can write yourself. A model can help you draft the sentence. It cannot attach the sentence to your kettle. Keep that humility or the coach becomes a second motivational quote, which environment design already beat.

Charles Duhigg popularised inspecting cue, routine, reward. A local summary that says "the reward is the streak, and you are protecting the streak instead of the walk" can be useful. A cloud chat that role-plays a trainer is theatre. Theatre leaks.

The airplane-mode test, done honestly

Turn on airplane mode. Disable Wi-Fi. Open the coach. If it still produces a summary from the habits already on the phone, the inference is local (or the text was cached, which you can check by logging something new and asking again). If it spins, errors, or asks you to connect, it is not on-device, whatever the marketing heading said.

No account should be required for that test. A coach that needs you to be a logged-in user is already a network identity plus a model. Do you need an account to track habits covers the identity half. Here: guest mode plus local inference is the strict pair.

Device class matters. On-device Foundation Models need a phone that actually runs Apple Intelligence, on a recent enough iOS. On older iPhones the honest product still tracks, still logs, and falls back to rule-based insights computed locally (if this, then that). That fallback is not "AI." It is a flowchart, and flowcharts are fine. What is not fine is silently shipping the same rows to a server because the silicon could not run the model.

A cloud coach can still be the right tool for a generic question ("what is an implementation intention") that does not include your private log. Do that in a browser, without pasting the CSV. Do not paste mood scores into a chat box and call it wellness. If you want a second opinion on the architecture, ask the vendor where inference runs, then ignore the adjective in the App Store subtitle. "Smart," "personal," and "coach" are not locations. Airplane mode is a location.

Design	You can verify	You cannot verify
On-device Apple Intelligence	Airplane mode, new log, new summary	Weights, but they never left the phone
Rule-based local insights	Same test, drier prose	Nothing interesting: it is if-then
Cloud coach	It fails in airplane mode	Where the prompt is stored, for how long, who fine-tunes

BJ Fogg's tiny habits need a celebration small enough to be honest. A generated paragraph that calls you a legend is not a celebration. It is a slot machine. Prefer a summary that names the skip. Prefer a suggestion that rewrites the cue, not your personality.

Do not title this as a brand fight. Named competitors are a different page. The architecture is the article: local model, local rows, optional backup of the rows if you choose an account, never a requirement that the model live elsewhere. If a vendor will not say "airplane mode" in a support article, believe the spinner.

Habit AI is a useful example of the local design: the coach runs on Apple Intelligence on the iPhone, habit data is not sent to an AI server because there is no AI server in that path, and you can verify it by logging in airplane mode. AI features need a capable device; everything else still works on iOS 17. No account is required to track. That is the whole claim.

If your current tracker fails the test, you do not need a personality crisis. You need a different coach, or no coach, and a cue in the kitchen. The model was never the habit.

Read next

Privacy in habit apps is the four-test version of this argument. Do you need an account to track habits is the identity half. For on-device coaching you can test with radios off, see how Habit AI works.