Apple put a language model in every new iPhone. Here's what it's good for

Writing a routine by hand is the reason most people never write one. So the first screen of Routine Run takes a sentence instead, and Apple's own model turns it into steps.

That model runs on the phone. No account, no server of ours, no bill per request, and it works in airplane mode. On iOS 27 a bigger version runs on Apple's Private Cloud Compute when the phone's model isn't enough. For an app that promises "private by default", this changes what is affordable: an AI feature that used to need a backend, an API key and a privacy page now needs none of them.

What it did well

Writing something new from one sentence. "Morning before a flight" comes back as a sensible list of steps with times, in the shape the app expects. Short, structured generation from a clear request is where a small model shines.

Where it failed, quietly

Asked to change an existing routine, it mixed up steps in ways that were hard to notice. Nothing crashed. A step moved, a time shifted, and the result still looked like a routine.

That is the dangerous kind of failure: plausible. So on the phone, the model writes new routines only, and changes go to a bigger model. We didn't decide that from a benchmark. We decided it from watching our own task fail.

Keep the prompt and the parser in one place

The prompt that asks the model for a routine is built from the same type that reads the answer back. And one test takes every example in the prompt and checks that it reads back into a routine.

That test doesn't judge the model. It stops a quieter bug: someone edits the prompt's example, the parser no longer accepts that shape, and the model starts producing answers the app throws away.

If you're adding it to your app

• Pick tasks that are short and structured: write, sort, extract, rename, summarise.
• Test your own task, with your own data, before choosing between the phone, Private Cloud Compute and a bigger model.
• Design the fallback first. Phones without Apple Intelligence should get a working screen, not an empty one.
• Watch for plausible failures. Measure what changed, not only whether something came back.