Cleanups/On-device AI
Most apps would call all twelve of them AI. We would rather tell you exactly where a model is involved, what it is allowed to see, and what stops it from being wrong - partly because it is more honest, and partly because it means the other ten rules work on every phone.
ContactOne uses Apple's on-device system language model, reached through Apple's Foundation Models framework. It is part of Apple Intelligence, it ships with the operating system, and it runs entirely on the device. There is no ContactOne server in the path, because there is no ContactOne server that touches contact data at all.
It requires iOS 26 or newer on Apple Intelligence-capable hardware - iPhone 15 Pro and later. On any other device the AI section of the app is simply hidden, and nothing about the standard rules changes.
One job: proposing new cleanup rules from patterns already present in your address book. In practice that means two of the twelve rules - the dead-service link patterns behind old service links, and the label equivalences behind label normalization. It proposes. It never applies.
Aggregate, de-identified signals only. For the link rule, that means URL scheme prefixes with occurrence counts - something like “a scheme evernote:// appears seventeen times” - thresholded so one-off values never make it into the input at all. It never sees a specific URL, a name, a phone number, an email address, or a note.
The interesting engineering is not the prompt, it is the filter. Model output is checked deterministically before you ever see it, and a proposal is discarded if:
Correctness rests on those checks, not on the model behaving. That distinction is the whole design: a model that occasionally invents something is fine if invention cannot survive the filter, and it is not fine if the filter is a hopeful prompt.
That ContactOne uses ChatGPT or any cloud model - it does not, and no contact data leaves the device. And that the AI does the cleaning - it suggests rules, the deterministic engine does the work, and nothing is written until you approve it.
Which is why in the app those ten deterministic rules are labelled Recommended rather than filed under AI. Calling them AI would overstate the model and hide the more useful fact: they work on every supported phone, whether or not it can run Apple Intelligence.
