On-device AI

AI that reads your email without sending it anywhere

What ChatInbox’s AI does, where it runs, what it never does, and why it would rather say nothing than guess.

When a mail app adds AI, the first question isn’t what it can do. It’s who else gets to read your email.

Many AI features work by sending your messages to a company’s servers, where a large model reads them. Your email, the words other people wrote to you, your receipts and your appointment confirmations can then pass through, and sometimes be stored by, a company you never chose to give them to.

Where ChatInbox’s AI runs

On your device, by default. ChatInbox uses Apple Intelligence, the language model built into iPhone, iPad and Mac on the devices that support it. Your email is read where it already is, and goes nowhere else.

A larger model, only if you choose it. A few jobs are too big for a model that fits on a phone, like summarizing a very long conversation or turning rambling notes into a polished email. For those, ChatInbox asks once whether to use Apple’s Private Cloud Compute, the same system that powers Apple Intelligence. Your data isn’t stored there, and it isn’t accessible to anyone, including Apple. It never passes through ChatInbox’s servers. Say no, and everything stays on your device. You can change your mind in Settings › AI Assistant › Where AI Runs.

Or not at all. Turn AI off in Settings › AI Assistant, and everything else in the app works the same.

What it does

  • Waiting on You: what people asked you, with the sentence they asked it in.
  • Catch Up: who’s waiting for a reply and what’s due, in one list.
  • Summaries, when you ask: of a long conversation or an attached document.
  • Questions about your mail, like “when is the dentist appointment?”
  • Writing: turning typed or dictated notes into an email, and suggesting answers to the questions in an email you’re replying to.

What it never does

  • It never moves, deletes or sends anything. A draft waits for you to keep it and press Send. Your mail moves only when you move it, or when a rule you wrote does.
  • It never invents a quote. When it says what an email asks, it has to point to the exact sentence, and the app checks the sentence is really there. If it isn’t, nothing is shown.
  • It doesn’t do the counting or the dates. “How many emails did the school send this week?” is answered by counting your mail, not by the model. Which day “next Friday” means is worked out from the day the email was sent.
  • It doesn’t cry wolf. Scam warnings need proof. An earlier version that leaned on hints flagged 14 of 987 real emails, every one of them wrongly. The current check flagged none of 1,120.

Why it would rather say nothing

Language models are fluent, and fluency is persuasive: a wrong summary sounds as sure as a right one. In email that’s a real problem. A question from your boss read the wrong way, or a to-do that was meant for someone else on the thread, is worse than no AI at all.

So ChatInbox gives its AI narrow jobs and checks everything it says. Sometimes that means a summary is shorter than you hoped, or a question gets no answer. We think that’s the right trade. A missing answer costs you a few seconds. A wrong one teaches you to stop trusting the feature.

It writes like you

When you ask it to draft an email, ChatInbox looks at how you write, on your device: how you usually greet people and sign off, how long your emails run. The draft sounds like you, not like an assistant. And anything you mentioned that didn’t make it into the draft, a name, a day, a number, is pointed out before you send.

Speaking and translating, on your device too

When you dictate notes for an email, your voice is turned into text on your device. When an email arrives in a language you don’t read, ChatInbox can translate it in place, keeping its layout, with the translation built into your iPhone, iPad or Mac. Neither needs a server of ours or a cloud model.

ChatInbox is a mail app for iPhone, iPad and Mac. Download it free on the App Store