How does an AI email client work?

An AI email client is first of all a normal mail program. It signs in to your provider over IMAP, downloads headers and bodies, and sends over SMTP. That is the same plumbing Outlook and Apple Mail have used for decades.

One practical consequence often gets lost: your mailbox stays where it is. An AI email client is a different view onto the same account, not a new provider. Delete it again and your mail sits unchanged with your old provider. The exception is clients that only connect Gmail or Outlook: those require an account with exactly those two.

As soon as a mail arrives, its text goes to a language model with a narrow job: what is this. The result is not prose but a handful of fields: a category (offer, invoice, newsletter, meeting request), a read on urgency, and often a derived task with a deadline when the text contains one.

Those fields are the real difference from a filter. A filter knows senders and keywords. A model understands that "we need a decision by Friday" is a deadline, even though the word deadline never appears.

Because this runs per mail and costs compute, good clients use the small fast model here and save the large one for the jobs that are worth it.

Summarising a twenty-message thread and writing a usable draft take considerably more compute than classifying. So they do not run automatically for every mail; they run when you open a mail or hit reply.

A draft reply is not a template. The model gets the thread, your earlier replies to the same person, and the language the mail arrived in. Out comes text in your register and in the right language, which you read, adjust and send. The decision stays with you, and it should.

The daily brief many clients show in the morning is the same mechanism applied to the whole mailbox rather than one thread: what is due today, who has been waiting how many days, which promise you made and have not kept.

Here is the honest part. For a model to summarise a mail, it has to read the mail. The text therefore leaves your device and goes to the model vendor. That is true of every AI email client, without exception, and anyone claiming otherwise is either computing entirely on the device, which is a different quality of result, or is not telling the truth.

The useful question is therefore not whether data is processed, but three others: which model, in which data centre, and what happens to the text afterwards. With a cleanly built client the answer to the third is: nothing. The request goes through, the answer comes back, nothing is kept.

Two things should never reach the vendor’s server: your mail password and your mail archive. Both belong encrypted on your device, in the operating system keystore. If a vendor stores your IMAP password to fetch on your behalf, that is a different architecture with a different risk.

• It connects your existing mailbox over IMAP instead of demanding a migration.
• It says which model it uses and which country it runs in. If that is written nowhere, that is the answer.
• It stores no mail contents and no passwords on its server, and writes that down in a checkable way.
• It shows you the draft instead of sending it. Autopilot belongs on routine, not on decisions.
• It names an allowance in numbers rather than "unlimited". There is no unlimited compute, only an unstated limit.

BoxBee is built this way. IMAP and SMTP, so your existing mailbox. Classifying on arrival, summarising and drafting on demand. The model is Mistral, developed in France, and requests run on servers in Europe. Mail contents and passwords do not sit on the server; credentials stay encrypted in your device keystore.

The free tier includes 30 AI actions a month, which is enough to try everything described above without a payment method.