Perhaps you could have a privacy policy which clearly states no text from user emails will be looked at by humans.

It's possible, though not easy, for a limited form of client-side encryption to be used while still offering autocomplete; e.g. if each word is encrypted client-side with a per-client secret, an encrypted "next word" could be determined and returned without the server ever knowing the specific words and sentences it operated on (other than length). There are other caveats here.

> Privacy and security are extremely high on our priority list

The lack-luster privacy policy doesn't show that at all.