Autopilot sounds great until the first wrong send
The pitch for fully automatic email is easy to like. Mail comes in, an AI answers it, you go and do something else. Nobody enjoys their inbox, so nobody argues.
We'd argue a little. The average reply isn't the problem, because models write a perfectly good average reply now. The problem is the one in a hundred that restates a price you never agreed, goes to the whole distribution list, or leaves from the wrong address. Once the other person has read a sent email, you can't unsend it in any way that matters.
So the question we'd ask of any AI email assistant isn't how well it writes. It's who decides when something leaves your mailbox. In Otto, by default, you do, every time.
Speed isn't the real prize anyway. The prize is skipping the boring part, and asking first keeps it. Drafting and finding are the boring parts, and so is getting things ready. Otto does those without waiting for you.
What asking first looks like
Ask Otto to reply, to delete, or to change a lot of mail at once and it prepares the work, then stops at a review card. The card shows the sending account and the recipients, plus the wording. For a bulk job it shows the scope too: which search, which mailbox, which action. Approve it, or edit it first, or throw it away.
Everything before that point moves freely. Otto reads threads and searches and summarizes without asking. It drafts and proposes without asking too, because none of that changes anything outside your own screen. The checkpoint sits where something becomes real: a send, a delete, a bulk change.
It's one narrow gate instead of a hundred small prompts. The fewer places you're asked, the more attention each one gets.
A walkthrough: invoice 2041
Say Maya at Studio Luma writes: "Can you confirm invoice 2041 goes out Friday? Priya needs it for month end."
You open the thread, press ⌘J and type: "Reply to Maya, confirm invoice 2041 goes out Friday, and thank her."
Otto drafts the reply in your voice and shows the card. Three things on it are worth reading, and we'd read only those.
The account. Is it going from your work address or your personal one? The text won't tell you, which is why this is the field that bites hardest.
The recipients. Priya was on the thread. If the draft is addressed to both of them and you meant to answer Maya alone, change it now. Reply and reply all read exactly the same in the body.
The facts. Invoice 2041 and Friday are what Maya will hold you to. If the draft says anything else, a different date or an amount you didn't mention, it came from somewhere in the thread, and that's worth tracking down.
Say Priya is on the card. You remove her and approve, and the reply goes. Then you glance at Sent to confirm. All of that takes about as long as rereading the subject line.
Or say the draft promised the invoice for Thursday. Type "make it Friday" and the draft changes before anything leaves. The card is a place to fix a draft, not only to accept or reject it.
Where automatic replies fit
Some replies you never want to see again. "Got it, thanks, I will look today" to the same supplier, every time. For those, Otto offers opt-in automatic replies, and opt-in means something specific here.
They're off until you accept a policy for one mailbox. You choose the allowed senders or domains and the topics. You set the active hours, a daily limit, and a cancellation delay of at least five minutes. Before you grant any permission to send, you can keep the policy draft-only and read samples of what it would have written.
A few more limits are built in. Replies go to the original sender, never to a group, and only for new incoming mail. Activity shows every candidate as queued, preparing, ready for review, sending, sent or uncertain, and you can cancel one before its window closes. Pausing the policy stops new candidates from starting.
That's the gap between this and autopilot. You pick the limits in advance, you can see them, and you can undo them, instead of finding out after something went wrong. Automatic replies are a Plus feature. The automatic replies guide has the full setup.
What asking first costs you
Attention. Approve every card without reading it and you've built autopilot with extra clicks. The safety was never there.
Two habits help. Shrink the request when you're unsure: "draft a reply" before "reply", "show me the messages" before "clean these up". And save your attention for the cards that touch a customer or a lot of mail, where a mistake is expensive.
Approval isn't completion either. It starts the work. A queued send still has to go out. A large bulk job runs in the background and can come back partial. Pausing an automatic-reply policy doesn't recall a send the provider already confirmed. If an outcome comes back uncertain, check Sent before you retry by hand, or you'll send the same email twice. AI that waits for your approval goes through those checks in detail.
Plenty of people will reasonably pick a more automatic tool, especially when most of their mail is routine and low stakes. The five questions in how to choose an AI email assistant will tell you whether one fits.
Try it on a reply you've been avoiding
Pick one low-stakes thread today. Ask for the reply, read the card, approve it, then look at Sent.
Free to start: one connected Gmail or Outlook account, with core mail, calendar and tasks and an AI allowance. Read the assistant guide or open Otto and try it on a real thread.
First published on the Otto Mail blog.
Top comments (0)