AI that waits for your approval

The scary part of AI email is not the writing

It is the sending.

Any decent model can produce a polite, plausible reply. That stopped being the hard problem a while ago. The question that decides whether you can actually trust an assistant with your inbox is narrower: what sits between "here is a draft" and "that message has gone, to fourteen people, from your work address, quoting a price you never checked"?

For ordinary assistant sends, deletes and bulk changes, the answer is a card. It shows recipients, wording and scope. You approve it, edit it or throw it away. Optional standing rules are separate: you explicitly set their limits before they can act without another card, and you can disable them.

Six seconds, three things

Reading one of those cards properly takes about as long as re-reading the subject line. There are three things worth your attention, and they are always the same three.

The sending account. This is the one that bites hardest, because the mistake is invisible in the text. Your reply is flawless and it went out from the wrong identity. If you run a work address and a personal one, or Gmail on one side and Microsoft 365 on the other, this is the field to look at first.

The recipients. Reply, or reply to all, or the alias that fans out to a customer's whole team. The draft body will not tell you which of those you are about to do.

Any number or date. A model writing from a long thread will happily restate a price, a deadline or a commitment it only half understood. Fluent sentences hide wrong facts beautifully. If a figure appears in the draft, it came from somewhere: go and look at where.

Scope is its own category

Bulk actions deserve a different kind of attention, because the damage is not one bad email, it is four thousand.

"Archive every message from that vendor since March" is a sentence with a range hiding inside it. Before approving, read what the job says it will touch: the filter, the count, the folders. A preview showing the first page of results is not proof of what the whole query matches.

When the scope is vague, narrow the request rather than approving and hoping. Ask Otto to find the messages first, look at what came back, then ask it to act on that.

Approval is not completion

This is the part people skip. Approving an action starts it; it does not finish it.

A queued send still has to go out. A bulk job of several thousand messages runs in the background and reports progress. A job can also come back partial, with some items done and some not. If something fails, check what already succeeded before you retry, or you will send the same email twice or create a duplicate event.

So the habit is two steps, not one: read the card, then read the result. The second step takes a glance.

Ask for the smaller step

When you are not sure, shrink the request. "Draft a reply" instead of "reply". "Show me the messages" instead of "clean these up". "Propose a time" instead of "book it". You lose a few seconds and you get a decision small enough to actually inspect.

Over a week or two you will stop doing this for the boring work, because you will have watched it behave correctly a hundred times. Keep doing it for anything that goes to a customer or touches a lot of mail.

Use this to judge any assistant

It is a useful lens when you are comparing tools, including this one. Where is the review point? What does it show you at that moment: recipients, wording, scope, counts? What does it tell you afterwards about what actually happened?

A draft button on its own answers none of those questions. Plenty of products have one.

Start where it does not matter

Pick a low-stakes reply today. Watch the card, approve it, check the result. Then try a small bulk change, and watch that too. Trust in a tool like this should be earned in the places where a mistake costs you nothing.

Core mail, calendar and tasks are free on one connected account, with an AI allowance on top. Read the assistant guide and the bulk actions guide for the ordinary review controls and result states.