The inbox that started replying

A quiet thing happened in work email this year: the assistant stopped waiting for the compose box.

It drafts while you are in another tab. In some clients it offers to send. In a few, “helpful” is the default, and the undo is a toast you will miss if you look away.

That is not a productivity story. It is a permissions story.

What actually changed

People already lived with autocomplete. Autocomplete guesses the next word. A reply agent guesses the next relationship.

The difference is social. A wrong next-word is a typo. A wrong reply is a commitment, an apology, a no, a yes — issued in your cadence, with your signature, toward someone who will not ask whether a model wrote it.

If the product ships that loop on, the human default that moved is this: your silence used to mean you had not spoken. Now it might mean the software did.

Why a human should care

Junior staff feel it first. They cannot tell a client “the model sent that.” External partners treat the email as the person. Managers discover the tone after the damage.

The launch blog will talk about hours saved. Hours are not the unit. Reputation is.

What to watch

• Does send require a second, boring confirmation, or is confirmation the toast?
• Is the model allowed to see the whole thread, including the thing you would never forward?
• Who is on the hook in the audit log — you, or “assistant”?

If the reverse gear is smaller than the send button, the product has chosen for you.

We will not cover another “AI inbox” until that gear is legible.