In short: autocomplete assumes a reply is mostly a typing problem. It usually isn’t — it’s a decision problem. And in a search box a wrong suggestion costs a keystroke; in a reply to your manager it costs a position you didn’t choose. So we built the part before the words instead of the words themselves.
Autocomplete is the first feature anyone suggests for a tool like this, and we understood why. You type half a sentence, it finishes it. It’s the most visible proof that the AI is doing something.
We tried the shape of it early and then stopped. Here’s the reasoning, because it explains most of the other decisions in the product.
Wasn’t autocomplete the obvious thing to build?
Yes. And in plenty of places it’s the right call — code editors, search boxes, spreadsheets. Those are domains where the next token really is predictable from the previous ones, and where a wrong guess is free: you ignore it and keep typing.
A reply is different in both ways. The next words are not predictable from the previous ones, because what comes next depends on a decision that hasn’t been made yet. And a wrong guess isn’t free, because these words go to a person who will remember them.
So the feature that demonstrates the AI most clearly is also the one that sits at exactly the wrong point in the process — at the typing, rather than at the decision.
What does autocomplete assume about a reply?
Three things, and none of them hold:
That the hard part is producing words. For most difficult replies it isn’t. The words are the cheap part; what takes ten minutes is deciding whether to commit, refuse, or ask for time. Autocomplete makes the cheap part cheaper and leaves the expensive part untouched — which is why a faster draft doesn’t make the message feel easier, as we tried to show in the twelve minutes before you hit send.
That the next words follow from the last ones. In a reply they follow from context that isn’t in the message: how long you’ve known this person, whether you already missed a deadline, what you’re willing to be held to. A model completing your sentence doesn’t have that, and will fill from the most likely pattern instead — fluent, generic, and slightly wrong.
That a suggestion is easy to reject. Suggestions are not neutral. Once a sentence appears, it anchors: you’re now editing someone else’s position rather than choosing your own, and the default becomes agreement. That’s a real shift in who is making the decision, disguised as a convenience.
Why is being wrong here expensive?
Because the failure isn’t a typo.
In a search box, a bad suggestion costs a keystroke. In a reply, a suggestion that guesses a date, a commitment, or a level of warmth gets sent — and now it’s a promise you didn’t make, in writing, to someone who’ll hold you to it. There’s no undo that doesn’t involve a second, more awkward message.
And the errors are quiet. Autocomplete doesn’t produce obviously wrong text; it produces plausible text, which is precisely what makes it hard to audit. You’d have to read every generated clause as if you hadn’t written it — and reading carefully is slower than writing.
Put those together and the feature has an odd property: it saves time when it’s right and costs more time than it saved when it’s wrong, because now you’re checking. That’s a bad trade for a tool whose whole purpose is to reduce the weight of a reply.
What did we build instead?
Not “less AI.” AI pointed at a different moment.
- Read the situation first. Warmth or firmness, brief or detailed — settled out loud, early. Most of the steering disappears once that’s decided.
- Offer directions, not a verdict. When a reply could reasonably go three ways, you see three ways and choose in a click. You’re still deciding; you’re just no longer deciding alone and unprompted.
- Ask when it doesn’t know. If a date or a number is yours alone, it should stop and ask rather than invent one. That’s the whole argument in why it asks instead of guessing, and in you don’t need a smarter AI, one that asks.
- Never send for you. You decide, you send. The tool has no ability to reply on your behalf, which is the one guarantee that makes the rest safe to use.
The cost of this design is real: it’s slower than autocomplete when autocomplete would have been right, and it asks you things you’d sometimes rather not answer. We think that’s the correct trade for words that go to a person. Plenty of tools will disagree, and for some kinds of writing they should.
Will you add autocomplete later?
Not in the form described here — completing your sentences as you type toward a person. That would put the tool back on the wrong side of the decision, which is the whole thing this design exists to avoid. It isn’t a feature we’ve postponed; it’s one we think belongs to a different kind of tool.
We’d rather be trusted than be impressive. Join the waitlist — BreatheReply is a Windows app that reads the situation, offers directions, and asks when a fact is yours alone. It never sends anything for you: you decide, you send.