I went shopping for AI that drafts email but doesn't send, and I was surprised how hard it was to find. Almost everything on the market wants to do the opposite — read your inbox and act on it, autonomously, in your name. Auto-reply. Auto-triage. Auto-schedule. "Let the agent run your inbox." Every product seemed to be racing toward the same promise: hand us your accounts and we'll handle it so you don't have to think about it.
I closed every one of those tabs. Not because the technology doesn't work — it often does — but because the thing they were offering to take off my hands was the one thing I wasn't willing to give up: the final call on what goes out under my name. This is the case for the other model — draft-only AI that you approve — and why I think it's the only model an accountable operator should trust.
Why "AI that drafts email but doesn't send" is the right default
Here's the asymmetry that decided it for me. When an assistant drafts something and gets it slightly wrong, the cost is a few seconds of editing. When an assistant sends something and gets it slightly wrong, the cost is a client who now thinks you said something you didn't, a deal you have to walk back, or a teammate you've quietly insulted. The downside of drafting is trivial. The downside of autonomous sending is unbounded, and it lands on you, because it went out in your name.
No model has the context you have. It doesn't know that this customer is one bad email away from churning, that this deal is politically delicate, that this teammate reads bluntness as hostility, that you'd phrase this one carefully because of something that happened three months ago in a meeting that was never written down. You hold all of that. An autonomous agent acting on your behalf is, by definition, acting without it. The draft is where that missing context gets supplied — by you, in two seconds, before it matters.
What you actually want from AI at work
Strip the hype away and the genuinely valuable thing AI does for an overloaded operator isn't doing the work. It's removing the cold-start cost of the work. The expensive part of replying to forty things isn't the typing — it's the context-switching, the re-reading, the "what was this thread about again," the staring at a blank reply box deciding how to start. A good draft collapses all of that. You go from a blank box to a near-finished message you just edit and approve. That's the leverage. And it requires zero autonomy to deliver.
So the question to ask any AI tool isn't "how much can it do without me?" It's "how much does it reduce the cost of me doing it?" Those sound similar and they are opposite philosophies. The first one is trying to remove you from the loop. The second one is trying to make the loop cheap. Only the second one is safe to point at your real relationships.
The control checklist
When I evaluate any AI that touches my work tools now, I run it against four questions. They're worth stealing:
- Does it act, or does it propose? Anything that takes an irreversible action — sending, replying, updating a record, booking time — without my explicit tap is disqualified. Proposing is fine. Acting is not.
- Whose account does the action go through? If it sends, it should send through my account, as my tap, so the sending itself is an act I performed — not something a third party did on my behalf with my credentials.
- What access did it take to do its job? Reading across my tools to brief me requires read access. It does not require the ability to send, delete, or modify. If a tool asks for write access it doesn't need, that's a tell about what it intends to do.
- Can I revoke it in one move? Read-scoped OAuth I can revoke from my own account settings, anytime, is a very different trust posture than handing over a password.
That last cluster is the whole difference, and it's a phrase I keep coming back to: a good tool gives you your context, never your keys. It reads across everything to tell you what needs you. It does not hold the ability to act as you.
"But isn't approving every draft just more work?"
This is the fair objection, and the honest answer is: only if the drafts are bad. If you're rewriting every message from scratch, the draft added nothing. But a draft written in your own voice, with the full context of the thread already loaded, is a two-second decision most of the time — approve, or nudge one sentence and approve. The approval isn't overhead. It's the moment your judgment enters, which is exactly where it should enter, and nowhere it shouldn't.
Compare that to the autonomous model's hidden tax: every message it sends on its own is a message you now have to monitor, because you're still accountable for it. You haven't removed yourself from the loop — you've just moved your attention from "approve before" to "audit after," which is slower, more anxious, and catches mistakes only once they've already gone out. Draft-and-approve puts your attention in the cheaper, safer place.
How this shapes what I built
Every morning, Standfast reads across my email, calendar, project tracker, and CRM and hands me one brief of what needs me — and for each item, the reply or update is already drafted in my voice, ready to approve. I read, I edit where it matters, I tap send. It never sends or acts on its own. Not as a feature flag I could turn off — as the design. The product's entire job is to make the loop cheap, never to remove me from it.
That choice costs something. A fully autonomous agent demos better; "it just handles everything" is a stronger sentence than "it drafts and you approve." But I'm not building a demo, I'm building something an operator points at the relationships their livelihood depends on. For that, restraint isn't a limitation — it's the entire value. It's the same reason I think of this less as automation and more as a chief of staff for operators: counsel that prepares the decision, never a system that takes it from you. And it's why, when you're trying to stop spending all morning on email, the answer isn't to let something else write your emails — it's to make writing them nearly free while keeping the call yours.
If you've also been looking for AI that drafts email but doesn't send — that gives you leverage without taking the wheel — that's exactly what I'm building. Join the early-access list.