What to Hand to an AI Email Assistant, What to Review Before Send, and What to Keep Yourself

Most of the trouble with an AI email assistant is figuring out where to draw the line. Hand off too little and you keep doing work it could have handled for you. Hand off too much and one bad draft lands in front of the wrong person at the wrong moment.

A useful way to think about it: three buckets, and a deliberate ramp from one to the next, the same way you'd onboard a new hire instead of giving them the keys on day one.

The three buckets

Hand off completely. The assistant does the work end to end and you don't read what it produces. The cost of a small miss is small, and you can recover in under a minute.

Hand off, then read. The assistant drafts, files, or proposes. You read before it goes out. The cost of a miss is real but recoverable, the way a typo in an internal email is recoverable but a typo in a client email costs more.

Keep yourself. The assistant doesn't do this work. The cost of a miss is the kind you can't take back, a relationship bruised, a message you wish you hadn't sent.

Most people try one of these tools, put almost everything in the first bucket, and then pull work back the first time something feels off. That's the wrong direction. Start almost everything in the middle bucket, watch what the assistant gets right, and move pieces up as you see it earn the move. Keep a clear list of what stays in your hands on purpose.

What you can hand off completely by week two

Some of the work an assistant does is genuinely low-stakes. Summaries of long threads you've forwarded. Filing attachments into the right project. The morning brief that tells you what's open and what's gone quiet. Telling the assistant "remember that Jane is the new GC at Acme" and trusting it to do that. A weekly synthesis of your calendar.

All of these have a small recovery cost. A summary that misses one nuance reads slightly off, you catch it in ten seconds, you move on. A misfiled attachment can be moved back. The morning brief shaping up wrong for a week is a tuning conversation, not a crisis.

There's a reason this is safe with Deck specifically: every reply the assistant produces goes only to you. Even if the summary is wrong, nothing leaves your inbox. The worst case of a miss in this bucket is one email you read and a thirty-second correction.

What you should always read before it goes out

Anything that touches another person directly belongs in the second bucket. Drafts of replies before you send them. Replies to clients, your team, anyone you actually care about. Anything with numbers in it: prices, dates, capacity, deadlines. "Yes" or "no" to a request. Anything that sets an expectation with someone outside your head. Memory the assistant proposes writing on its own, before it gets baked in.

The assistant is competent at all of this. The read-before-send step is cheap insurance. With Deck it's the only path the product offers anyway: every draft lands in front of you before it goes anywhere.

This is the bucket most people live in for the first two weeks. By the end, you'll have a good sense of which categories you trust enough to move up.

What you keep yourself

There's a short list of work you should keep doing in your own voice no matter how good the assistant gets. Political judgment. Bad news to a person who deserves to hear it from you. Apologies. Personnel calls. The opening of any new important relationship. The hard conversation you've been putting off.

These aren't things an AI is bad at. They're things that prove you still run your relationships. An assistant that drafts the apology for you is a bad use of an assistant, no matter how well it writes. Keep those for yourself.

How the ramp usually goes

The first week is the boring one. Everything in the middle bucket. You read before send. The point isn't efficiency yet, it's learning where the assistant lands well and where it doesn't.

By the second week, two or three categories tend to move up: thread summaries, file routing, the morning brief. These are the easy wins.

In the third week, one or two more usually graduate, and routine drafts in your own voice move to a lighter read where you skim instead of editing word by word.

By the fourth week, most people land somewhere around six things they hand off completely, three or four things they read before send, and four to six things they keep entirely. That mix is stable for a long time after that.

When the calibration is off

You can tell you've moved too fast when something embarrassing actually gets sent because you skipped the read step, or when a memory entry the assistant wrote on its own is wrong and starts showing up in drafts, or when you delegated a judgment call and got back an answer that doesn't sound like you.

You can tell you're moving too slow when you're still reading every draft word by word in week four, when you haven't asked for a single routine because you don't trust the timing, or when one small miss in week one makes you pull everything back to the middle bucket.

The right calibration looks like trust that climbs steadily over weeks, not trust that swings.

What makes the ramp safe

The thing that lets you climb without anxiety is that the assistant only ever replies to you. The other people on a thread receive nothing from it, ever. That's not a setting you have to keep checking, it's how the product works. Even a bad draft in your hand-off-completely bucket stays inside your own inbox. Nobody on the outside sees the miss.

Combine that with the fact that the assistant only sees what you choose to forward or CC, and the floor under every category in every bucket is the same: a miss is a thing you correct privately, not a thing you have to apologize for.

Without that floor, every step up the ramp costs more nerve. With it, you can move at the pace you actually need to move.

For the longer story on how the assistant fits together, see why your inbox needs a sidekick.

Try it

Reserve your address. Start everything in the middle bucket. Move work up once you've watched the assistant earn it.

Pro is $29.99 a month billed annually, $34.99 a month billed monthly.

Frequently asked

Questions we get.

How fast can I move things to the hand-off-completely bucket?

Thread summaries and routine filing usually move up by the second week. Drafts in your voice tend to take three or four weeks of watching before a light read feels like enough.

What if the assistant never earns enough trust to move past the middle bucket?

That's a stable place to live. You're still saving the time it takes to write the first draft, and you're still in control of every send. Plenty of people are happy there for months.

Is the third bucket the same for everyone?

Roughly. The political work and the bad-news work belong to you no matter who you are. Some people also keep fundraising or personnel decisions out of the assistant's hands. Others are comfortable letting it draft those with heavy review.