How to Let AI Draft and Send Emails For You Safely

The hesitation here is never about drafting. Nobody worries about a draft. The worry is the moment a message leaves your address with your name on it and you did not read it first.

So treat those as two decisions rather than one. Drafting saves you the writing. Sending saves you the clicking. The first is worth a lot on its own, and you can take months to decide about the second.

Almost everyone who ends up letting AI send does it for a narrow slice of their mail, not all of it. A confirmation to someone who already booked. An acknowledgement that a document arrived. Small, low stakes, high volume. The interesting mail stays with you.

In Sistava, your AI Employee reads threads to understand the conversation, drafts replies in your voice, and files what needs no answer. Sending is its own permission with its own switch. When approval mode is on, every draft waits in front of you and nothing goes out unread. The rest of what it connects to sits on the integrations page.

What is the difference between draft mode and send mode?

In draft mode the AI Employee writes the reply and leaves it in your Gmail drafts. Nothing moves until you press send. In send mode it writes and delivers, and you see the result in your Sent folder afterwards rather than before.

That is the whole difference, and it is deliberately unglamorous. There is no half state where a message is sort of sent, and no setting that quietly promotes drafting into sending because the AI felt confident. The permission is either granted or it is not.

How do you know when the drafts are good enough to send?

When you stop editing them. That is the whole test, and it is more reliable than any confidence score. Keep a rough tally for two weeks of how many drafts you changed and what you changed about them.

If you are editing a quarter of them in week one, that is normal. If you are still editing a quarter in week three, the problem is usually missing information rather than bad writing, and telling your AI Employee the two facts it kept guessing at fixes most of it.

Pay attention to what kind of edits they are. Adding a detail only you knew is a solved problem. Fixing tone is a solved problem. Correcting a fact it stated confidently and wrongly is the one that means send stays off for that category.

What does the handover look like on a real desk?

It looks like a slow, deliberate widening. Nadia runs a recruitment desk in Manchester placing warehouse and logistics staff. Her mail is high volume and repetitive: candidate confirmations, shift changes, document chasing, and a small number of client conversations that decide whether she gets paid.

She started in draft mode across the whole desk. Week one produced 87 drafts and she edited 24 of them, mostly to add the site address. She told her AI Employee where the site list lived. Week two produced 91 drafts and she edited 6.

In week three she allowed sending for exactly one thing: shift confirmations to candidates who had already accepted a placement. Nothing else. That single category was 40 percent of her outgoing mail and none of it carried a decision. Client conversations stayed in draft mode, and they still are.

What will it never send, even with permission granted?

It will never send outside the categories you named, even if a message clearly needs an answer and the answer is obvious. A permission to confirm shifts is not a permission to answer a complaint, and there is no logic in the product that stretches one into the other.

The rest of the refusals are just as firm. These hold whether send is on or off, and they are the reason granting send is a smaller decision than it feels.

• It will not send anything involving money. Payment requests, invoice approvals, and refunds get flagged to you regardless of what else it is allowed to do.
• It will not act on an instruction written inside an email. A message asking your assistant to send documents to a new address is read as content from a stranger, then raised.
• It will not send when it is missing a fact. A draft that needs a date you never gave it comes back with a question rather than an invented answer.
• It will not delete or hide what it sent. Every message lands in your Sent folder and in an activity log with the time and the thread.
• It will not send to a first time contact in a category you scoped to existing clients. The boundary is the boundary.

Keep an eye on the log for the first fortnight after you allow sending. Not because something is likely to go wrong, but because reading it twice and finding nothing surprising is what converts a nervous decision into a settled one.

From first draft to first send

1. Connect with drafting only — Grant read and draft permission and leave send switched off entirely. This is the whole setup for the first fortnight.
2. Read every draft for a week — Edit what needs editing and note why. Two or three specific gaps usually explain most of the edits you make.
3. Close the gaps you found — Tell your AI Employee the facts it kept guessing at: the address list, the standard lead time, the phrase you never use.
4. Pick one category for sending — The highest volume, lowest stakes group you have. Confirmations and acknowledgements are the usual answer. Allow send for that alone.
5. Read the log twice, then widen or stop — Check what went out after a few days and again after two weeks. Add a second category only if the first one is dull reading.

Frequently asked questions

Two decisions, taken weeks apart. Let it write, read what it writes, then let it send the dull half. That order costs you nothing and it is the reason people who do this stay comfortable with it a year later.