Drafts, never sends: keeping a human in the loop

The quickest way to lose a team's trust in AI is to let it do something in public that nobody checked. One odd email to a customer and the whole project gets switched off, along with everything useful it was doing.

So when I built an AI-run sales pipeline for a software startup, the first decision wasn't which AI to use. It was what the AI is never allowed to do.

The four rules

The pipeline runs three times a day. It picks up new signups, matches replies, notices bookings and no-shows, turns call transcripts into notes, and works out which follow-ups are due. It does all of that inside four rules.

1. Drafts only. It never sends.

Every email it writes lands as a draft. The founder reads it and presses send herself. This costs her a few seconds per email and buys something valuable: nothing reaches a customer that she hasn't seen. It also means a bad draft is a small annoyance instead of an incident.

2. Never email the same person twice about the same thing.

Automation fails by repeating itself. So the process has to check what has already gone out to a person before it drafts anything new. A double follow-up looks careless in a way that's hard to take back.

3. Ask, don't guess.

When a reply could belong to two people, or a transcript is unclear, it stops and asks. A question in the founder's report is cheap. A confident wrong guess buried in a customer record is expensive, because nobody knows it's there.

4. Never mark someone lost without a reason.

Closing a lead loses information. If the record says why, the founder can see patterns later: price, timing, a missing feature. If it just says "lost," that learning is gone.

Report back every time

After each run, the founder gets a short report: what came in, what was matched, what was drafted, what it wasn't sure about. When people can see what the AI did, they keep using it. When they can't, they quietly stop.

Rules belong in the process, not in your hopes

These limits aren't a line in a prompt saying "please be careful." They're part of the written process the AI follows. Where your tools allow it, enforce them there too: give the AI a connection that can create drafts and nothing else.

That's also what made it possible for a non-technical founder to run the whole thing herself. The rules are written in plain English, so she can read every one of them and change them.

When should AI send on its own?

Sometimes it should. Internal reminders, a receipt, a calendar confirmation. My rule of thumb: let it send when a mistake is cheap, easy to notice, and easy to take back. Customer conversations fail that test, so a person stays in the loop.

Start strict. It's easy to loosen a rule once the drafts have been good for a month. It's very hard to win trust back after the first bad send.