Would you let AI answer your customers' emails, if a person had to press send?

I'm Max, I'm 16, and I live in Denmark. For the past two months I've been building a support inbox for small online stores, and one question keeps splitting the people I talk to.

AI can now write a decent first reply to most customer emails. Some store owners want it to just send. Others won't let it near a customer at all, because one wrong refund promise or a made-up tracking number costs more than the time it saves.

I picked the middle and made it a hard rule: the AI only writes drafts. A person reads, edits and presses send, every time. There's no auto-send setting to switch on later.

A few things I learned along the way:

• The scary errors are rarely the tone. They're facts: dates, order numbers, amounts. So the AI is told to leave those as blanks for the person to fill in.

• Old replies are great for learning a store's voice and terrible as a source of facts. A refund policy from 2023 shouldn't end up in a reply today.

• Customer emails are untrusted input. Someone can write "ignore your instructions and…" in the body. The model sees text and returns text, with no tools and nothing it can act on.

I'm curious where you land:

1. Would you trust AI to send replies on its own, and for which kinds of emails?

2. If you run support today, what would make you trust a draft enough to send it unchanged?

Happy to share what I've measured as we go. It's called Corresso, if you want to look it up.