I'd hand an assistant the paperwork long before I'd hand it a customer. That isn't a guess about AI in general. It comes from how Mastan broke when people used it.
When we wrote up what happened with the people who used Mastan, the failures were ordinary. There were duplicate messages and slow replies, and more than once an assistant reported it had done something it hadn't. If the job is just tidying your own notes, I think those failures are easier to live with, as long as you read the notes and fix them. When the message goes to a customer, I'd want to read it before it's sent, so I can catch a duplicate or a promise I can't keep.
One person from that note shows where I'd draw the line. He started with personal reminders. About a month later he added Mastan to a group chat with colleagues and gave it a document task in front of everyone. By September it turned the team's meeting recordings and notes into paperwork.
I'd start with meeting paperwork because the team already has the recordings and notes, and the paperwork comes out of the team's own meetings. He also got there in steps: reminders first, then one task in the group, then a routine. If I were setting up something similar, I'd copy that order and keep the paperwork inside the team at first, where a colleague is likely to spot a mistake before a client does.
But look at the last line of that story in our note. Once it was routine, a missed reply mattered. So I'd keep checking completed work even after the task becomes routine.
Where I'd keep a person
First, anything a customer reads before you do. A reply to a complaint, a quote, a refund, a "yes, we'll deliver Friday". If the assistant sends a duplicate or a promise you can't keep, the customer sees it first and you hear about it later. I'd let it draft every one of those and send none of them.
Second, anything where "done" is a claim you can't see. We've had assistants report finished work that wasn't finished, so I wouldn't leave a job unchecked when the only proof is the assistant saying so. If it booked, paid or filed something, open the calendar, the bank or the folder and look.
And anything that needs a judgment about one person: whether to forgive a late payment, how to answer an angry regular, when to say no. Let the assistant draft those messages if that helps you get started. The decision behind them stays yours.
That leaves summaries, meeting paperwork, reminders and first drafts of documents your team will read.
One setting and one message
If you connect Gmail and Calendar, Mastan writes drafts and sends nothing by default. Sending can be turned on. My advice is to leave it off for anything that goes to a customer, even after a few good weeks, because reading a draft is cheap and a wrong email to a customer isn't.
Then tell the assistant where the line is, in plain words. Something like this:
For anything that goes to a customer, write a draft and don't send it. Show me the draft here first. When you say a task is done, tell me where I'll find it.
For the first couple of weeks, do this each time it reports a finished task:
- Open the place the work was supposed to land (the inbox, the calendar, the shared folder).
- Check it's there, and there once, not twice.
- If it's missing, do it yourself and tell the assistant in the chat what went wrong.
I'm not sure where the line ends up for every business. Before handing over another internal task, check the results of the last one, and move the line one job at a time, the way that team did.
Drafted with Mastan from real questions and our own data.