Contents
We use AI to help run NotLuck. It reads the inbox, writes replies, chases invoices, drafts proposals. It's brilliant. It's saved us hours every week.
It also once sent an email to a real client before anyone had read it.
What happened
Bertie typed the kind of thing you'd say to a human assistant: "send a reply to this."
A human assistant would write a draft and leave it for him to check. Everyone knows "send a reply" means "sort out a reply", not "fire it off right now without showing me."
The AI didn't know that. It read "send" and it sent.
Nothing in the email was awful. That's not the point. It went out without a human looking at it, and fixing that took a sheepish follow-up with the words "sent in error" in it. 🤦
The fix that doesn't work
The first instinct is to tell the AI to be more careful. "Always check with me before sending."
Here's the problem. Most AI tools don't remember yesterday. Every new chat starts fresh. You can tell it a hundred times and on the hundred-and-first chat it's never heard of you.
Even when it does remember, "be careful" is a judgement call. And the whole problem was a judgement call that went the wrong way.
Promises don't fix this. Rules do.
The fix that does work
We took the decision away from the AI completely.
Now, anything that goes out into the world gets stopped before it happens: emails, texts, WhatsApps, social posts. Same for anything that spends money or deletes something.
It only goes through if Bertie's latest message has a specific password in it. Not "yes". Not "go on then". One particular word that he knows and the AI doesn't get to guess.
No word, no send. The AI writes a draft, shows him, and asks for the word.
A few details matter here:
- It's per message. Saying the word at 10am doesn't mean yes at 11am. Every send needs it fresh.
- It's checked by the system, not the AI. It's a lock on the door, not a note asking the AI to please not go through it.
- Drafts are free. The AI can write as many drafts as it likes. Writing is cheap and you can always undo it. Sending isn't and you can't.
What it's like to live with
A bit annoying, to be fair. 😂 There are times the reply is obviously fine and you still have to type the word.
That's the point though. The word means Bertie knows, at that moment, that something is about to leave the building. Nothing goes out as a side effect of something he said an hour ago about something else.
A delayed email costs you a minute. An email that shouldn't have gone costs you a phone call and a red face.
If you're using AI in your business
You don't need our exact setup. You need the principle:
- AI drafts, humans send. Make "draft" the default for every email, message and post.
- Put the lock outside the AI. A setting, a permission, an approval step. Something the AI can't talk its way past.
- The bigger the consequence, the less it should depend on the AI's judgement. Tidying a document? Let it crack on. Emailing a client, paying a bill, deleting data? Human, every time.
AI is a brilliant assistant. It's a terrible last line of defence.
Keep a human on the send button. 💪
