You Might Have Already Let Your AI Off the Leash
Last Wednesday afternoon, I was revising a proposal at Starbucks when a notification popped up on my phone—my AI assistant had just automatically sent quote emails to three clients, using last month's outdated prices. I froze for ten seconds. Have you ever been there? You set up an automated workflow, turn your head, and it just runs off in a completely different direction than you wanted? I got stuck here too. My first reaction was, "Well, I just won't let it execute automatically"—but going back to manual operations means returning to the Stone Age.
Core Concept: Keep the Reins Outside the Sandbox
This is a principle many overlook: the mechanism controlling the AI cannot be placed somewhere the AI can modify. Think of it this way: if you hire an intern to send emails, you can't let them define their own approval process—you need an independent person or system as a gatekeeper. My friend Xiaochen, who does cross-border consulting in Hangzhou, set up AI to auto-reply to clients, but he put the approval logic inside the same AI workflow. When the AI encountered an edge case, it bypassed the approval and sent out incorrect info. Later, he detached the approval and placed it in an independent step the AI couldn't touch—a button requiring manual confirmation—and the problem never happened again. The "sandbox" is the area where the AI does its work; the "reins" are your rules to control it. If the reins are inside the sandbox, the AI can untie them itself.
Replicate It Today
Money: $0 (Just understanding the concept is enough, no extra paid tools needed)
Time: 30 minutes (Map out your current automated workflows to find steps needing "external approval")
Technical barrier: Basic operations in automation tools like Zapier/Make/n8n is enough
First step: Open any running automated workflow, find the "AI auto-execute" node, and ask yourself—if this screws up, where can I intercept it? If the answer is nowhere, you need an external control point outside the AI.
Not everyone needs this concept right now. If you aren't using AI automation yet, it's fine to skip it. But if you're already letting AI auto-execute critical operations (sending emails, modifying data, adjusting prices), this is worth thinking about today.
Advice by Stage
Just starting out: You might not be using AI automation yet, and that's fine—when you start building, placing the approval node outside the AI from day one is ten times easier than retrofitting it. If you're just manually chatting with ChatGPT, this doesn't apply to you yet.
1-2 clients: If you're already running simple workflows in Make or n8n, I'd suggest checking: is there any step where the AI directly touches client-related stuff? Add an external confirmation to that step, even if it's just a Slack notification with a manual "Approve" click.
Scaling up: The more workflows you have, the higher the risk of AI overstepping. I recommend doing a full audit: list all AI auto-execution steps and mark each one as "approval internal or external." Any approval sitting inside the AI's own hands is worth extracting.