Last Wednesday, I almost blew my only big client
That afternoon, my auto-reply email bot flagged a VIP's refund request as fraud and replied: "We don't handle requests like this." The client sent the screenshot straight back at me. I went pale. My first instinct was to say "the AI made its own call" — but I couldn't get the words out.
Why "the AI went rogue" is a terrible excuse
Eoin Higgins' piece, "There are no 'rogue' AI agents," cuts right at this. He points out Big Tech loves this dump: see, our product "accidentally" crossed a line — we're shocked too. But there are no rogue AIs — only people who didn't draw the line.
Same thing happened to us that day — I hadn't written "all refunds go to a human" into the prompt. This isn't a Big Tech-only problem. If you use AI support, AI writers, AI assistants, the rule's the same: tell it what it can and can't do, and it runs inside that line. Cross the line, and that's on your line — not the AI's.
What you can do today — 30 minutes
Cost: $0. Time: 30 minutes. Skill barrier: knowing how to edit a prompt.
Step 1: Pull up the prompt you last used to have AI reply to a customer or write copy. Look at it — does it say "don't do X"?
If not, add it now. For a support bot: "Anything involving refunds, complaints, or amounts over $X — escalate to a human." For a writer: "Don't fabricate facts the client never said. Don't use 'guaranteed' or 'definitely.'" — I got stuck here too, kept thinking "the AI is smart enough to figure it out," but it isn't.
How to use this at your stage
Just starting out (no clients yet): Not urgent, but do this one thing — write "what the AI must NOT do" on a sticky note and stick it on your screen. When your first client lands, just copy it over instead of patching after something breaks.
Got 1-2 clients: Spend 30 minutes this week reviewing every AI workflow. Ask yourself: "If it said the wrong thing tomorrow and a client screenshotted it online, could I excuse it with 'the AI went rogue'?" If not, fix the prompt now.
Scaling (5+ clients or a team): You need an "AI behavior boundary checklist" taped next to every tool's config page. We have 12 "AI must-not-do" items on a single A4 sheet now, and we recheck every prompt change against it. Not everyone needs this — but once your team crosses 3 people, you should have one.
The bottom line: AI isn't a scapegoat. It does what you told it to do. What you didn't spell out, it makes up for you.