Last Wednesday night, a message that made my back go cold
Bloomberg reported: the Pentagon admitted that over-relying on AI for targeting led to missiles hitting an Iranian school.
My first reaction wasn't political—it was panic. I've done the same thing, and almost didn't recover.
Three months ago, I let AI reply to Lin-jie's first email
Lin-jie (a prospect who'd found me through our WeChat official account) emailed asking for a quote. I was slammed, so I told AI to "reply using our standard template."
Next morning I opened my inbox. Lin-jie had replied: "Feels like you didn't actually read my needs—I'll find someone else." A 50K deal, almost gone.
Lucky break: my partner Xiao Chen spotted the exchange in the backend and flagged it before I resent a human reply. I owned the mistake, sent a custom proposal, and Lin-jie stayed. Today she's my highest-value client.
If the Pentagon can't go full-auto, neither can we
The Pentagon has hundreds of billions in budget and a bench of AI experts—and still messed up. For 1-5 person teams like ours, judgment can never be fully handed to AI.
Judgment. Relationships. Risk. These three stay human. AI can draft, research, build tables. AI cannot set prices, make commitments, reply to sensitive emails, or handle complaints.
One thing you can do today: turn every "full-auto" into "half-auto."
- Cost: $0
- Time: 30 minutes
- Technical barrier: no code needed—just log into your tool's backend
- First step: open the tools you use daily (email, support, social backend) and switch every "auto-send," "auto-reply," "auto-publish" toggle to "draft for my review"
Advice by stage
Just starting out: If you haven't plugged AI into client comms yet, no rush—bookmark this and come back when you do.
Got 1-2 clients: If you want to act today, flip off "auto-reply" first and read every AI output yourself before it goes out.
Scaling up: If you're already scaling, consider a dedicated AI reviewer role—even part-time—to gate every AI output before it goes external.