Last Tuesday at 11pm, I got stumped on a Tencent Meeting call

Lao Zhou was a client who hired me last month to run his Xiaohongshu account. That night he had a board meeting at the last minute and pulled me into the call to "chat casually." He pointed at our monthly data report and asked: "Xiaowei, why did this curve drop 30% on the 18th? How are you going to recover next month?"

I stared at the screen, frozen. Because I hadn't analyzed that curve — ChatGPT had written it. I hadn't even looked at the backend data myself, let alone figured out why it dropped or how to fix it. I could only mumble "let me go back and check."

Lao Zhou didn't press, but I knew he'd mentally tagged me with a question mark.

There's an academic version of this: homework A's, exam F's

MIT and Harvard recently ran a study. They had students use AI for assignments — the AI group scored 30% higher on the homework than the control group. AI really works, right?

But on the real exam — no AI allowed — the AI group scored 17% lower than the control. In plain terms: AI makes what you turn in look better and better, but what you actually know keeps shrinking.

A recent piece in The Atlantic calls this "cognitive outsourcing": we outsource "thinking" to AI, but our brains don't outsource, our skills don't outsource — the moment we leave AI, we show our true colors.

This is what got me on Lao Zhou's call. I outsourced the analysis, not the thinking.

How us small teams avoid being eaten by AI

After I messed up, I set three homemade rules. Not everyone needs all of them — pick one to try first:

1. Think for 3 minutes before opening AI
Before letting AI do anything, jot down three sentences in a note: what I'm trying to solve, what my guess is, what I'm worried about. Those three minutes force your brain to move first. After AI writes it, use your own judgment to poke holes.

2. One "no-AI day" per week
Say Wednesday — no AI writing all day. Hand-type client emails, hand-write outlines. That day shows you where you actually know your stuff versus where you're faking it. I keep Wednesdays; when my "email reply speed" drops, I know that muscle is weak.

3. Leave "my signature" on AI output
Before sending any AI-written work, add a hand-written section — even just two or three sentences: "My own judgment is..." This isn't for the client to see, it's to keep your brain in the room.

Replication cost: $0 money, 10–20 extra minutes daily, tech barrier = knowing how to use a notepad, first step = pick your "no-AI day" tonight.

What to try at different stages

If you're just starting out (no clients yet): I'd suggest holding off on AI. Walk the basics yourself — write 5 WeChat articles solo, build 3 proposals solo. Once you've felt the pain, AI becomes "acceleration" not "driving for you."

If you already have 1–2 clients: Use AI, but pick one workflow each week to do purely by hand. Like client comms by hand, content production with AI. This protects your most valuable muscle — talking to clients.

If you're scaling: The real bottleneck is no longer "doing the work" speed but "judging whether the work is good." Don't outsource that. Outsource that, and your company is a hollow shell no matter how big. I'm not there yet, but I hear my friends who've scaled keep saying this.

One last thing: you don't have to try this now. I only started thinking about it after stepping into Lao Zhou's trap that night. But if one day you find yourself losing confidence around AI, come back to this piece.