Scene hook
At 11 p.m. last night, I was still tapping “approve” on my phone, one item at a time.
That feeling was painfully familiar: AI had clearly saved us the heavy lifting up front, but in the end it dumped the fiddliest, most mentally draining confirmation work right back on us. I got stuck here too. I used to think being “careful” meant checking every single item. The result was clients during the day, catch-up approvals at night, and this constant feeling of being tied to the process.
What this method is + who’s already using it
Lately I’ve become much more convinced by a different approach: don’t put human review at every step; put it only at the critical checkpoints. For example, let low-risk content go out directly from AI. But if it involves pricing, delivery promises, or client privacy, then a person gives it a final look. In plain terms, this is not about removing humans. It’s about not letting small tasks drain all our human attention.
At 3 p.m. last Tuesday, I met Zhou Yan in a coworking space in Binjiang District, Hangzhou. She runs an English résumé service. She used to have her assistant watch every client reply generated by AI, sometimes well over a hundred in a day. Later, she switched to “auto-classify first, then review only the items with red labels.” Her assistant stopped spending the day copying and pasting nonstop and started focusing only on the 20% that could actually go wrong. I messed this up too. I used to chase 100% control. Later I realized that for a small team, protecting our energy matters more than surface-level perfection.
What it costs to replicate today
If I were trying this today, I’d put the setup cost as plainly as possible: money: 0-99 RMB; time: 30-60 minutes; technical barrier: knowing how to use a spreadsheet and an off-the-shelf automation tool is enough—no coding required. The first step is simple: open a Feishu Bitable, click “new blank table,” and split the AI-handled items from the past week into two columns: can be sent automatically, and must be reviewed by a human.
Then add just one rule: anything containing an amount of money, a delivery date, or a client’s personal information goes straight into “human confirmation”; everything else gets completed automatically first. The effect shows up fast: the amount that really needs our eyes on it is usually smaller than it feels. Of course, this isn’t necessary for everyone. If we’re only handling five items a day right now, doing it manually is still fine. There’s no need to rush into a process for the sake of having one.
Advice by stage: just starting / 1-2 clients / scaling up
If I were just starting out, I’d do only one thing first: list “high-risk content” separately. I wouldn’t jump straight into full automation. That usually feels steadier, and it’s easier to keep going with.
If I already had one or two clients, I’d split “client replies” from “pre-quote checks” first. Automate the parts of the replies that can be automated; always keep the quote-related check in human hands. That’s the setup I’d trust most.
If I were scaling, I’d start building a sampling system instead of full review: check only 10%-20% of AI output each day, record the types of mistakes, and hand that back into the next round of rule optimization. This isn’t about cutting corners. It’s about saving our limited attention for the judgments that are actually worth money.