On Monday morning, I almost deleted a batch of client records
Last Monday morning, I was in Hangzhou setting up an automated workflow for a client who sells knowledge-based courses. As soon as the AI recognized the “awaiting payment” tag, it prepared to archive the old records. I saw something was off and clicked stop immediately. I had also gotten stuck here before: trusting what merely “looked like it worked” nearly scrambled six months of client data. So I added a human checkpoint and started previewing every batch action. Honestly, for a small team like ours, the most expensive part may not be the tools. It is having nobody who can immediately understand what went wrong.
What I added wasn’t a new tool, but a brake
The method is simple: let the AI read the materials, propose a plan, and list everything it plans to add, modify, or send. It executes only after I confirm. Chen, who runs a secondhand bookstore, has AI compile consultation requests every day, but both “reply to the customer” and “record the order” require her approval. She reviews the list on a tablet at her Chengdu shop, then syncs it to her spreadsheet. Ajie, an independent designer, uses a similar approach for contract summaries in his Xiamen studio, with a human always checking the original files. Not everyone needs a tool, and it does not make the work easier overnight. It simply makes the places where things can go wrong visible.
Today’s test: 0 yuan, about 30 minutes
Money: 0 yuan. Use the AI chat tool already in your workflow. Time: about 30 minutes. Technical barrier: no programming required, but I need to distinguish “preview” from “execute.” First step: start a new conversation and write the rule, “List everything you will add, modify, or send. Do not execute directly. Wait for my confirmation.” Then drop in one non-sensitive test document and check whether the list is accurate. If your platform supports projects, I’d create three steps: organize, review, execute. If it doesn’t, use the chat history for manual confirmation. When payments, deletions, or outbound messages are involved, I would rather check twice. There is no urgency to try this today. The real test comes when AI changes something on its own and you need that brake.
Here’s how I would set it up at each stage
If I’m just starting and still handle most services manually, I would use it only to organize outlines and summarize consultations, while confirming all external content myself. If I already have one or two clients, I would separate client details, orders, and messages into different lists, with anything involving money, deletion, modification, or sending set as a fixed review point. If I’m scaling and my team has more than one person, I would keep a visible activity log for each person and define who approves and who executes. Connecting it to an automation platform can come later. I would run this for a week, then decide whether it is worth keeping.