Last Wednesday, AI’s “smart answer” fooled me again
At ten that night, I was helping a freelance client in Hangzhou streamline her quoting process. AI quickly produced a polished solution, but what she actually needed was simple: for projects over 5,000 yuan, ask about the budget before sending a formal quote. I had handed the entire decision to AI, so the answer circled for half an hour and cost more than usual. I only realized afterward that not every question needs a large language model. A spreadsheet or a few simple conditions can handle plenty of rules.
How I use it now: triage first, then bring in AI
Simply put, my decision model starts by asking, “Does this question need AI?” When a client asks, “Should we add a feature?”, I can decide using fixed conditions such as the project amount or stage. I bring in AI only when there are many documents, ambiguous language, or a response that needs variation. Our team now writes one clear rule before deciding whether to call a model.
I know a friend, Xiaolin, who runs a personal IP and sells educational content in Chengdu. She reviews a large volume of reader comments every week. Since last November, she has used fixed rules for common cases such as refunds and course inquiries. Only complaints and comments with obvious emotion go to AI for a draft reply, which she then reviews. Her takeaway: AI organizes and drafts, while people make judgments and commitments. She has had far less rework.
Try it today: spend one hour, buy nothing yet
You can replicate this for 0 yuan. Start with the spreadsheet, notes app, or AI chat page you already use; there is no need to buy dedicated software. Budget about one hour: take 20 real customer questions and sort them into two columns—“can be handled by a rule” and “needs AI assistance”—then turn five into simple rules. The technical barrier is one sentence: if you can copy, paste, and edit a condition, you can do this; no coding required. First step: open your usual AI tool, create a file called “Customer Question Triage Sheet,” and add the most common questions.
If I were just starting out, I would begin with 10 questions and avoid building a complicated system until they outgrow that. With one or two clients, I would turn quoting, delivery boundaries, and refund terms into fixed rules, leaving AI to handle only the variable parts. When scaling, I would log why each case was handed over and what a human changed before deciding whether automation is worth buying. Not everyone needs this setup. If you receive only two or three questions a day, keeping your current process may be perfectly fine. There is no urgency to try it now.