At 10 a.m., I Almost Sent an AI Answer to a Client
Last Wednesday at 10 a.m., I was revising a client proposal in a Hangzhou café. In three minutes, AI had listed five recommendations in a confident tone, complete with a smooth risk disclaimer. I nearly copied them straight into a quote email. Then I checked two external sources and found that one recommendation was built for a different type of company. I panicked for a moment: the scariest part wasn’t that it was wrong, but that it sounded right.
I’ve probably seen you somewhere in this: AI drafts customer-service replies, contract explanations, or content topics, and then we spend half an hour fixing them. It saves typing time, but the responsibility for judgment still stays with us.
I Treat It Like an Intern Who Might Mislead Me, Not an Answer Machine
Now I use Claude by asking it to give the conclusion first, then list the evidence, counterexamples, and anything that needs human confirmation. Sometimes I also ask, “Which parts are just speculation?” It isn’t a fact database. Whether a citation is real and whether the answer fits your client still has to be verified.
I know an indie developer, Lin Zhe, who builds small websites for four clients. At 2 p.m. on Thursday, he had AI generate a project quote explanation, then checked every line against his previous service records. Any estimate without historical data behind it was marked “to be confirmed” rather than sent to the client. He says the extra step makes communication slightly slower, but reduces the pressure of having to explain pricing on the fly.
This doesn’t require a paid subscription. The free chat version can handle the initial structure just as well; paid plans simply tend to offer more usage and file-processing capacity. It isn’t a tool everyone needs either: if I’m occasionally polishing a paragraph, an ordinary chat assistant is enough.
Replication Cost: ¥0, 30 Minutes, No Coding Skills
Money: ¥0. A publicly available conversational tool is enough; extra spending isn’t required.
Time: About 30 minutes for the first setup, then 10 extra minutes to check each important output.
Technical barrier: No coding knowledge, but I must be willing to verify the evidence.
First step: Open the AI chat page I already use, start a new conversation, and paste these three lines: “Give the conclusion first; then list the evidence; finally mark any speculation, gaps, and points that need my confirmation.”
I’ve made this mistake too: when I first launched, I adopted a recommendation without even opening its links. When the client asked, “How do you prove this?”, I had to rewrite the whole section. Now I keep three steps: read the conclusion, check the evidence, retain human judgment. For important content, I switch sources and cross-check at least once. There’s no pressure to try today; starting with a low-risk message is enough.
Use It for Your Stage—No Need to Copy One Fixed Workflow
If I’m Just Starting Out
I’d use it to organize client requirements and draft replies, but I’d confirm every promise myself. The cost is low, so I wouldn’t rush into a long-term paid plan.
If I Already Have 1–2 Clients
I’d build a one-page “confirmation checklist”: pricing basis, delivery scope, and payment milestones, asking AI separately to flag omissions. I’d also keep real client feedback so it can inform the next round of checks.
If I’m Scaling
I’d bring AI into templates and first drafts, reserving human time for key decisions while tracking which steps fail most often. A shared team process can wait until recurring problems show a stable pattern; there’s no need to make it heavy from day one.
AI sounding polished doesn’t mean it has taken on the responsibility. Treating it like a new colleague whose work needs review feels safer.