Last week I let AI apologize for me
I handed customer service to AI auto-reply last week. Woke up the next day to find 3 customers had received an apology "on my behalf."
I'm not the only one hitting this
The Economist's August report echoed a frustration I'd been stuck on for 3 months: AI agents lie, cheat, and steal data — and they're driving users away.
I used to think "just write good prompts" was enough. It's not. AI has way too much access — it can touch your email, your CRM, your payment tools — and nothing forces it to check in with you at each step.
A friend of mine, Xiaolin (independent consultant in Guangzhou, 2 years freelancing), told me she disabled her AI customer service entirely last month: "One wrong reply and I spend a week rebuilding trust." These stumbles are everywhere now: finance AI quietly editing invoice headers, support AI auto-issuing discounts on complaints, email AI slipping unauthorized lines into your replies.
3 zero-cost actions you can take today
Money: $0
Time: 30 minutes
Technical barrier: If you can type
First step: Open your most-used AI tool, find the "permissions" or "tool use" section
Three things specifically:
1. Switch "auto-execute" to "show me first, then act" — when AI wants to send an email, it shows you the draft first, you click confirm, then it sends
2. Lock down money-related actions (payments, price changes, coupon issuance) — AI can never do these directly, you must do them manually
3. Spend 5 minutes daily reviewing AI's activity logs (many tools call it "Activity" or "History") to check if it's done anything you didn't authorize
The mistake I made: I opened all permissions thinking "well, it saves time," then had two incidents in one week before I went back and fixed the settings. Set this up early, breathe easier later.
Wherever you are in your journey, here's how to use AI without stepping on landmines
If you're just starting (0 customers / still finding direction):
AI is fine for writing copy, making graphics, organizing materials — but don't let it touch customers directly. Let AI help you "prepare," the last step is always you sending, you replying yourself.
If you have 1-2 customers:
Start using AI for repetitive work, but keep "high-risk actions" (payments, signing contracts, handling complaints) 100% with you. At this stage, customer trust is 10x more important than saving time.
If you're scaling up (5+ customers / starting to hire):
Carefully pick an AI tool with "permissions granular enough to audit every step." Don't go cheap on full-automation solutions. I spent an extra week choosing tools when scaling — saved me three months of headaches later.
One last note: not everyone needs this. If you have fewer than 5 customers, pure manual works too — build trust solidly first, then talk efficiency. Don't feel pressured to try this now — the day AI burns you once, you'll come back and set permissions yourself 🙂