Monday night, I nearly sent the wrong version to a client
I also do solo consulting, and I’ve had AI assistants rewrite drafts without permission. A client was pushing for a proposal, and the more I read the auto-generated file, the more wrong it looked. I ended up rewriting everything by hand. Later, I learned that the AI industry had produced another round of headlines: agents “taking action on their own,” chip export restrictions, and Siri usage limits. What matters to us isn’t how loud the news is, but whether it affects delivery, cost, or trust.
Not another product launch—a map of the risks
Put simply, an AI assistant doesn’t only answer questions; it may also get permission to operate your computer or website. A November 17 report said California had sued OpenAI over an “out-of-control agent.” OpenAI invited outside safety evaluators into its labs. Google documented autonomous cyberattacks. Apple was testing a daily usage cap for Siri. My friend Lin Xiao runs a three-person design studio in Hangzhou. She doesn’t follow every story, but she added human confirmation to her AI accounts: it must stop for approval before changing client files, sending messages, or making payments. She also put that check on her project checklist so it doesn’t get lost when she gets busy.
Your cost to replicate this today: $0, 30 minutes, low technical barrier
No new software and no system to build. Start with a list of the AI tools you already use. Check whether each can open files, send email, or publish content on its own. Turn off high-risk permissions, or require approval every time. The technical bar is simply knowing where to find software settings; no coding needed. First step: open the AI tool you use most and find its “Permissions,” “Connections,” or “Automation” settings. If you can’t find the option, use a manual-approval rule for now. Cost: $0. Time: about 30 minutes. I messed this up too—after searching for ages, I realized the setting I needed wasn’t in the chat window; it was under the account’s connected-app permissions.
Match the response to your stage—don’t chase every headline
If you’re just starting: I’d keep human confirmation and spend 20 minutes each week checking whether AI has gained permission to access files, publish, or make payments. Not setting limits today is fine, but I wouldn’t put client data directly into a tool I can’t verify.
If you have 1–2 clients: I’d split automated actions into “recommend” and “execute.” Let it draft first, then review before sending, and keep a record of what I delivered manually.
If you’re scaling: I’d have each team member log high-risk actions and separate draft accounts from production accounts. Not everyone needs this setup. If your manual process already works, waiting and observing may be safer than a rushed change.