OpenAI, Google and Anthropic Wouldn't Promise Their AI Agents Follow the Rules. Can You Shut Yours Off?

Monday, three of the biggest AI companies sat in front of the New York City Council and got asked a simple question. Will your AI agents always follow your safety rules? The Council Speaker, Julie Menin, asked them to say it under oath. OpenAI, Anthropic and Google all passed. OpenAI's rep said it isn't possible to guarantee any technology is without risk. Anthropic's said the science of doing this is "fundamentally hard and unsettled." Google's said promising perfection wouldn't be possible with any product on the market.

I get it. That's an honest answer. But think about what it means if you run a business around here. The people who built the thing won't promise it behaves, and you're the one who put it on your inbox, your booking calendar, or your customer chat.

What I'd do about it this week

I'm not saying pull the plug on anything. I use AI agents myself. What I tell every group I train is that you can't count on the tool to police itself, so you need two things in place that don't depend on it. Neither one takes long.

First, can you see what it did? Open every AI tool that can act for you, meaning anything that sends email, books appointments, answers customers, or changes records. Find the activity history. Some of them have a clean log and some bury it three menus deep, but most have something. If you can't find one, that's your answer on that tool, and I'd keep it on a short leash until you can.

Second, do you know how to turn it off, and who's allowed to? I mean the actual steps. Where's the switch, who has the login, and could somebody do it at 7pm on a Friday when a customer calls upset about a weird text they got? I've walked into shops where the person who set up the AI chat left a year ago and nobody had the password. That's not a tool problem, that's a Monday problem waiting to happen.

Spot check it once a week

Once you can see the log, read five items from it every Friday. Just five. You're looking for anything you wouldn't have said or done yourself. It takes about ten minutes, and it's the thing that catches a problem when it's one wrong email instead of two hundred.

The council members were asking for accountability, and I think that's fair. But a hearing in New York isn't going to fix what your AI sends to a customer in Maumee tomorrow morning. That part's on you, and the good news is it's a small job. Pull up the log, find the off switch, write down who can use it.

If you want help going through the tools your team already uses, that's most of what I do in my workshops and one-on-one consultations for businesses around Toledo, Northwest Ohio, and Southeast Michigan. We can sit down and do this check together in an afternoon.