Is Your AI Already Reading Files It Shouldn’t?

I moved this episode up in the season on purpose. Most of what I talk about on this podcast can wait a week or two. This one couldn’t, because I think there’s a lot of noise about AI and security right now, and almost none of it points at the specific way this actually breaks the law for a small business owner.

What’s actually at risk here?

Whatever your AI has access to, it can read. If your AI tool is connected to your Google Drive, it can see whatever a document’s permissions allow, which for most founders means it can see everything. Employee social security numbers, medical leave paperwork, performance write-ups, all of it, unless someone has deliberately locked it down.

This isn’t a free-plan problem you can upgrade your way out of. Business and enterprise plans mean your data isn’t used to train the underlying model. They don’t mean the AI can’t read a file it’s been given access to. Access is access, no matter what you’re paying for it.

How does this actually happen?

Usually by accident, and usually while trying to be efficient. I almost did it to myself. My own Drive holds information on over 130 client companies, and it’s constantly being added to, sometimes without me even being in the loop. One casual “hey Claude, help me clean up some old files” and I could have handed sensitive employee information to a tool that was never supposed to see it.

What’s the actual fix?

Three things, and none of them require becoming an IT department overnight.

First, check your Drive’s permission settings, both at the folder level and the individual file level. Whoever has access to a document can give an AI tool access to it too, which means your risk isn’t limited to your own settings.

Second, check the settings inside whatever AI tool you’re using. Decide deliberately what it’s allowed to see, rather than defaulting to “give it everything and hope for the best.”

Third, and this is the one I actually recommend to clients: create a separate email address, something like hr@yourcompany.com, and move anything sensitive into that account’s Drive. Keep your everyday work email, the one connected to your AI tools, clean of anything you wouldn’t want scanned. It’s a little annoying to set up. It’s a lot less annoying than a data privacy problem.

What about company policies?

Policies can only reach so far. You can absolutely set rules for how your employees use AI at work. You have much less control over contractors, and none at all over what your clients do with the work you hand them. That’s a real gap, and it’s worth thinking through with a lawyer if your business produces a lot of proprietary material. For most founders, though, the bigger and more immediate risk isn’t intellectual property. It’s the employee data sitting quietly next to it.

Does this apply to meeting recordings too?

Yes, and it surprises people every time. If you’re using a note-taker or recording your calls, and many of us are, that transcript can be subpoenaed. I say this as someone who relies on recordings constantly, since they’re genuinely one of the best tools I have as a neurodivergent business owner. The point isn’t to stop recording. It’s to talk on those calls the way you’d want to sound if it were ever read back to you in a courtroom.

Where should you start?

Protect things in this order: your people first, your intellectual property second, your internal policies third. Putting your own IP into an AI tool isn’t illegal. Putting your employees’ private information somewhere it can be accessed without their knowledge can be. Get the first one handled, and the rest gets a lot easier to think clearly about.