It's leaking from the tab your team already has open.
Most people think that their data is safe, or they don't think about it at all. That's the biggest problem.
When someone asks me whether AI is safe for their business data, I say that it's safe for anything that you would put out on the internet and make public yourself, if you're just using it normally. Anything that you wouldn't put out publicly, don't put into an AI that's connected to the internet. If you want to use it for your personal, private or financial data, I suggest using an in-house LLM.
Yes, use security measures in your AI settings, of course. Use guardrails in your prompts and things like that for sure, and in your setup and everything you do. But if it's going out to the internet, it's going out to the internet.
You also have to be really careful about what comes back from the internet. We ran into an issue where we had false information coming back that was trying to trick my AI into doing malicious things. Luckily, my AI and I both caught it.
The part everybody gets backwards
AI can't help with your business if it knows nothing about your business. So what matters is whose AI your information goes into.
My information goes into my own account, always.
Now picture somebody on your team pasting a client contract into a free tool on their phone at nine at night. The same information lands on an account nobody at your company manages, with retention settings nobody has read.
What we actually do here
My AI can see my calendar, my email and my Notion database, but only the databases that I allow it to see. That's my contacts, my networking information, my task lists, or anything else that I have it work on.
Now, don't get me wrong, my AI does have access to my calendar, email and Notion database but does not have rights to send email or permanently delete any of my information. It can only make drafts. That's guardrails at work. You just have to make sure you have guardrails and use your common sense.
What can it do on its own? It can create calendar events, create draft emails, and add to my databases. It can create databases and update databases. I approve actions as they come through.
The training setting is off on every AI tool we use, so our information isn't used to improve anybody's model.
What a governance framework actually asks
It starts with three questions about any tool anyone at your company is using. What can it see? What can it reach? What can it do on its own?
You need to know what it can reach on your own machine or your own phone. If you own a business, you also need to know what other AI systems your employees are using that aren't approved. You need to have policies that follow up on that, and you need to let your people know the guidelines.
Anything that touches money, client data, or something you can't undo should have a person in front of it.
Where we started
When we worked through this for ourselves, the training setting came first. Most tools have a switch controlling whether your conversations are used to train the model, and in our experience almost nobody has looked at it.
The thing to actually stop doing
Stop assuming nobody on your team is using AI yet. They are, on their phones and personal accounts, with your work in the prompt.
Check and see what's being used. Ask your people what they are using. Be sure they're not using their own personal AI and putting work information into it, and that they're only using the approved tools, the stuff that the business says they can use.
You can't govern something you don't know about.