I often get asked: “Is it safe to upload this document to AI?” That question made me wonder how many people aren’t asking and are doing it right now.
AI is at the frontier of the workforce, and people are excited about its capabilities without pausing to consider the dangers. Right now, companies worldwide are sending personal, sensitive, and company data into AI without thinking twice about the repercussions. If your customer list gets out, your competition gains an edge, and it could create a PR disaster for your clients.
In a previous article, I wrote about the benefits of using AI this way, like uploading a contract and having AI review it. I’m not contradicting myself by saying you should never do it. I’m saying we need mechanisms in place before you do it.
Three questions people blur into one
We need to figure out what “protected” means. What does “safe” mean when you upload documents? People often blur three really big questions into one. Does the tool train on the data I give it? Who at the AI companies can see that data? Where does it live, and how long does it stay there? Those three questions define your data safety landscape.
Does the tool train on your data?
When you use a free model, like OpenAI or Claude, and don’t pay anything, that data is likely used to train their models. Here is what that means: When you ask a question or give it data, it responds. When you challenge it, it uses that to say, “Uh-oh, why did I get that wrong… why did the user challenge me?” and learns to get better and stronger. They claim they anonymize the data to keep sensitive information from getting out. But they have no contractual obligation to do that. You’re not paying anything. They have no reason to protect you.
This goes for some consumer AI as well. Even if you pay for it, there might be a checkbox you have to turn off to make sure the AI isn’t training on your data. That’s not the end of the conversation; that’s just the beginning.
Next, check your business account. If you have a business and you’re using AI, you should have a paid business account with your provider. Again, check your settings for a toggle to make sure any data training is turned off.
Once AI uses your data for training, it becomes their data, not yours. There’s no deleting. There’s no removing. You can’t ask for it back.
This ties into the other articles I’ve written. Here’s why. You can’t protect any data you can’t find. If you don’t know where it lives or who has access to it, you can’t even know what your team put in, because there’s no central repository to audit.
When I wrote about insurance questionnaires, one question is “how you protect confidential data and control access to it”. If an employee pastes confidential data into an AI model you don’t pay for, or that has training turned on, you should answer no on that form. Suddenly, your risk factor goes through the roof.
Then we have to talk about how AI can use the data inappropriately. In the first article, I wrote about how AI deleted a bunch of data out of a system we gave it access to. This is AI doing the same thing in a different pattern. It’s doing things beyond what you want it to because it has limited context or a rule set you can’t control.
Who else can see it?
As for who at the vendor can see it, that’s really unknown. We’ve seen many stories from other companies about employees accessing servers and looking at data. So far, we haven’t seen anything major from the big AI players, and they often build these so employees can’t do that. But again, it’s a trust factor. That doesn’t mean we need to stop using the tools. It means we need to understand the risk and accept the risk for what it is.
How long does it stay there?
Here’s the biggest piece. How long is your data stored on their servers? Personally, I’m not so worried about an employee at these companies looking at data. I’m more worried about a bad actor breaching an AI company and exfiltrating the data. Now all those documents, files, and chats you had are theirs. We know they use strong encryption and keep their infrastructure secure, but we’re all human, and we make mistakes. Heck, even AI makes mistakes.
So the last question is: how long do they store that data before deleting it? In technical terms, we call that a retention policy. Most of the large AI companies let you set a retention policy, but only if you pay for the enterprise version, which gets costly.
Not all data is sensitive
Here’s the flip side. Not all data is sensitive. AI drafting an email, helping you set up a job post, or answering general questions about business ideas doesn’t really need protection. If your team is using AI for those things, I wouldn’t worry about locking them out completely.
That said, being paranoid about everything is how you end up with what we call shadow AI. People use it anyway, but they hide it from you by setting up their own free accounts. Letting people use AI in your company in a controlled way is better than letting them use it on their own, uncontrolled. The goal is to distinguish between safe and unsafe, and to accept or tolerate the risk rather than getting rid of the tool altogether.
What to do this week
As always, I want to leave you with some action items.
First, find out which AI tools your team uses. Ask them. Let them know the goal is to improve it, not take it away. You’ll probably be surprised by what they’re using and what data they’re entering. In most cases, it’s not malicious; it’s because they just don’t know.
Second, make sure you’re paying for the business tier of whatever you’re going to use. Read the terms about training. Make sure it’s turned off and deployed to your team. Yes, it costs money. But the inverse costs your business.
Third, draw a really simple line. Here’s what’s allowed to go in. Here’s what’s not. This depends on what you pay for, your risk tolerance, and the model’s retention policy.
Fourth, connect any AI model you’re using to your tools directly, not by copying and pasting. Why? Because when you connect it directly, you can control what it can and can’t do. When you copy and paste, you’re just copying and pasting. Having someone copy an entire email thread that might include financial figures you don’t want to share is a bad idea. Giving AI access to that person’s email, with constraints on who it can read emails from and who it can’t, is a little better.
And lastly, think about this. If you’d be embarrassed to see your business in a headline about something you put in AI, it probably doesn’t go into a chatbot. Unless we’re talking about a specific enterprise use case where you can control the retention policy. Even then, the risk is still there, but it’s very low.
Here’s the path: set it up right, and your team can use the tools safely and securely, without having to be paranoid about security.
Remember, the security work you do isn’t supposed to be the boring stuff before you get AI working. It’s what makes AI safe to use on anything real in your company. It’s about knowing who holds the keys, what insurers require, where your data lives, and how long you retain it. If you set this up right and secure it, AI stops being a gamble. It becomes a powerful tool for your business.
In two weeks, I’ll get into automating a broken process. Most businesses have existing, broken processes they want to automate. Automation mimics you and your process, good or bad. If that sounds interesting, subscribe so we can send it to you.
If you have a tech decision in front of you, like AI, automation, security, whatever it is, and you want a straight read on it, that’s the kind of thing I like doing. You can grab some time with me here.