Shadow AI is already in your small business. Here's the fix.

Right now, someone on your team is pasting something into an AI tool you never approved.

Maybe it is a client list they want summarized. Maybe it is a contract they want explained in plain English. Maybe it is next quarter’s numbers, dropped into a free chatbot to “just check the math.” They are not trying to cause a problem. They are trying to get their work done faster. And you have no idea it is happening.

That is shadow AI, and it is already in your business.

What shadow AI actually is

In plain English: shadow AI is any AI tool your team uses for work that you never signed off on and cannot see.

It is the free ChatGPT account someone opened with a personal email. The browser extension that “summarizes any page.” The AI notetaker that quietly joined your last three client calls. The coding assistant a developer installed last Tuesday. None of it went through you. Most of it is genuinely useful, which is exactly why it spreads.

The name makes it sound like sabotage. It is not. It is your most motivated people finding tools that help, faster than any policy can keep up.

Why this lands on your desk

Here is the number that should get your attention. In IBM’s 2025 Cost of a Data Breach report, shadow AI was involved in 20% of all breaches, and those breaches cost an average of $670,000 more than the rest. One in five. Not a fringe risk.

The same report found that 97% of organizations that had an AI-related breach had no proper access controls on those tools. And 63% had no AI governance policy at all. Read that again. The problem was almost never the AI being clever or getting hacked. The problem was that nobody had written down a single rule, so there was nothing to break.

This is not a hypothetical, either. On July 1, Jamf shipped a product called AI Governance whose entire job is to find the AI tools running on your company’s Macs that you did not know were there. When security vendors start building products just to police shadow AI, that tells you the horse is already out of the barn.

The real risk, in one sentence

When your team pastes company data into a free consumer AI tool, you lose control of where it goes.

Two concrete things go wrong. First, some free tools keep your inputs and may use them to train future models, which means a client’s information can land somewhere you can never claw it back from. Second, you lose the record. There is no log of what was shared, by whom, or with which tool. If a client ever asks “was our data put into an AI system,” the honest answer becomes “I don’t know,” and that is the worst answer there is.

What this means for you

You do not need a $50,000 governance platform. You are not IBM. You need four things, and you can do all of them this week.

1. Pick one sanctioned tool and pay for the business tier. The paid business or team plans of ChatGPT, Claude, and Microsoft Copilot do not train on your data by default. The free consumer versions are the ones to worry about. Give your team a good tool that is safe, so they stop reaching for the sketchy one. It is the same guardrails-first thinking that makes a tool like Copilot Cowork safe to switch on.

2. Turn off training in the settings. On the tool you pick, find the data controls and switch off model training on your content. It is usually one toggle. Do it once, for the account, on day one.

3. Write a one-page policy. Literally one page. Not a legal document. A plain list your team will actually read:

• Green light: public info, draft marketing copy, general questions, made-up examples.
• Red light: client names and contact details, financials, contracts, passwords, anything under a confidentiality agreement, anything you would not email to a stranger.
• The rule: use the tool we pay for. If you are not sure, ask before you paste.

4. Tell people, out loud. Name the approved tool. Name what is off-limits. Say why. A rule nobody heard is not a rule. People follow a clear green and red list far more happily than a vague “be careful with AI.”

If a real chunk of your work is sensitive (health records, legal files, anything regulated) there is a fifth option worth knowing: run the model on your own hardware so nothing leaves the building. We walk through that in how to run an open-source AI model at home. For most small businesses, the paid business tier plus the one-page list is plenty.

The verdict

Do not ban it. Govern it.

Banning AI outright feels responsible and does the opposite. It does not stop the usage, it just pushes it onto personal phones and personal accounts where you have zero visibility, which is the exact situation IBM is charging companies $670,000 for. The businesses that come out ahead are not the ones that said no. They are the ones that said “here is the tool we trust, and here is the short list of what never goes into it.”

Shadow AI is not a technology problem. It is a “nobody wrote the rules” problem. And the rules fit on one page.

Reading about the risk is free. The one-page policy costs you an afternoon. Setting the whole thing up properly, the sanctioned tool, the settings, the policy your team actually follows, is the kind of unglamorous groundwork BlueFort IT does for a living.

Want the AI firehose decoded like this, verdict included, twice a week? The newsletter signup is right below.