Shadow AI is your real breach

Ask a company what its biggest AI risk is and you'll hear about the model — bias, hallucination, a misbehaving agent. The actual incident, the one already occurring across most organizations, is much more mundane: your own employees feeding confidential information into AI tools that IT never approved and can't see. It's called shadow AI, and the numbers are not subtle.

71% of workers have used unapproved AI tools at work, 57% actively hide that they're doing it, and roughly 98% of organizations report some unsanctioned AI use. It's not just the interns: 90% of security professionals admit to using unapproved AI tools, and a strong majority of executives openly prioritize speed over data privacy. Meanwhile one in five companies has already suffered a breach tied directly to shadow AI, at an average of $670,000 in extra cost per incident. That's the breach most companies aren't looking at, because they're busy watching the model.

It's shadow IT, and the dynamic is identical

None of this is new in shape. We've seen it before: people couldn't get the tool they needed through official channels, so they used Dropbox, or their personal Gmail, or a SaaS app nobody vetted — shadow IT. Shadow AI is the same story with a faster, more capable, more data-hungry tool, and the same root cause: people route around friction. They're not trying to leak data. They're trying to get their work done, the sanctioned path is slow or doesn't exist — 63% of organizations have no AI usage policy at all — so they reach for the fast unofficial one and keep quiet about it.

That reframing matters, because it tells you the problem isn't really a security problem first. It's a product-and-process problem that becomes a security problem. Your people want to use AI more than you've given them a safe way to, and shadow AI is the size of that gap.

Banning it is the worst move

The instinct, once leadership sees these numbers, is to clamp down: block the tools, forbid the use, mandate it away. That is precisely the move that created the problem. A ban doesn't remove the demand; it just removes your visibility into it. The productivity is real and people won't give it up, so a blanket ban drives the usage underground, where you can't monitor it, can't set boundaries, and can't even know what data went where. You trade a manageable risk you can see for an unmanageable one you can't.

The data backs this up bluntly: bans get circumvented, but when companies provide approved tools, unauthorized use drops by 89%. The lever isn't prohibition. It's a sanctioned path that's good enough that nobody needs the shadow one.

What actually works

The answer to shadow AI is the same as the answer to shadow IT: make the safe way the easy way. Concretely:

• Provide a sanctioned, fast, safe option. The single most effective move. If your approved AI is as good and as quick as ChatGPT, the reason to go around you evaporates. Most shadow AI is people reaching for capability you didn't offer.
• Set clear data boundaries, not vague fear. People hide usage partly because the rules are unwritten. Tell them plainly what they can and can't put into AI — customer data, secrets, source code — so the careful majority can comply instead of guessing.
• Coach in real time; don't just hard-block. A warning at the moment someone pastes something sensitive ("this looks like a customer record — here's the approved tool") changes behavior without breaking the work. Hard blocks just send them to their phone.
• Keep a living inventory and audit. You can't govern what you can't see. Discovering what AI is actually in use is step one — and it's a moving target, so it's ongoing, not a one-time scan.

This is the same lesson as the bigger safety point: the dangerous part of AI usually isn't the model, it's the system and the human process around it — here, the gap between sanctioned and actual use.

The bottom line

The AI breach you should worry about most isn't the sci-fi one where a model goes rogue. It's the deeply ordinary one already happening: a well-meaning employee, under deadline, pasting something they shouldn't into a tool you don't control, and not telling you. It's widespread, it's hidden, and it's expensive — and it exists because the demand for AI at work has outrun the safe supply.

So stop treating shadow AI as a discipline problem and start treating it as a supply problem. Don't ban the thing people obviously need; give them a version that's safe and fast enough that going around you stops making sense, draw the data lines clearly, and watch for the rest. Close the gap between how much your people want to use AI and how safely they can — because right now that gap is wide open, and it's leaking.