The CTO's AI Playbook – Part 4: Your Employees Are Already Using AI. Do You Know What They're Doing With It?

​Picture a scene. It’s a Tuesday afternoon. A member of your finance team is preparing a board report and they’re running behind. They open ChatGPT, paste in a chunk of internal financial data, and ask it to help draft the executive summary. It takes thirty seconds. The summary is good. They’ve done this dozens of times.

You don’t know about it. Your security team doesn’t know about it. Nobody’s policy covers it.

This is Shadow AI and according to Menlo Security’s 2025 State of Shadow AI Report, 57% of employees are already inputting sensitive data into free-tier AI tools. Not maliciously. Not carelessly. Just practically, because the tool works, it’s free, and nobody told them not to.

I want to be careful here, because I think the instinctive CTO response to this “lock it down immediately” is often the wrong one. Or at least, it’s incomplete.

The employees using these tools aren’t trying to cause problems. They’ve discovered that AI genuinely makes them more productive, and they’re using the tools available to them. If your organization hasn’t given them a governed, safe alternative, that’s a supply problem as much as a behavior problem.

Banning external AI tools without replacing them with something better doesn’t eliminate Shadow AI. It drives it underground, where it becomes invisible to you and therefore more dangerous.

That said the risks are real, and they’re not trivial.

When an employee pastes confidential data into a public AI interface, that data is processed on third-party servers operating under terms of service your legal team has almost certainly not reviewed recently. Where it sits, how it’s used for training, what happens to it varies by provider and changes with updates that nobody reads.

The IBM Cost of a Data Breach Report 2025 found that AI-associated incidents cost organizations over $650,000 per breach, with organizations experiencing high levels of shadow AI usage incurring an additional $670,000 on top of that. These are not rounding errors.

Beyond the immediate data risk, there’s a compliance dimension that’s growing teeth. The EU AI Act, GDPR, financial services regulations, healthcare data law these frameworks create real obligations around how AI processes data that your employees are currently navigating in an information vacuum.

And then there’s the output problem. AI tools hallucinate. They confidently produce inaccurate information. Without oversight, that information enters workflows, gets sent to clients, or informs decisions with nobody checking whether it’s actually true. I’ve seen this happen with contract summaries, with market analysis, with customer communications. The consequences range from embarrassing to costly.

The Menlo Security report also found that 68% of employees are using free-tier AI through personal accounts specifically to avoid corporate monitoring systems. So even if you have visibility tools in place, they’re not seeing the full picture. This isn’t a technology problem you can solve with better tooling alone. It’s a culture and enablement challenge.

And here’s the counter-intuitive finding that should give any CTO pause: the organizations with the strictest AI bans tend to have the worst shadow AI problems, not the best. When you prohibit something that people have decided makes them more effective, you don’t stop behavior. You just make it covert. The employees who were previously using ChatGPT openly on a work laptop are now using it on their phone, on a personal account, with no visibility whatsoever. A blanket ban feels like risk control. In practice, it often creates more risk than it prevents.

Here’s the reframe I’d offer: Shadow AI is a signal, not just a risk.

It’s telling you that your people see value in these tools. It tells you that your organization hasn’t yet given them a governed way to access that value. And it’s telling you that the gap between what people want to do and what IT has sanctioned is wide enough that they’re crossing it on their own.

The response to that signal isn’t a memo. It’s a capability. Deploy enterprise AI tools Microsoft Copilot, Google Workspace AI, or purpose-built solutions within your security perimeter. Create an AI Acceptable Use Policy that’s readable and practical, not a 40-page legal document designed to cover every edge case. Classify AI use cases by risk level and apply governance proportionally using AI to draft a blog post is a very different proposition to using AI to process customer PII.

And train your people. Not just on what the policy says, but on why it exists, what the real risks look like, and how to use AI in a way that helps them and doesn’t hurt the organization. Education is genuinely your first line of defense here.

There’s a final point I want to make about governance and competitive advantage, because I think it gets lost.

Organizations that build robust AI governance frameworks aren’t moving slower. They’re moving more confidently. They can scale AI faster because the guardrails are already in place. They earn the trust of regulators and customers that lets them do things their less-governed competitors can’t yet attempt.

Shadow AI is already in your organization. That’s not a prediction it’s a fact of modern working life. The only variable is whether your response to that reality is a policy document that nobody reads, or a capability that works.

The most dangerous AI risk in most organizations right now isn’t the model. It’s the gap between what your governance says and what your people do. Close that gap and you’ve solved more than a security problem. You’ve built a foundation for AI that scales.

Next in the series: Part 5:  Escaping pilot purgatory: how to move from AI experiments to AI