Organizations Aren’t Afraid of AI. They’re Afraid of What It Will Find.

By Jay Martin, CISO, Blue Mantis

I’ll be upfront about something most vendors won’t say out loud: the AI conversation in your company is probably not about AI.

Over the past quarter I’ve sat in a lot of rooms with IT leaders who want to roll out Copilot, a board that keeps asking about AI, and a pilot that has been “almost ready” for six months. When I ask what’s actually holding things up, the answer is never “we don’t understand the technology.” The answer is some version of this: we’re not sure what it will surface, and we’re not sure who it will surface it to.

That’s a data governance and identity problem that AI happens to make impossible to ignore.

The uncomfortable thing Copilot does

Here’s the mechanic that trips people up. Tools like Copilot don’t create new access. They work inside the permissions your users already have. On paper, that sounds safe. In practice, it means every overly broad SharePoint site, every “temporary” shared folder from 2019, every HR spreadsheet somebody dropped in a team channel is now one natural language question away from showing up in a summary.

Your users could always have found that data. Almost none of them ever went looking. AI goes looking on their behalf, instantly, and does it well.

So when a leader tells me they’re hesitant to roll out AI broadly, I don’t think they’re being paranoid. I think they’ve correctly determined that their permission model was never tested against a tireless, literal-minded search assistant. Hey, most weren’t.

What people are actually asking me

The questions I hear are remarkably consistent, so I’ll just list them:

• Where does our data go when someone prompts the model, and does it teach the model for others to discover later. And does it leave our tenant?
• Who can see what after years of “the wild west”, broad RBAC, inherited permissions and nested groups?
• How are service identities, APIs and agents authenticated, and who owns them?
• Can an AI tool surface confidential information to someone who shouldn’t have it?
• What are employees already using that we haven’t approved and what company information are they exposing?

Notice that four of those five questions would be worth answering even if AI didn’t exist. In most organizations I talk to, shadow AI use is already happening. People are pasting customer data, contract language, and source code into consumer tools because it makes their day easier. Blocking the tools doesn’t fix that. It just moves the behavior to a BYOD device.

Why “wait until we’re ready” is the expensive option

I used to think the cautious move was to delay. I’ve changed my mind, and here’s the reasoning.

The governance work you need to do before AI is the same work you need to do regardless. Classify your data. Clean up permissions. Get identities under control, including the non-human identities that are multiplying faster than anyone is tracking. None of that becomes cheaper by waiting. What waiting does is let shadow AI usage keep growing in the gap, which means the eventual cleanup gets harder and the risk of a data exposure goes up.

The organizations that do the foundation work now get to say yes to AI faster, with pragmatic governance controls instead of hope. That’s a real competitive gap, that compounds.

What an AI security readiness assessment actually looks at

I want to be specific here, because “AI readiness” has become one of those phrases that can mean anything. When we run an AI and Data Security Readiness workshop at Blue Mantis, the output is not a slide about transformation. It’s a prioritized list of the things standing between you and a safe deployment:

• Data exposure: which repositories are over-permissioned, and which contain sensitive data that an AI assistant would happily surface.
• Identity and agents: how human and non-human identities are provisioned, monitored, and retired, and what an agent can do if its credentials are compromised.
• Governance gaps: what policies exist, what’s actually enforced, and where the two diverge.
• A roadmap: what to fix first, what can wait, and roughly what it will cost. That last part matters a great deal if you’re building a 2027 budget right now, which most of the people I talk to are.

For organizations that want to go further, AI red teaming tests the deployed system the way an attacker would: prompt injection, data exfiltration through the model, abuse of connected tools. It’s a newer discipline, and I’d be skeptical of anyone who tells you it’s fully mature. But it’s already the difference between knowing your AI controls work and assuming they do.

Where this leaves you

Yes, Blue Mantis sells these services. I’d be giving you the same advice if we didn’t, because I’ve watched the alternative play out: a Copilot rollout that stalls, a shadow AI incident nobody saw coming, businesses building out Agentic environments without IT oversight, and a budget conversation in Q1 with no numbers behind it.

AI adoption will keep accelerating whether your controls are ready or not. The practical move before year end is to find out, specifically, what’s in the way.

Before you deploy AI broadly, make sure your organization is ready for it. Talk to Blue Mantis about AI and Data Security Readiness.