AI That Earns Its Place

AI That Earns Its Place, an article by Hessam Zahedi

Why secure AI adoption should start with the controls you already pay for

Almost every conversation about AI at work starts as an argument with two sides.

One side wants to block it. There are too many unknowns, too much sensitive data, and too many stories about prompts ending up where they shouldn’t. Put the popular tools on the deny list and revisit next year.

The other side wants to let it run. The productivity gains are real, competitors are moving, and nobody wants to be the department that said no to the future.

Both positions are understandable. Both fail, and for the same reason: they treat AI adoption as one decision, when it’s really a sequence of them. Each step should be small enough to reverse and clear enough to explain. And every tool, whether it’s the assistant itself or the guardrail around it, should earn its place before it gets one.

The Problem Isn’t AI. It’s AI No One Can See.

The case against doing nothing is now hard to argue with. IBM’s Cost of a Data Breach Report 2026 found that shadow AI, meaning staff using AI tools the organization never approved, was involved in 43 percent of the incidents studied, up from 20 percent a year earlier. Those incidents averaged $5.39 million, above the global average of $4.99 million. And among organizations that suffered an AI-related breach, 92 percent lacked proper access controls on their AI systems.

The case against banning is just as clear. Surveys this year keep landing in the same place: about half of employees say they use AI tools their employer hasn’t approved (one survey of 2,000 workers put it at 49 percent). Another, of 2,001 UK and US employees, found that senior decision-makers were more than twice as likely to use unapproved tools as the people they manage. The people writing the policy are often the first to work around it.

Put those findings together and the picture is consistent. The risk isn’t that people use AI. It’s that they use it where no one can see what goes in, what comes out, or where it’s kept.

Why the Ban Fails

A block on a handful of domains feels decisive. In practice it moves the work, not the risk. The employee with a deadline opens the same tool on a personal phone or a home laptop, outside every control you have, and the sensitive document goes with them. You’ve traded a risk you could see for one you can’t.

Bans also age badly. New AI tools appear every week, and AI features keep arriving inside software you’ve already approved: the meeting recorder, the CRM, the design tool, the note-taking app. A deny list can’t keep pace with a market moving that fast, and the people maintaining it end up in a permanent game of catch-up.

And a ban carries a cost that never shows up in a risk register: the hours your teams spend doing by hand what their peers elsewhere now do in minutes.

Why Buying First Fails, Too

The market’s answer to shadow AI is, predictably, a new product, and some of them are very good. But buying before you’ve measured anything is how security stacks get bloated: overlapping agents on every laptop, competing consoles, alerts nobody has time to tune, and renewals nobody remembers signing.

There’s a more basic problem. You can’t evaluate a tool against a gap you haven’t measured. Without a baseline, every demo looks like the answer, because the vendor gets to define the question.

Meanwhile, most organizations already own more AI-relevant control than they use. If you’re on a Microsoft 365 E5-class license, as many mid-size companies are, you’re already paying for cloud app discovery, conditional access, data classification, data loss prevention down to the endpoint and the browser, and insider-risk analytics. Google Workspace and other platforms have their own equivalents. The capability is often sitting there, licensed and switched off.

A Razor for the Security Stack

Philosophy has a name for the instinct I’m arguing for. Occam’s razor holds that entities should not be multiplied beyond necessity: start with the simplest explanation that fits the evidence. Applied to a security stack, it becomes a working rule. Don’t add a tool until the tools you have can’t do the job, and until you can say exactly which job that is.

That doesn’t mean never buy. It means buy last, and buy narrowly. In practice, the rule becomes four phases and a gate.

PhaseThe question it answersControls you likely already ownWhat “done” looks like
SeeWhat AI is in use, by whom, and with what data?Cloud app discovery, AI activity in data security posture management, a staff surveyAn inventory with owners and risk ratings
SteerWhich tools do we support, and how do people reach them?Single sign-on with conditional access, unsanctioned-app blocking on managed devices, a one-page policyMost AI use flows through a few approved tools
ProtectWhat data must never leave, and what stops it?Sensitivity labels, endpoint and browser DLP, an oversharing cleanupSensitive content is blocked or flagged at the prompt, with a reason
ProveIs it working, and how would we know?Posture reports, audit logs, insider-risk policiesA short scorecard leadership can read
Then buyWhat can’t we do yet?Nothing new until the first four are in placeOne sentence that names a real gap
The framework at a glance. The examples come from Microsoft 365 E5; most enterprise platforms have equivalents.

1. See: Find Out What’s Actually Happening

Start with discovery, not policy. Cloud app discovery shows which AI services are in use, how heavily, and from which devices; Microsoft Defender for Cloud Apps keeps a dedicated Generative AI category with a risk score for each app. Purview’s data security posture management adds a view of AI interactions that touched sensitive data.

Then do the simplest thing of all: ask. A short survey with no penalty attached (what AI tools do you use, for what, and why?) will tell you more about real demand than any log. What people use without permission is a requirements document they’ve written for you, for free.

The output of this phase is an inventory: which tools, which teams, which kinds of data, and how risky each combination is.

2. Steer: Make the Secure Path the Easy One

Choose a small number of sanctioned tools: enterprise versions that commit in writing not to train on your data, support single sign-on, and give you retention and audit controls. Put them behind your identity provider with conditional access, so people reach them with MFA from managed devices.

Then deal with the rest. Tag unapproved AI apps as unsanctioned. On devices onboarded to Defender for Endpoint, with the integration turned on, that tag becomes an actual block, or a warning that explains the policy and points to the approved alternative. Warn-and-educate is underrated: the first time someone hits it, they learn the rule and the right tool in the same moment.

Finally, write the policy people will actually read. One page, plain language: what’s approved, what data never goes into any AI tool, and how to ask for something new.

3. Protect: Guard the Data, Not Just the Doors

Access controls decide who reaches an AI tool. Data controls decide what they can bring to it, and that’s where most of the real risk lives, because the most common way sensitive information reaches a prompt is the humblest one: copy and paste.

Sensitivity labels give your data a classification the rest of the stack can act on. Endpoint DLP can then block or warn when someone pastes or uploads sensitive content to AI websites, and browser-level controls in Edge can inspect what’s typed into a prompt before it’s submitted. Run all of it in audit mode first. A week of data on what would have been blocked will make your policy smarter and your rollout quieter.

One more step belongs here, and it’s the one most often skipped. Before you turn on an AI assistant that can read your internal content, fix oversharing. These assistants respect existing permissions, which sounds reassuring until you remember how many sites, folders, and files are shared with everyone, or with anyone who has the link. AI doesn’t create that exposure. It makes it searchable.

4. Prove: Measure It, Then Report It

Controls you can’t measure are controls you can’t defend at budget time. Posture reports, audit logs, and insider-risk policies give you the raw material. Turn it into a few numbers leadership can follow:

  • The share of AI use happening in sanctioned tools. It should rise.
  • DLP events on AI destinations. They’ll climb as coverage grows, then fall as people learn.
  • Time to answer a request for a new AI tool. It should be days, not months.
  • Standing exceptions. Each one is a decision someone should revisit.

A word of caution on licensing: names and boundaries move fast. Microsoft folded its two posture tools into one this year, and some of the newest agent-governance features require licenses beyond E5. Check what your tenant actually includes before you build a plan around a feature.

Then, and Only Then, Buy

Once you’ve seen, steered, protected, and measured, you’ll know something most buyers don’t: exactly where the native controls stop.

Write that gap down in a single sentence. Maybe your browser coverage ends where unmanaged devices begin. Maybe you need real-time coaching inside the prompt box across dozens of AI apps, not just the big few. Maybe AI features embedded in other SaaS products are invisible to your current controls, or AI agents are starting to act on your data in ways your tools weren’t built to watch.

Those are real gaps, and there are good products built to close them. The difference is that now you can evaluate a tool against a gap you can name, with a baseline to measure it against. You’ll buy something smaller, better targeted, and easier to justify.

If you can’t write the sentence, you don’t need the tool yet.

The Policy Is a Control, Too

It’s tempting to treat all of this as configuration. Mostly, it isn’t. Shadow AI is rarely malicious; it’s usually someone trying to do good work faster with whatever is in front of them. Treat it as a demand signal, not a disciplinary problem.

That means a request process with a fast lane: a short form, a published turnaround, and a review checklist that covers the questions that matter. Does the vendor train on your data? How long do they keep it, and where? Does it support single sign-on? It means training that teaches judgment rather than fear. And it means a help desk that can answer “Can I use this?” the same day.

When the process is faster than the workaround, most people take the process.

The Business Case Writes Itself

Framed this way, the conversation with finance gets easier. The first four phases mostly use licensing you already pay for; the cost is time and attention. The risk reduction is measurable within weeks. And if you do come back for budget, you arrive with a documented gap, a baseline, and a narrow scope, which is exactly the kind of request a CFO can say yes to.

Where to Start on Monday

  • Pull the generative AI category from your cloud app discovery and sort it by number of users.
  • Choose your approved tools and publish a one-page policy.
  • Turn on DLP rules for AI sites in audit mode, then read the results after a week.
  • Run an oversharing review before you enable any assistant that reads internal content.
  • Write the one-sentence gap, or admit that, for now, there isn’t one.

Earning Its Place

I’ve spent two decades watching new technology arrive faster than the policies meant to govern it. Virtualization, the cloud, mobile, SaaS: each began with the same argument between the people who wanted to block it and the people who wanted to let it run. Each time, the organizations that came out ahead did neither. They looked first, decided deliberately, and built controls that let the useful thing in on terms they could explain.

AI is the same test, taken faster. The goal was never to keep it out. It’s to let it in where we can see it, guard what matters, and prove that it’s working, and to hold everything we add to the same standard, from the assistant to the guardrail around it. Each one should earn its place.

More on how I work with leadership teams on AI strategy and cybersecurity.

Hessam Zahedi

WRITTEN BY

Hessam Zahedi

IT leader, infrastructure architect and vCISO. Two decades making technology dependable; now helping leadership teams bring in AI and emerging tech in ways that hold up over time.