Shadow AI is what happens when the approved way to use AI is slower or harder than the unapproved way. You stop it by making IT the department of how, not the department of no: give your most productive people a fast, approved outlet, and they stop hiding. Block them, and they'll keep using AI on devices and accounts you can't see.
This post is for IT directors, and for the executives they report to. I'm not arguing that security doesn't matter. I'm arguing that in 2026, saying no to AI doesn't produce security. It produces shadow IT.
What is shadow AI?
Shadow AI is any use of AI tools your organization hasn't approved or doesn't know about. Usually that means personal ChatGPT, Claude or Gemini accounts, browser extensions, or AI features quietly switched on inside other apps.
It's widespread. A 2026 PagerDuty survey of 1,250 office professionals, conducted by Wakefield Research, found that two-thirds (66%) had used AI tools at work despite believing they were not permitted. In the same survey, 34% said they had entered customer data or information into public AI tools. The survey covered companies with at least $500 million in revenue, but nothing about the pattern depends on size. A 40-person firm in Cedar Rapids has the same incentives, with fewer people watching.
Why can't IT just block AI tools?
Because the people you're blocking are your highest AI adopters, and they've found real productivity. They're not going to give it up.
Think about who goes looking for a workaround. It isn't the person who tried ChatGPT once and lost interest. It's the analyst who cut a weekly report from three hours to thirty minutes, or the project manager who now drafts status updates in five. When you take that away, they don't go back to the old way. They move the work somewhere you can't see it.
That means company data and intellectual property leave approved devices and platforms. A file gets emailed to a personal account so it can go into a personal AI tool. A client document gets pasted into a phone app. Every block creates a new workaround, and IT ends up in an expensive game of whack-a-mole, chasing tools it can't see with policies nobody follows.
The cost is real. IBM's 2025 Cost of a Data Breach research found that a high level of shadow AI added an extra USD 670,000 to the global average breach cost.
How should IT handle shadow AI?
The most effective way to reduce shadow AI is to give people an approved outlet that's at least as fast as the workaround. Banning AI tools pushes your most productive employees, the ones who've found real gains, to move company data onto personal accounts and devices to keep those gains, and every ban creates a new hiding place. Instead, IT should become the department of how. That means one simple request form for new tools, automations or access, with every request answered by a set date with a yes, a no, or a not-yet and the reason. It means starting with a no-penalty list of the tools people already use, offering a one-time amnesty so people disclose what they've built, and configuring approved tools safely, because safety depends on account type and settings, not on the brand. When someone still steps outside the lines, treat it as a sign the safe path was too hard, and fix the path.
What does the "Department of How" look like in practice?
| Department of No | Department of How | |
|---|---|---|
| Approval time | Weeks or months, or never | A set, published date for every answer |
| What users do | Use personal accounts and hide it | Ask first, because asking works |
| Where data goes | Personal accounts, phones, home computers | Approved, configured business accounts |
| What IT can see | Almost nothing | What's in use, by whom, with which data |
| Cost | Ongoing whack-a-mole, plus the breach risk | A small, predictable intake and review effort |
Fix the latency first
Latency is the whole thing. A six-week approval guarantees shadow IT. Nobody with a deadline waits six weeks for permission to save an hour a day.
Build one front door
One form to IT for a new tool, an automation or access: what, who, which data, and why. Every request gets an answer (yes, no, or not yet, with the reason) by a set date. A clear no with a reason is far better than silence, because silence reads as "figure it out yourself."
Start from what's already in use
Don't begin with a blank approved list. Begin with a no-penalty list of the tools people already use, then sort them into approve, configure or replace. You'll learn more in a week of honest answers than in a year of network monitoring.
Offer a one-time amnesty
Tell people: show us what you've built and what you use, and nobody gets written up for it. Then set the rules. You can't secure what you don't know about, and people won't tell you while they're afraid.
Treat boundary breaks as a process fix
When someone puts data where it shouldn't go, find out why the safe path was harder. Maybe the approved tool lacked a feature, or the request took too long, or nobody knew the rule. Fix that, and the next person won't need the workaround. Punishment teaches people to hide better.
What makes an AI tool safe for company data?
Safe is a configuration, not a property. The same AI model can be fine for client work in one setup and a real risk in another.
The one thing to remember is the difference between consumer and business accounts. Business plans from the major AI vendors generally don't use your data to train their models by default. Consumer accounts often do unless someone changes a setting. That's why the approved list should name account types, not just brands.
It also helps to know that processing is not training. A tool that reads your document to answer a question isn't necessarily learning from it. What matters is the vendor's terms on retention and training, and the settings you choose.
And when AI tools connect to email, files or other systems, permissions matter most. The damage a manipulated AI assistant can do is limited by the access you gave it. Grant the least it needs. We go deeper on these questions for law firms in Is Claude safe for legal work?
Where does shadow AI fit in your AI policy?
Your AI policy should own the guardrails: which data goes where, which accounts are approved, and how fast requests get answered. But a policy that only says no is how shadow AI starts. We cover where the rules should come from the top, and where change has to come from your people, in AI policy vs. AI strategy.
IT's job isn't to stop people from finding better ways to work. It's to make sure they can do it safely, in the open. If you want help designing that front door for a business in Iowa, see our AI strategy and governance work on the Advise, Educate & Lead page.