Why visibility is becoming the foundation of AI governance

Employees aren't waiting for IT to approve an AI tool before they start using one. They're reaching for whatever's fastest and already open in a browser tab, often through a personal account where IT has no visibility. That means sensitive company or customer data flows through a tool nobody vetted to handle it.

The real problem in 2026 isn't that AI adoption is growing. Nobody should find that surprising. But it should be acknowledged that most of that growth happens where IT can't see it, let alone govern it.

When shadow AI becomes a reportable incident

On May 5, 2026, an employee at a small Pennsylvania bank ran nonpublic customer information (names, Social Security numbers, and dates of birth) through an AI tool that no one had ever cleared for that use. No hacker broke in. No outsider stole anything.

Within 48 hours, the bank's leadership determined the incident was material, and on May 11, CB Financial Services filed a Form 8-K with the United Securities and Exchange Commission (SEC). It was the first known public filing to name unauthorized AI use as the cause of a reportable cybersecurity incident.

The bank never named the AI application involved, and it hasn't disclosed how many customers the incident affected. What is public is how it happened: an employee doing ordinary work ran regulated customer data through a tool outside the bank's approved AI stack, causing a threat. No malice. Just a shortcut with real consequences.

What is shadow AI?

Shadow AI is the term security teams use for AI tools that IT never approved, never vetted, and in most cases doesn’t even know about. It can be as simple as an employee opening whatever tool is fastest in a new tab because it works better than what the company officially sanctioned.

The numbers show just how wide that gap is. The Omnissa State of Digital Workspace 2026 report found AI assistant usage inside the enterprise grew nearly 1,000% year over year, the fastest-growing app category the report tracked. Almost none of that growth happens through sanctioned tools. On iOS, 96% of AI usage runs through unsanctioned apps, mostly ChatGPT rather than the officially provisioned Microsoft Copilot. On Android, the unsanctioned share climbs even higher, to 99%, led by Google Gemini.

Other research confirms the trend. Verizon's 2026 Data Breach Investigations Report, built from more than 22,000 confirmed breaches, found that regular AI use on corporate devices tripled in a single year, from 15% to 45% of employees. Shadow AI is now the third most common non-malicious insider action, a fourfold increase from the year before. Additionally, two-thirds of employees accessing AI on a company device are doing it through a personal account.

The behavior isn't new, either. The Omnissa State of Digital Workspace report also tracks activity on communication apps, and the same pattern shows up there, which predates AI entirely, one that predates AI entirely. Roughly 45% of communication apps running on managed desktops are consumer tools rather than sanctioned enterprise software.

Employees have routed around approved tools for years, whenever the approved tool was slower or clunkier than what they already trusted on their own time. Shadow AI is just the fastest-moving version of that same habit.

The employees driving those numbers aren't trying to cause harm. By every account, they're trying to work faster. What changes by industry is what's at stake when AI use goes wrong.

Shadow AI in regulated industries

In regulated industries, a productivity shortcut can quickly become a compliance event. Here's how that risk plays out across three sectors.

Financial services

In most companies, shadow IT tends to concentrate in marketing or HR, departments with lighter oversight and more room to experiment. Banking breaks that pattern. Security researchers found ungoverned AI tools outnumbering monitored one roughly four to one in banking, concentrated in treasury operations, payments processing, and correspondent banking. Those departments carry the heaviest document workloads and the fewest approved tools to keep up.

What's flowing through those unapproved tools isn't incidental either. It's transaction records, counterparty communications, customer files, and sanctions-screening outputs: the kind of information regulators expect a bank to keep tightly controlled. CB Financials' SEC filing shows what happens when someone discovers that gap instead of managing it.

Healthcare

Healthcare carries a version of the same exposure, tied to a different rule. Any AI tool that touches protected health information (PHI) needs a signed Business Associate Agreement (BAA) under the Health Insurance Portability and Accountability Act (HIPAA) before anyone can legally use it with patient data. Standard consumer ChatGPT doesn't offer one. That means a nurse who pastes a patient's chart into a consumer AI tool to draft a summary may have triggered a HIPAA violation, regardless of how careful the prompt was.

This isn't hypothetical. Paubox's 2026 Shadow AI report found that 95% of healthcare organizations suspect their staff are already using generative AI for work-related tasks, and 62% of IT leaders say they've directly observed employees experimenting with unsanctioned tools like ChatGPT. Only 42% of those organizations have a BAA in place covering any AI tool at all.

Government

In January 2026, the acting director of the Cybersecurity and Infrastructure Security Agency (CISA), the federal government's own cybersecurity watchdog, uploaded sensitive contracting documents to public ChatGPT. The files weren't classified, but they were marked For Official Use Only, so they were never meant to leave government systems.

The upload tripped automated federal security alerts and triggered an internal review, all under an exception to a policy that was supposed to block the tool. If the person responsible for protecting the nation's critical infrastructure can still trigger a shadow AI incident through an approved exception, the theory that better rules alone will fix this doesn't hold up.

By May, the Department of Homeland Security (DHS) stopped experimenting with exceptions and shut the door entirely. The department told staff to stop using commercial generative AI tools, including ChatGPT and Claude, reversing the policy that had conditionally allowed them. The agency is steering employees toward internally built tools instead, including a DHS-specific assistant already available to about 20,000 headquarters staff. Will the ban actually reduce outside AI use? Evidence from other organizations isn't encouraging. 

Why bans fall short

AI moves faster than policy. Even the most regulated organizations struggle to keep up, and those growing pains show up as policy violations and exposed data. Bans don't close the gap—they just push usage somewhere harder to see.

Consider what happened when a financial services firm audited browser activity after banning ChatGPT outright. It found more than a third of employees still reaching the tool through personal accounts and mobile data.

What a sanctioned alternative looks like

The more durable fix looks less like a policy and more like infrastructure. For public sector agencies, virtual desktop environments can let staff use AI tools without the data ever leaving agency-controlled boundaries. That approach preserves the convenience employees already want while keeping the data inside the perimeter regulators expect.

Omnissa Horizon delivers this kind of secure virtual desktop environment, and Omnissa holds FedRAMP High authorization, the standard for the government's most sensitive unclassified workloads.

Healthcare offers a working example of what this looks like when organizations deliberately instead of reactively. Ambient documentation tools that listen to a clinical visit and draft the note automatically, with a signed BAA already in place, are now in use in roughly 62% of hospitals running Epic, meaningfully cutting documentation time.

Health systems didn't get there by banning consumer AI and hoping clinicians would comply. They got there by approving a tool fast enough that it beat the appeal of anything unsanctioned. Ambient tools are built to sit in the note, not the diagnosis, and that's exactly where clinicians want AI to stay.

Seeing it before it becomes a problem

Every story here, from the bank to the nurse to the nation's own cybersecurity agency, shares the same breaking point. It isn't the absence of a rule. It's the absence of visibility into whether people follow the rule, and the absence of a sanctioned option good enough to meet employees' needs.

Pair clear visibility into which tools employees open with a sanctioned option fast enough to actually use, and there's nothing left to guess. The organizations getting ahead of this aren't writing tougher rules. They're building visibility first.

How Omnissa helps you see and govern AI use

Visibility and control work best on one platform. The Omnissa platform helps IT teams:

• See what's really running. Omnissa Intelligence surfaces app usage data across managed devices, so you know which AI tools employees actually use, not just which ones you approved.
• Deliver sanctioned AI fast. Workspace ONE UEM lets you deploy approved AI apps across Windows, macOS, iOS, and Android, and enforce policies that keep sensitive data in managed apps.
• Keep regulated data contained. Omnissa Horizon gives employees secure virtual desktops where they can work with AI tools while data stays inside controlled boundaries.

Start with visibility

Shadow AI won't wait for your next policy review. Every day without visibility is another day of unmonitored data flowing through unvetted tools. The fix isn't a stricter ban. It's knowing what's happening on your devices and giving employees an approved option they'll actually choose.

Ready to bring shadow AI into the light? Download the Omnissa State of Digital Workspace 2026 report for the full data, or talk to your Omnissa or partner sales team to see how the Omnissa platform delivers the visibility and control AI governance demands.