Why the People Who Write AI Policy Are the Ones Breaking It

Most shadow AI conversations assume the risk sits with frontline employees quietly pasting customer data into a free chatbot. A 2026 survey of enterprise security leaders suggests the assumption has the hierarchy backward. The executives who set AI policy are nearly twice as likely as the frontline staff who answer to them to prioritize speed over security when using AI tools themselves.

That single finding reframes the entire shadow AI problem for anyone building a governance program today. If governance is failing because frontline employees do not understand the policy, training fixes it. If governance is failing because the people who wrote the policy do not follow it, no amount of employee training will close the gap, because the behavior being modeled from the top contradicts the rule being enforced from below. Enterprises spending heavily on awareness campaigns aimed at individual contributors while leaving executive AI use unmonitored are solving the wrong half of the problem, and doing so with a budget and a sense of confidence that the underlying numbers do not actually support.

This post looks at what the research actually found, why the gap is structural rather than a matter of a few bad actors, and what a governance model has to do differently once it accepts that enforcement cannot depend on goodwill at any level of the organization, including the top. It also looks at why that structural reality hits regulated industries harder than most, since a compliance officer cannot write off a senior leader's shadow AI use as a one-off lapse when the underlying statute makes no distinction based on seniority.

Key Takeaways

  1. Leaders bypass AI policy more than staff, not less. Roughly 69 percent of C-suite executives prioritize speed over security when using AI tools, compared with 37 percent of frontline employees.
  2. Most AI activity still runs outside sanctioned accounts. A majority of enterprise AI usage happens through personal accounts on company-licensed platforms, invisible to IT by design.
  3. Visibility is the deeper problem, not awareness. The large majority of organizations report no meaningful visibility into how data moves to and from the AI tools employees actually use.
  4. Restriction without alternatives drives the behavior underground. A large share of employees find workarounds when AI tools are restricted, and many say they would keep using AI even if it were banned outright.
  5. The generational assumption does not hold up either. A majority of younger employees admit to hiding their AI use at work, which undercuts the idea that shadow AI is mainly an older-workforce training problem.

What the Research Actually Found

The clearest data point comes from the Shadow AI Behavior Report, published in 2026 by the workforce analytics vendor Teramind (read the report), based on a survey of 300 enterprise security leaders alongside a separate survey of more than 600 senior HR professionals. The headline finding is stark. Sixty nine percent of C-suite executives say they prioritize speed over security when using AI agents and other AI tools, compared with 37 percent of frontline employees asked the same question. Leaders are not accidentally less careful. They are nearly twice as likely to choose speed deliberately, in roles where their own example is supposed to set the standard for everyone reporting to them.

The same report found that 67 percent of enterprise AI activity runs through personal accounts on platforms the company itself has licensed, meaning the organization paid for a sanctioned tool and most of its actual usage happens outside the account structure that would let IT see it. An additional 86 percent of organizations surveyed reported lacking visibility into how data moves to and from the AI tools their employees use day to day. That combination, licensed tools used through unlicensed accounts, with no visibility either way, means most enterprise AI governance programs are built to monitor a usage pattern that represents a minority of what is actually happening, a dynamic documented separately in research on a 509 percent surge in shadow AI activity that traditional data loss prevention tools were never built to catch.

Why Restriction Alone Makes the Problem Worse

A natural response to these findings is to tighten access, blocking more domains, restricting more tools, and enforcing stricter use policies. The same research suggests that response backfires more often than it helps. Forty five percent of employees surveyed said they find workarounds when an AI tool is restricted, and 48 percent said they would keep using AI even if their organization banned it outright. Those are not fringe numbers from a disengaged minority. They describe close to half the workforce treating a restriction as a problem to route around rather than a rule to follow, which is exactly the dynamic that blocking AI apps without addressing the underlying workflow need tends to produce in practice.

This is where the generational assumption breaks down as well. It would be convenient if shadow AI were mainly a function of younger employees moving fast without thinking about risk, since that framing suggests a training fix aimed at a specific demographic. The data does not support that story cleanly. Sixty two percent of Gen Z employees surveyed admitted to hiding their AI use at work, which is a meaningfully high number, but it describes concealment, not recklessness, and the research on shadow AI training gaps suggests the executive figures above show the same instinct to prioritize speed showing up most strongly at the top of the organization, not the bottom. browser-based shadow AI is the venue precisely because it is easiest to hide there, regardless of who is doing the hiding.

The Cost of Getting This Wrong Is Not Hypothetical

Separate research on enterprise data breaches gives this governance gap a dollar figure rather than leaving it as an abstract risk. IBM's 2025 research into AI-related breaches found that 97 percent of organizations that experienced a breach of an AI model or application lacked proper AI access controls (see the IBM findings), and shadow AI activity specifically compounds that exposure because it creates data flows that incident response teams were never built to trace. When a breach originates from a personal account on a licensed platform, the forensic trail does not run through any system the organization logs by default, which extends both detection time and the eventual cost of containment and disclosure.

That matters directly for the compliance and audit function that eventually has to explain what happened to a regulator or a board. A written AI use policy that the organization cannot demonstrate was actually followed, by anyone, at any level, is not a defense in that conversation. It is evidence that the organization knew the risk and did not close the gap between the rule and the behavior.

Governance That Depends on Compliance Will Not Survive Contact With Executives

The uncomfortable implication of the Teramind findings is that a governance model built on voluntary compliance cannot work if the people least likely to comply voluntarily are also the people setting the policy. An acceptable use policy is a statement of intent, not a control. Naming an AI owner on an org chart does not change what happens the moment a senior executive opens an unsanctioned AI tool to move faster on a deadline, because nobody in that moment is positioned, or often willing, to intervene.

What has to change is where enforcement happens. A policy that depends on every individual, from the newest analyst to the chief executive, choosing compliance over convenience in the moment will fail at a predictable, measurable rate, and the research above suggests that rate is higher at the top than anywhere else in the organization, which is exactly why guidance on how to govern shadow AI now treats enforcement, not awareness, as the starting point. A control that evaluates and governs the content itself, at the point an AI tool actually touches it, regardless of whose account is doing the touching, does not depend on anyone's willingness to follow a rule. Entity-aware security that enforces based on what the data is and who or what is requesting it works the same way whether the requester is a new hire or the person who approved the policy in the first place.

What a Content-Layer Control Actually Catches

Consider what each of the Teramind findings looks like from the perspective of a control built at the content layer rather than the account layer. A control that only governs sanctioned accounts never sees the 67 percent of AI activity happening through personal logins on licensed platforms, because the account itself looks legitimate even when the authentication behind it is not. A control built around the content leaving the organization, regardless of which account or browser tab it travels through, catches that activity because it does not depend on the account being the sanctioned one.

The same logic applies to the executive behavior at the center of this research. Blocking a specific AI domain for everyone except a carved-out executive exception list, a common real-world compromise, recreates exactly the governance gap the research describes, since the people most likely to bypass policy are explicitly granted the exception that removes them from monitoring. Legacy DLP built around blocking specific tools or domains cannot adapt fast enough to cover the next unsanctioned tool an executive or employee adopts next quarter, which is precisely why the control needs to live at the level of the content and the data relationship, not the application being used to move it.

Why Visibility Has to Come Before Restriction

The 86 percent of organizations reporting no real visibility into their own AI data flows is, in some ways, the most actionable number in the entire report, because visibility is the one prerequisite that has to exist before any other governance decision can be made correctly. An organization cannot scope a policy exception sensibly, cannot calculate real risk exposure, and cannot explain to a regulator what actually happened during an incident if it does not know, in the first place, where its own sensitive content is flowing, a problem closely related to the broader shadow data gap most enterprises carry even outside their AI tools specifically. A related pattern shows up in shadow AI identity research, which found that unmanaged AI access often rides on the same privileged accounts nobody is reviewing regularly.

Data Surface Visibility has to precede any restriction decision, not follow it, because a restriction built on incomplete visibility just pushes the uncovered activity further out of sight, exactly as the workaround statistics above describe. Getting visibility right first also changes the politics of the rollout: a security team that can show a specific executive, accurately, what sensitive content moved through an unsanctioned tool last month has a very different conversation than a security team asking that same executive to simply trust a new restriction.

That conversation matters more than it might seem, because the Teramind findings imply that executive buy-in cannot be assumed and has to be earned with evidence specific enough to be undeniable. A generic warning that shadow AI is a growing risk competes poorly against the convenience an executive feels from using whatever tool gets a deadline met fastest. A specific finding, naming the actual document, the actual tool, and the actual date, is much harder to wave away, and it reframes the conversation from a policy debate into a concrete, individual risk the executive now has to own. Security teams that have made this shift report that the second conversation, after the first piece of specific evidence, goes considerably better than the first one ever did.

Why Regulated Industries Cannot Treat This as a Culture Problem

For most consumer-facing businesses, a shadow AI incident is primarily a competitive and reputational risk. For regulated industries, specifically healthcare, financial services, and the defense industrial base, the same incident is a compliance failure with statutory consequences attached, and the governance gap documented above lands differently in each.

A healthcare compliance officer reading that the C-suite is the group most likely to bypass AI policy has to reckon with the fact that HIPAA's unique-user-identification and audit-control requirements do not carve out an exception for executives. A senior clinician or administrator who pastes patient information into a personal account on an unsanctioned AI tool has created exactly the kind of untracked, unauthorized access the Security Rule was written to prevent, and the organization's inability to produce a record of that access is itself the violation, independent of whether any harm resulted. The same logic applies under GLBA and SEC recordkeeping obligations in financial services, where an executive's shadow AI use is not a training lapse but a books-and-records gap that examiners are specifically trained to look for.

Defense contractors face an even sharper version of the problem. Controlled unclassified information routed through a personal account on a consumer AI tool falls squarely inside the assessment boundary that CMMC compliance now requires organizations to defend, and the research above suggests the people most likely to make that mistake under deadline pressure are the people least likely to be questioned about it in the moment. A C3PAO assessor evaluating access controls will not accept "the policy existed" as evidence that controlled information stayed inside the authorized boundary. The assessor wants the access record, and the record either exists at the point the data was touched or it does not exist at all.

What unites all three verticals is that none of them can treat this as a matter of better leadership example-setting. A compliance program has to assume that example-setting will fail at a measurable, predictable rate, exactly as this research describes, and build the control at the layer where it cannot be bypassed by seniority, urgency, or good intentions.

This is also why the usual remediation sequence, write a policy, train employees on it, audit compliance annually, tends to understate the actual exposure in a regulated environment. An annual audit samples a narrow window of activity months after the fact, while the behaviors this research documents happen continuously, across every account, sanctioned and unsanctioned, every day in between. A compliance function that only looks once a year at a problem that is happening daily will keep discovering, each time it looks, that the gap it is measuring has already grown since the last review. Continuous, content-level visibility closes that lag, because it does not depend on a scheduled audit catching a pattern that a daily control would have caught the first time it occurred.

Frequently Asked Questions

1. Is shadow AI mainly a problem caused by junior or less experienced employees?

No. The research central to this post found the opposite pattern: 69 percent of C-suite executives prioritize speed over security when using AI tools, compared with 37 percent of frontline employees, meaning senior leaders are nearly twice as likely to bypass the caution their own policies require. Shadow AI is better understood as an organization-wide behavior pattern than as a junior-employee training gap, and any governance program built on the opposite assumption is likely targeting the wrong population first.

2. If most AI activity runs through personal accounts, how would a company even see it?

Account-level monitoring will not catch it, since a personal account on a licensed platform looks legitimate from an identity standpoint even when it falls outside the organization's managed structure. Visibility has to come from monitoring the content and data itself as it moves toward and through AI tools, independent of which account or browser session is carrying it, which is a fundamentally different approach from traditional access-log monitoring.

3. Does restricting or blocking AI tools actually reduce shadow AI use?

The research suggests restriction alone tends to push the behavior further out of sight rather than eliminating it. Forty five percent of surveyed employees said they find workarounds when a tool is restricted, and 48 percent said they would keep using AI even under an outright ban. Restriction paired with no visibility into what employees actually need the tools for tends to produce exactly the concealment the data describes, and it does little to change the executive behavior that the research identifies as the larger driver of the overall gap.

4. Is this just an awareness and training problem that better onboarding would fix?

Training addresses a knowledge gap, but the data describes a behavior gap at the leadership level specifically, and leaders already generally understand the policies they themselves wrote. A training program aimed at frontline staff does not change the incentive an executive feels to move faster than the policy allows, so training alone will not close the part of the gap the research found to be largest.

5. What does an effective response to this research actually look like in practice?

It starts with visibility into where sensitive content is actually flowing through AI tools, sanctioned and unsanctioned alike, because the vast majority of organizations currently lack that visibility entirely. From there, enforcement needs to happen at the level of the content itself rather than depending on account-level restrictions or policy compliance, since the research shows the people most likely to need an exception from account-level rules are also the people least likely to follow them voluntarily. In regulated industries specifically, that means building an access record that holds up under examination regardless of whose account generated the activity, rather than a policy document that only describes intent.

To learn more about shadow AI governance, schedule a custom demo.