More than half of enterprises are now actively deploying artificial intelligence, and barely a quarter believe their governance has kept pace with that deployment. That is not a prediction about where shadow AI risk is headed. It is a measurement of where enterprise AI governance stands today, published by researchers who surveyed the organizations actually living through the gap. The number that should concern every compliance leader is not the adoption figure. It is the distance between it and the governance figure sitting next to it.
Shadow AI, the hidden threat of unsanctioned or unmonitored AI tool use against enterprise data, has been discussed as an emerging risk for several years. What has changed is the scale at which it is now a documented, measured condition rather than an anecdote security teams swap at conferences. Fresh research from Smarsh and FTI Consulting puts hard numbers behind what CISOs and compliance officers have been describing qualitatively. Adoption is real, broad, and already in production, while the governance structures meant to contain it remain, for most organizations, aspirational.
The reason this gap persists is not that organizations lack awareness. It is that AI deployment happens at the speed of individual teams adopting individual tools, while governance happens at the speed of policy committees, legal review, and procurement cycles. Those two speeds were never going to match without a deliberate architectural choice to enforce policy at the point where data actually moves, rather than trying to govern every tool and every team through a review process that cannot keep up with how fast new AI surfaces appear.
The Smarsh and FTI Consulting study surveyed enterprises on how AI deployment and AI governance are actually tracking against each other, rather than asking separately about intent or aspiration. The headline numbers are stark set against each other, with 55 percent of organizations actively deploying AI, and only 26 percent reporting that their governance frameworks are fully aligned with that deployment pace. A further 30 percent reported comprehensive capability to detect and manage shadow AI specifically, meaning seven in ten organizations cannot reliably see the unsanctioned AI activity happening across their own workforce.
James Condon, who leads insights for FTI Consulting's strategic communications practice, framed the finding directly. AI readiness is not primarily a technology challenge, it is a data governance problem. That framing matters because it reorients the conversation away from model selection, vendor evaluation, and capability benchmarking, the questions that dominate most AI strategy conversations, and toward the much harder question of whether an organization actually controls what happens to its data once an employee opens an AI tool, sanctioned or not.
Smarsh's own leadership echoed that shift in a different register. Communications data, historically treated mainly as something to preserve in case a regulator asked for it later, is becoming a foundation that responsible AI programs, internal investigations, and business intelligence functions all depend on simultaneously. That dual role, compliance archive and active governance substrate, means the quality and completeness of an organization's content governance now directly determines how defensible its AI governance claims actually are.
The gap between 55 percent deployment and 26 percent governance alignment is not best explained as organizational negligence. It reflects two processes moving at fundamentally different speeds. An individual team adopting a new AI tool, whether a sanctioned enterprise license or a free consumer product opened in a browser tab, can start using it the same day. A governance framework capable of covering that tool, meanwhile, typically requires legal review, a risk assessment, a policy update, a training rollout, and often procurement sign off, a cycle that reliably takes months even in organizations that prioritize it.
Kiteworks' own research into this dynamic found a similarly stark asymmetry. In Kiteworks' AI policy gap analysis, roughly 90 percent of organizations reported active AI use while only about a quarter had documented acceptable use policies covering it. That figure, drawn independently from the Smarsh and FTI numbers above, points at the same structural mismatch from a different survey population, which is itself worth noting. When multiple, independently conducted studies converge on the same order of magnitude gap between adoption and governance, the finding stops looking like measurement noise and starts looking like a structural feature of how enterprise AI adoption actually happens.
This is also why blocking tools outright has proven to be an ineffective governance strategy on its own. An organization that blocks a specific AI application at the network layer does not eliminate the underlying demand that drove an employee to use it. It only removes one visible channel, often pushing the same behavior toward a personal device, a browser extension, or a tool that has not yet made anyone's blocklist. As Bonfy has documented, blocking AI apps tends to make shadow use harder to see, which is in many respects a worse outcome than the unmanaged visibility organizations started with.
It is worth separating two distinct failures that the research captures in adjacent statistics. A policy gap means an organization has not yet written the rules. A detection gap means an organization could not enforce those rules even if it had written them, because it cannot see where AI tools are actually touching its data. The 30 percent figure for comprehensive shadow AI detection capability describes the second, more fundamental failure, and it is the harder one to close through policy alone.
Detection failure compounds over time in a way that policy failure does not. An undocumented policy can be written this quarter. Blind spots in an organization's visibility into its own data flows tend to widen as AI adoption spreads, because every new tool, every new browser extension, and every new personal device used for work adds another path that existing monitoring was never built to cover. Shadow AI security built around content-layer visibility, rather than trying to enumerate every application an employee might open, is one of the few approaches that scales with this problem rather than falling further behind it with each new tool that reaches the market.
The healthcare, financial services, and legal verticals carry a sharper version of this exposure, because a detection gap there is not only an operational risk, it is a direct compliance failure the moment regulated data crosses into an unmonitored AI tool. An organization that cannot demonstrate what data moved through which AI tool during a given period cannot produce the evidence a regulator, a plaintiff's counsel, or an internal auditor will eventually ask for, regardless of how well intentioned its written policy was.
Naming an AI owner is not the same as governing AI data, and the ownership question for AI governance remains genuinely unsettled across the industry, and the research cited throughout this post does not resolve it cleanly. What it does make clear is that the problem sits squarely at the intersection of two functions that have not always coordinated closely, the Chief Compliance Officer, whose job is producing defensible evidence that data was handled according to policy, and the CISO, whose job is making sure the technical controls exist to enforce that policy in the first place.
Kiteworks' broader research tracking this question, an annual governance gap survey, found that the gap had widened rather than narrowed year over year, with audit logging and kill switch capability frequently existing on paper without ever being tested against a live workflow. That finding pairs directly with the Smarsh and FTI numbers. A governance framework that exists only as a document, untested against the actual tools employees are using, is not meaningfully different from having no framework at all when a regulator asks for evidence rather than policy.
The shadow AI identity gap adds a further wrinkle specific to the CISO's side of this joint ownership. Many shadow AI tools get used through a human employee's own credentials rather than through any distinct AI-specific identity, which means the access an unsanctioned tool exercises is functionally indistinguishable, from an identity and access management standpoint, from the employee's own normal activity. That makes shadow AI use nearly invisible to controls built around monitoring privileged or unusual account behavior, because nothing about the account behavior looks unusual.
The instinct to close a governance gap by restricting AI adoption is understandable and, based on the evidence available, largely counterproductive. Organizations that have tried to govern shadow AI primarily through restriction have tended to push the behavior further from visibility rather than eliminating it. The more durable approach treats governance as something that has to operate at the same speed as adoption, which in practice means enforcement embedded in how data moves rather than enforcement that depends on reviewing each new tool before an employee can use it.
Contextual data enforcement is built around that premise directly, policy that evaluates content and context at the moment data would move into or through an AI tool, rather than policy that depends on first identifying and approving every tool capable of touching that data. This closes the structural mismatch described earlier in this post, because enforcement no longer waits on the months long cycle of legal review, policy drafting, and training rollout that a new tool would otherwise require before anyone could govern it.
Kiteworks has made a parallel argument specifically about avoiding the trade-off between governance and velocity, framed around shadow AI governance. The organizations that come closest to closing the gap the Smarsh and FTI research identifies are not the ones running the most restrictive AI policies. They are the ones whose data governance infrastructure makes the restriction largely unnecessary, because sensitive content is protected at the point of use regardless of which specific tool an employee happens to be running that week.
Return to the framing FTI Consulting offered. AI readiness is a data governance problem before it is a technology problem. The practical test of data governance, for a CISO or compliance officer who will eventually sit across from a regulator, an auditor, or opposing counsel, is not whether a policy document exists. It is whether the organization can produce, on request, a defensible account of what data moved through which AI tools, sanctioned and unsanctioned, over a specified period.
Communications and content data sit at the center of that evidentiary burden, which is precisely why Smarsh's own framing treats that data as a foundation rather than a retention afterthought. An organization with only 30 percent confidence in its shadow AI detection capability, the figure this research found to be the norm rather than the exception, is an organization that cannot currently make that accounting with confidence. Closing that specific gap, moving from policy that exists on paper to enforcement that operates continuously at the data layer, is the single highest leverage step available to the 74 percent of enterprises this research found still running governance behind their own AI deployment.
Research from Smarsh and FTI Consulting found 55 percent of enterprises actively deploying AI in production, while only 26 percent reported governance frameworks fully aligned with that deployment pace. That leaves a majority of organizations with AI in active use ahead of the controls meant to govern it, a gap that held consistently across the surveyed population.
Shadow AI refers to AI tools used inside an organization without formal sanction, review, or monitoring, ranging from free consumer AI products opened in a browser to AI features embedded inside other software that employees enable without IT's knowledge. Sanctioned AI use runs through tools the organization has formally approved, licensed, and, ideally, governed with policy and monitoring. The risk with shadow AI is not that the underlying technology differs, but that the organization has no visibility into what data reaches it.
Evidence suggests blocking alone tends not to solve the problem and can make it harder to detect. When an organization blocks a specific AI application, employees who still want that capability often move to an unblocked alternative, a personal device, or a browser extension that was not anticipated, which removes one visible channel without removing the underlying demand or improving overall visibility into where sensitive data is actually going.
Because the risk that shadow AI and ungoverned AI deployment create is fundamentally about what happens to an organization's data, not about any particular model's capabilities or a specific vendor's practices. An organization that can enforce policy on its data wherever that data moves, including into AI tools it has not individually reviewed, closes most of the practical risk regardless of which AI tools employees ultimately use.
The research and the broader body of evidence it sits alongside point toward enforcement embedded at the data layer as the most scalable answer, because it does not require reviewing and approving every individual AI tool before governance can apply to it. Combined with honest measurement of current shadow AI detection capability, since 70 percent of organizations in this research could not claim comprehensive visibility, this lets an organization govern new AI tools as they appear rather than perpetually catching up to tools that are already in use.
To learn more about shadow AI governance, schedule a custom demo.