The gap between how organizations think their employees are using AI and how employees are actually using it is wider than most IT and security teams would find comfortable if they examined it directly. BlackFog research finds that 86% of employees use AI tools for weekly work tasks, and 58% of them prefer unapproved, publicly available applications over the ones their organizations have sanctioned. The more revealing finding is attitudinal: 65% of employees believe using unvetted AI tools is acceptable, and 60% consider it worth the associated risks when it helps them meet their targets. These figures describe a workforce that has made its own judgment about the risk-reward calculation of shadow AI and largely concluded that the productivity benefit outweighs the risk. That conclusion is understandable given the pressure most employees are under to deliver results, and it is also wrong in ways that the employees making it are not positioned to fully appreciate. The data they are feeding into unsanctioned tools- the trade secrets, client contracts, employee information, financial data- does not stay within the boundaries they assume it does, and the compliance exposure that results is not theirs to bear. Addressing shadow AI effectively requires understanding why it is happening before designing responses to it.
What Shadow AI Is and Why the Risks Are Not Theoretical
Shadow AI refers to the use of AI tools, applications, and services by employees or departments without the knowledge, approval, or oversight of the organization’s IT and security functions. The tools involved are typically consumer-facing AI products- ChatGPT, Claude, Midjourney, and equivalents- that are accessible without organizational credentialing and that operate under data handling terms designed for individual users rather than enterprise environments.
The risk categories that result from this usage pattern are distinct but compounding. Data exposure is the most immediate: when employees enter client contracts, proprietary research, internal financial information, or personal data into public AI models, that information enters an environment with data protection standards that may be substantially weaker than what the organization’s own systems provide, and potentially enters training pipelines that make the information available in ways the employee never anticipated. Security vulnerabilities follow from deploying AI tools that have not been assessed by IT, because those tools may introduce pathways into internal systems that would not survive a formal security review. The compliance dimension is particularly consequential for organizations operating under GDPR, HIPAA, or sector-specific data protection requirements: the use of unsanctioned tools to process regulated data can constitute a violation regardless of whether any breach actually occurs, because the violation is in the processing arrangement rather than the outcome.
The information integrity risk operates differently from the others but carries its own organizational cost. AI tools produce outputs that reflect their training data, their design, and the quality of the inputs they receive, and those outputs can be confidently wrong in ways that are not immediately apparent to users who are not positioned to critically evaluate them. When employees incorporate AI-generated analysis, summaries, or recommendations into business decisions or client-facing work without appropriate scrutiny, the errors those outputs contain become organizational errors, with reputational and operational consequences that the original AI tool bears no responsibility for.
The data on what employees are actually sharing through unsanctioned tools makes the risk concrete rather than theoretical. A third of workers acknowledge sharing data or research through unapproved AI platforms. Thirty-seven percent have disclosed employee data. Twenty-four percent have shared sales or financial information. These are not hypothetical exposures. They are occurring now, across organizations that may have no visibility into them.
Why Prohibition Is Not a Viable Response
The instinct to respond to shadow AI by banning unsanctioned tools entirely is understandable from a risk management perspective, but the behavioral evidence suggests it does not work. Employees who are using unapproved AI tools are doing so because those tools are helping them accomplish work that they are under pressure to deliver. Removing the tools without addressing the underlying productivity need leaves employees facing the same performance expectations with fewer resources, which predictably drives the behavior underground rather than eliminating it.
The 58% of employees who already prefer publicly available tools over sanctioned ones are making that choice in an environment where both options exist. Removing sanctioned alternatives while maintaining performance expectations does not change the underlying calculation that makes shadow AI attractive. It changes the visibility of the behavior, making it harder for IT and security teams to monitor and manage rather than actually reducing the exposure it creates.
The more useful frame is that shadow AI is a signal about unmet needs within the organization’s official AI offering. Employees are reaching for unapproved tools because those tools are doing something that approved tools are not doing, whether that is capability, accessibility, ease of use, or simply being available when an approved alternative is not. Understanding specifically which tools employees are using and what they are using them for is more actionable than knowing that unsanctioned usage is occurring, because it identifies the gap in the official offering that shadow AI is filling.
Building the Framework That Makes Sanctioned AI the Preferred Choice
The organizations managing shadow AI most effectively are those that have made the sanctioned path easier and more capable than the unsanctioned one, combined with policy and technical controls that address the exposure created by unsanctioned usage.
Transparency as an organizational norm is the cultural foundation that makes this approach work. When employees can report the tools they are using without concern about punitive consequences, the organization gains visibility into its actual AI usage landscape rather than its official one. That visibility is the prerequisite for understanding which unsanctioned tools are in use, what they are being used for, and what the gap in the official offering looks like. Framing this as a collaborative effort to build a better AI environment rather than a compliance enforcement exercise produces more honest reporting and more useful information.
A clear usage policy that defines permitted tools, prohibited tools, and the process for requesting evaluation of new tools gives employees a framework for making decisions rather than leaving them to make their own risk assessments. The employees who currently believe that using unvetted AI is acceptable are often not making that judgment in bad faith. They are operating without a clear organizational position on what is and is not permitted, and in that absence they are defaulting to their own judgment about risk. A policy that makes the organizational position explicit, and that includes a realistic process for employees to request tools they need, reduces the ambiguity that drives well-intentioned policy violations.
Technical discovery is the operational component that provides the accurate picture of current usage that self-reporting alone cannot. Endpoint logs, network monitoring, and SaaS discovery tools identify AI tool usage across the organization, with particular attention to departments handling sensitive data where the exposure consequences are most significant. This is not primarily a surveillance exercise. It is the baseline assessment that tells security teams where the real risks are concentrated and what intervention is most needed.
Data loss prevention tools and role-based access controls for more advanced AI capabilities provide the technical backstop that limits exposure even when unsanctioned usage occurs. No policy framework eliminates shadow AI, and technical controls that reduce the damage that unsanctioned tool usage can cause are the preparation that reflects that reality honestly.
The underlying opportunity in shadow AI, and it is a genuine one, is that the widespread adoption of AI tools by employees who are using them without being asked to reflects a workforce that is actively seeking ways to work more effectively. That energy is an asset if it can be channeled through sanctioned tools with appropriate oversight. The organizations that treat shadow AI primarily as a compliance problem to be suppressed tend to drive it further underground. The organizations that treat it as evidence of unmet need and invest in meeting that need through sanctioned alternatives tend to find that the compliance problem becomes more manageable as the productivity argument for using approved tools becomes stronger.
