Yes, the title is provocative. But let me explain. Looking back only several years, the amount of AI systems organizations managed was highly restricted. GenAI was not a thing, neither were agents. With the decisive transformation that tool place since then, AI seems to be everywhere. In all formats. Classical machine learning applications, GenAI applications, AI agents, cameras, robots, and so on. Externally purchased, internally developed, embedded in otherwise approved products, or just created by employees as DIY It. All of this is a net positive. From an AI governance standpoint, the question is whether those systems follow your organization’s AI governance. This is not nice to have, but rather a central pillar of your AI strategy. Yet, most organizations still get it wrong.
When talking about Shadow AI, let’s start with a clear definition. Of course, the term stems from the familiar term „Shadow IT“. There is no legal definition for Shadow AI, so normally the term is used based on the IBM definition: „Shadow AI is the unsanctioned use of any artificial intelligence (AI) tool or application by employees or end users without the formal approval or oversight of the information technology (IT) department.“ (What Is Shadow AI? | IBM). The central elements are as follows: a) The use is unsanctioned, meaning there are no checks in place to prohibit unwanted outcomes, b) Shadow AI systems are actually in use and not deactiveated, to they actually produce outcome, c) The AI Governance organization is not aware of those systems for various reasons (e.g. non-compliant behaviour by employees, lack of AI literacy, or because of unsufficient organizational capabilities; there is a myriad of different reasons). The weakness of IBM’s definitions is limiting the oversight responsibility to the IT department. In reality, different entities are involved in managing the AI portfolio.
How prominent is Shadow AI? According to the Cloud Security Alliance AI Safety Initiative (Shadow AI Apps: The Enterprise Attack Surface That Outpaces Monitoring – Lab Space), 80% of employees use Shadow AI systems. A staggering number. This includes the use of Chatbot solutions online not approved for business use, vendors shipping AI functions for SaaS solutions approved years ago, coding agents running on personal subscriptions, and so on (overview of different scenarios: Seven Shades of Shadow AI: A 2026 Field Guide for Pragmatists – Security Boulevard). What are the effects? According to findings by IBM (Cost of a Data Breach Report 2026 | IBM), Shadow AI incidents have doubled, result in higher costs (5,39$ millions), with regulatory sanctions on the rise. For EU companies, I emphasize it here clearly: Shadow AI is not merely a governance problem. It is a legal risk. Companies need to identify and classify all (!) AI systems in use.
If Shadow AI is such a challenge, then why stop eliminating it? Let me summarize the strategy that makes sense for most organizations. First, you will never fully eliminate Shadow AI. Sure, there are vendors marketing their tools as one-size-fits-all AI scanners. But realistically, there will remain gaps. Encrypted browser sessions, personal devices, capabilities in approved software that remain under the radar. What will happen in reality is that organizations with an overreach strategy will push Shadow AI even more underground while at the same time slow delivery and scaling. The reason is simple: Focussing on Shadow AI as a symptom instead of focussing on the causes will fail. This is the most fundamental success factor: Focus on the causes of Shadow AI use, and not on the symptoms.
Here are 7 core strategies I advice to follow in order to actually succeed in managing Shadow AI.
- Establish a common understanding. Use this excellent taxonomy as a starting point: Seven Shades of Shadow AI: A 2026 Field Guide for Pragmatists – Security Boulevard. Tailor it to your organizational characteristics. Ensure that everybody understands the definition of Shadow AI and its manifestations.
- Provide approved AI solutions for the business needs of your user base. Simply put: When your users need the capability to generate code using AI and there is no approved solution, there should be no surprise that they turn to personal Github Copilot use. Know your user base. Know their needs. Act upon those.
- Be quick when providing AI solutions. Once Shadow AI has been widely spread for a defined capability, users will hold on to it, as it contains their usage history and they are familiar for it. Have processes in place to quickly offer approved AI capabilities for new needs.
- Manage an AI inventory that works. Easily accessible, up-to-date, and implemented in all relevant processes. Incuding agents, data produces, models, and so on. This provides visibility, and visibility is the best medicine agains shadows.
- Include Shadow AI in your policies and processes. Make sure those policies and processes are known, and include trainings to communicate them.
- Establish and maintain a coherent lifecycle management, especially for agents. Decommissioning AI systems (including agents) must be a process and not a nice to have sideline activity. Systems will fall under the radar otherwise.
- Ensure human oversight. When accountability is clear, roles defined, and AI literacy in place, governance responsibilities are shifted to where the actual products are managed, and AI system owners are empowered to ensure that Shadow AI is prevented within their business scope.
- Implement effective technical controls. Specialized detection systems, managed browsers, DLP measures, and so on. However, be realistic on the shortcomings. Technicals controls are effective, but only when based on a functioning organizational governance.
So yes, stop trying to eliminate Shadow AI. Start enabling your organization to address the causes that lead to Shadow AI. Focus on these causes, not on the symptoms. Shadow AI can be successfully managed that way, and when following the principles I outline above, you are on the right track to tackle the issue.
As always, here are my recommendations for further deep dives in addition to the already mentioned sources above:
1. [2606.00088] From Frontier to Shadow AI: A Simmering Threat to Assurance and Security in Critical Infrastructure
2. (PDF) Shadow AI as Governance Failure: Unauthorized Model Use, Organizational Risk, and the Limits of Policy-Based Control
3. (PDF) Shadow AI in the Enterprise: Innovation with Risk and Compliance