Shadow IT
Tools, services and now AI agents used in an organization without IT approval or oversight. One in five organizations has suffered a breach linked to Shadow AI.

Shadow IT refers to all the digital tools, services and applications used within an organization without approval or oversight from the IT department. Its most recent and most worrying form is Shadow AI: the use of unmanaged artificial intelligence tools and agents, which combines sensitive data exfiltration with autonomous, unrecorded action. In 2025, one in five organizations suffered a data breach linked to Shadow AI, with an average additional cost of $670,000 per incident.
From Shadow IT to Shadow AI: a change in nature
Classic Shadow IT is a data problem: an employee stores files on a personal cloud, a team signs up for a SaaS tool without going through IT. The risk is exfiltration and loss of control. Shadow AI adds a new dimension: action. An AI agent doesn't just store or read — it acts, writes, sends, decides. When these agents proliferate without an inventory, the organization loses track not only of its data but of what machines are doing on its behalf.
The numbers show the scale of the risk: 71.7% of AI tools used in the workplace are high or critical risk, and 34.8% of the data entered into these tools is sensitive — a share that keeps rising year over year.
Why banning fails
The knee-jerk reaction — banning — backfires. Generative AI is one click away, free and immediately useful, while internal approval processes are slow. A ban doesn't remove the usage: it makes it invisible, and therefore more dangerous. Shadow AI thrives precisely in the gap between the usefulness employees perceive and the slowness of official channels.
Inventory, frame, oversee
Effective Shadow AI governance rests on three levers:
- Inventory — build a living inventory of the AI tools and agents actually in use. Standards are emerging, such as agent registries (the Linux Foundation is backing a DNS-based directory). You can't govern what you haven't inventoried.
- Frame — offer approved, secure alternatives that are more convenient than the rogue tools, with clear rules on which data is allowed.
- Oversee — log access, scope permissions tightly (keys with limited scope), and keep a human in the loop on sensitive actions.
Shadow AI isn't an employee discipline problem but a governance signal: it reveals the gap between how AI is actually used and the organization's ability to govern it.
Frequently Asked Questions
What's the difference between Shadow IT and Shadow AI?
Shadow IT refers to using digital tools or services (SaaS apps, cloud storage, personal accounts) without approval from the IT department. Shadow AI is its AI-specific variant: an employee pasting customer data into a consumer chatbot, a team plugging an agent into a database with no guardrails. Shadow AI is riskier because it combines two dangers: exfiltration of sensitive data to third-party services and the autonomous action of agents operating with no inventory and no oversight.
Why is Shadow AI exploding in organizations?
Because generative AI has become one click away, free and immediately useful, while internal approval processes are slow. The gap between perceived usefulness and the effort of official adoption pushes employees to work around IT. The result: organizations lose track of their own agents — an assistant plugged in here, a content agent deployed there, a workflow connected to a database without anyone keeping a register of what's running, with which access, on which data. You can't govern what you haven't inventoried.
Should Shadow AI be banned or governed?
Banning is ineffective and counterproductive: usage continues, but hidden, and therefore more dangerous. The approach that works combines three levers. Inventory: build a living inventory of the AI tools and agents actually in use (agent registries, automated discovery). Frame: offer approved, secure alternatives that are more convenient than the rogue tools, with clear rules on which data is allowed. Oversee: log access, scope permissions, keep a human in the loop on sensitive actions. Governing the practice beats denying it.