Shadow AI
Shadow AI is the use of AI applications, models, agents, or features for work without the visibility, approval, or governance of the organization responsible for that work. It can include an employee using a public chatbot, a team buying an unapproved AI tool, or a developer deploying a model outside established security, procurement, and monitoring processes.
What are examples of Shadow AI?
Shadow AI is defined by the lack of organizational visibility or control, not by whether the tool is free or paid. Examples include pasting internal material into a consumer AI assistant, using a personal account for company work, connecting an unreviewed agent to business systems, installing a local model on a managed device, or enabling an AI feature inside approved software without reviewing its data handling.
Microsoft describes Shadow AI as AI tools used without the knowledge, approval, or governance of IT or security teams (Microsoft Purview guidance).
Why does Shadow AI happen?
Employees often reach for an unapproved tool because it solves an immediate problem faster than the approved process. Teams may also lack a clear catalog of permitted tools, a practical review path, or an approved option that fits the task. Blanket restrictions can push usage further out of view when they do not address the underlying workflow need.
Discovery should therefore answer two questions: which tools are being used, and what work are people trying to improve? The second question can reveal useful AI demand as well as risk.
What risks does Shadow AI create?
The most direct risks involve sensitive data, access, retention, intellectual property, auditability, and regulatory obligations. An unapproved tool may process information under terms the organization has not reviewed. An unmanaged agent can also act with permissions that were never assessed for that use.
Microsoft’s Shadow AI guidance identifies data leakage, compliance violations, security vulnerabilities, and lack of auditability and governance as key risks (Microsoft 365 admin guidance). NIST recommends inventories of software, including AI, and monitoring of third-party AI tools within broader cybersecurity risk management (NIST Cybersecurity Framework FAQ).
Shadow AI also creates economic blind spots. Costs can sit on personal cards, departmental budgets, or bundled software contracts. Benefits and failures may never enter the organization’s measurement system.
How is Shadow AI discovered?
Use several evidence sources because no single inventory is complete:
- identity and access records for AI applications
- network, browser, endpoint, and cloud-application telemetry
- procurement, expense, and software asset records
- API keys, cloud resources, and model endpoints
- employee surveys and workflow interviews
The purpose is to create a usable inventory with an owner, workflow, data classification, access level, cost, and decision status. Discovery data should be handled proportionately and with approved privacy and employee-monitoring practices.
How should organizations manage Shadow AI?
Classify each use case rather than treating every discovered tool the same. A low-risk drafting task and an agent with production access require different controls. Common responses include approving the use, moving it to a sanctioned tool, limiting data or permissions, adding monitoring, or blocking it when the risk cannot be accepted.
NIST’s AI Risk Management Framework organizes this work through govern, map, measure, and manage functions. It also calls for third-party AI risks and benefits to be monitored and documented (NIST AI RMF Core).
How does Shadow AI affect AI value and ROI?
An AI measurement framework cannot produce a complete picture if material usage is missing. Shadow AI can hide costs, adoption, completed work, incidents, and quality problems. It can therefore distort both AI business value and AI ROI.
Once a use case is visible, measure it by workflow. Record what task it supports, who owns the outcome, the total cost, the accepted completed work, and the relevant quality and risk controls. That turns discovery into a decision about whether to approve, replace, improve, or stop the use.