Shadow AI is employee use of AI services outside an organization's sanctioned tools — on personal accounts, personal devices, or unauthorized SaaS subscriptions. It is a successor concept to "shadow IT," the employee use of unauthorized cloud applications that defined enterprise IT challenges in the 2010s.
Definition and framing in NIST AI 800-4
NIST AI 800-4 (March 2026) identifies shadow AI as a monitoring challenge under its visibility-and-transparency cross-cutting issues:
Not knowing who's using what systems where, including… on personal devices, poses a real challenge to monitoring in enterprise use cases.
In some enterprise cases, the workforce may have access to authorized AI tools but still choose to use AI services on personal accounts for certain queries. The result is that an organization's monitoring infrastructure, even when well-built, has a blind spot for some unknown fraction of actual AI usage.
Scale
Vanta, a compliance and security platform, reached $300M in ARR in April 2026, which a Fortune disclosure (April 29, 2026) attributed to what it called the explosion of shadow AI across corporate America. The disclosure framed the situation as one in which most corporate employees in America have signed up for ChatGPT, Claude, and Cursor without their employers' security teams having visibility or control (Source: click.mail.fortune.com). A single compliance vendor reaching $300M ARR offers one quantitative proxy for the scale of the corporate-purchasing response to shadow AI. The framing also aligns with NIST AI 800-4's qualitative warning, suggesting that by mid-2026 shadow AI had become a default state of enterprise AI adoption rather than an exception.
Comparison with shadow IT
Shadow IT primarily created data exfiltration and compliance risks: employees putting confidential data into unauthorized cloud apps, where it could be lost, leaked, or fall outside the organization's compliance regime.
Shadow AI carries those risks alongside additional dimensions:
- Decision propagation — outputs from shadow AI tools become inputs to authorized work products, decisions, and communications, so the organization's official outputs are silently AI-shaped.
- Skill attribution — employee perceived competence diverges from actual competence in ways the organization cannot measure (the LLM fallacy connection).
- Vendor risk — unsanctioned AI subscriptions create vendor relationships the organization has no visibility into and no contractual control over.
- Training data exposure — many free or consumer AI tools may train on user inputs by default, so unsanctioned use can expose organizational data to model training.
- Regulatory exposure — sector-specific regulations (HIPAA, GLBA, FERPA, EU AI Act high-risk classifications) may apply to data the employee discloses to a shadow AI tool, with the organization on the hook.
Factors that make it difficult to address
- Employees experience shadow AI as productivity-enabling rather than as a policy violation. Friction in approved tools (slower, more limited, fewer features) creates demand for unsanctioned alternatives.
- Personal devices and BYOD policies make endpoint-based monitoring infeasible.
- Browser-based AI and mobile AI apps have minimal network signatures.
- AI usage is increasingly treated as an employability signal, which can incentivize employees to use whatever tools maximize their personal output regardless of policy.
Proposed mitigations
- Sanction high-quality tools quickly — much shadow AI usage reflects deficiencies in the approved alternatives, so reducing friction in sanctioned tools reduces demand.
- Monitor outputs rather than access — instead of detecting AI use at the endpoint, examine the volume, style, and signature of outputs.
- Clarify policy around AI-shaped outputs — for many roles the issue is non-disclosure rather than use itself, so clear disclosure norms reduce the harm.
- Training and AI literacy — give employees frameworks for what to put through AI and what not to (see Cognitive Friction for the related individual prescription).
Relationships
- surfaced-in: NIST AI 800-4: Challenges to the Monitoring of Deployed AI Systems
- related: Post-Deployment AI System Monitoring, LLM Fallacy, Cognitive Friction, AI Literacy, Enterprise AI Deployment Gap
- instance-of: AI Governance (umbrella)
- predecessor: Shadow IT (general)