Fleet 1.13:Teams are now shipping 5x more PRs with autonomous pipelines.See what's new →
FleetFleet
Glossary

Shadow AI

Shadow AI refers to the use of AI tools, agents, and services by employees or teams within an organization without the knowledge, approval, or oversight of IT, security, or management.

The term is modeled on 'shadow IT' — the long-standing phenomenon of employees using unsanctioned software. Shadow AI has emerged as AI coding assistants, chatbots, and autonomous agents became accessible without any procurement process. An engineer might pipe proprietary code into an external AI API, a team might deploy an AI agent with no security review, or a department might use a consumer AI tool that retains training data.

The risks are concrete: proprietary source code sent to external models may be retained and used for training, violating IP agreements; agents acting without oversight may introduce vulnerabilities; and costs incurred outside approved channels may be invisible until a large bill arrives. Shadow AI also produces inconsistent results because different team members use different tools with different quality.

Addressing shadow AI requires making the sanctioned alternative more convenient than the unsanctioned one. Heavy-handed bans typically drive usage underground rather than eliminating it. Organizations that publish clear AI tool policies, provide approved tools with adequate capability, and create lightweight escalation paths for new tool requests tend to see better compliance.

How this relates to Fleet

Fleet is a self-hosted, auditable orchestration layer designed in part as an antidote to shadow AI. Because Fleet runs on the team's own infrastructure, uses only approved model endpoints, and maintains a complete audit trail, it gives engineering teams the productivity benefits of autonomous AI agents without the visibility and compliance gaps that define shadow AI.

Frequently asked questions

How do I know if shadow AI is happening in my organization?

Signs include: unexplained API charges on individual developer accounts, code commits that show stylistic patterns inconsistent with the developer's history, team members referencing AI-generated content they cannot explain in detail, and productivity improvements in individual contributors not reflected in team tooling metrics. Security tooling that inspects outbound API traffic can provide more systematic detection.

Is all unsanctioned AI use harmful?

Not inherently. An engineer using a coding assistant for personal side projects on personal hardware presents minimal organizational risk. The concern is when proprietary code, customer data, or production systems are involved. Policies should distinguish between these cases rather than blanket-prohibiting all AI use, which tends to be ignored.

Run your first agent fleet

One binary. Five minutes. See every agent, coordinate every handoff, and keep a full audit trail of what your fleet did.