An AI control plane is a logical management layer that configures AI policies, coordinates policy decisions, and provides visibility into how those policies are applied.
This glossary uses the term as an operational definition, not a universal product specification. It describes responsibilities that may span several components. A management service and an enforcement component can work together without every AI request passing through the management service.
Four responsibilities to distinguish
Configuration defines the rules, their scope, and the versions intended for particular environments. Saving a rule establishes an intended configuration, not necessarily the configuration currently in use.
Decision-making evaluates a request against the applicable policy and context. It determines an outcome such as permission, denial, or a requirement for approval.
Enforcement gives that decision an effect at a component able to constrain the operation. The decision and its enforcement can reside in different places.
Observability supplies information about configuration, decisions, execution, and failures. That information can support later review or feed another decision; it is not automatically a preventive control.
NIST's Zero Trust Architecture, SP 800-207 distinguishes the policy engine from the components that execute its decisions and enforce access. These are logical responsibilities, even when combined in a service. The distinction is useful here without requiring every AI system to adopt that reference architecture.
A saved policy is not a completed rollout
Hypothetical example: a reporting workflow changes its policy to require approval before an agent creates a public sharing link. An administrator saves the new rule, but one executor still uses the previous version.
The saved rule and the executor's effective policy now describe different states. A useful review checks which version the executor received, which rule informed the decision, and what constrained the operation. A console showing the new rule cannot establish those facts on its own.
What the label does not guarantee
A control plane does not necessarily intercept every prompt, tool call, or network connection. Coverage depends on the actual integrations and execution paths. The architecture should identify where decisions occur, which components apply them, and what happens when a policy update or decision is unavailable.
Central administration can support consistent rules, but consistency also depends on distribution, version handling, and enforcement. A single interface is not proof of a single effective policy across all environments.
Related concepts
AI governance sets the broader organizational responsibilities and rules. AI agent governance applies them to delegated agent authority.
The AI agent governance pillar explains how to review the execution paths, approvals, and evidence that connect those decisions to actions.