Agentic AI describes systems and approaches that delegate the choice and execution of steps toward an objective to AI agents, with varying degrees of autonomy.
Terminology varies. NIST's overview emphasizes independent decisions, adaptation, and behavior directed toward goals. The OECD's 2026 conceptual analysis places greater emphasis on coordination among multiple agents and sustained operation. This glossary uses the operational definition above; it does not require multiple agents or a large language model.
What is delegated
An AI agent is a participating system. Agentic describes how work is organized around the choices and actions delegated to agents. That arrangement can leave some steps fixed, give an agent discretion over others, and reserve particular decisions for people.
Describe the delegation at the level of the workflow: who sets the objective, who selects intermediate steps, who may change the plan, and where execution must wait. Simply adding another agent does not resolve any of those questions. Coordination and autonomy need their own descriptions.
An example of bounded delegation
Hypothetical example: an internal training workflow lets one agent examine available sessions, apply attendance constraints, select a suitable date, and reserve seats in an internal catalog. The agent can choose among eligible sessions; it cannot change the training requirement or enroll people outside the assigned group.
Requiring a person to confirm the reservation would preserve the agent's scheduling discretion while limiting what it can execute without review. Requiring a person to select the date would reserve a different decision. These are separate design choices, even if both versions carry the same agentic label.
Autonomy is not an assurance
The label does not establish that a workflow is safe, governed, or free of human supervision. It also does not show whether a documented approval requirement is enforced. The actual control needs to operate at the relevant step, with evidence that supports reviewing its effect.
For a concrete description, name what was delegated and what remains constrained. An agent's ability to choose a schedule, for example, says nothing about its authority to expand the participant list. A successful reservation does not prove that every earlier decision followed policy.
Related concepts
The term AI agent identifies the system that uses context to select and carry out actions. The distinction keeps the participating system separate from the way the broader workflow delegates work.
AI agent governance addresses authority, controls, and evidence. The governance pillar explains how to apply those concepts to an enterprise workflow's tools, approvals, and execution paths.