How much autonomy does an agentic AI system actually have?

Nielsen Norman Group's "Artificial Intelligence" glossary (Caleb Sponheim, 2026-08-21) defines "Agentic AI" not as a yes-or-no label but as "a description of how much latitude a system has between prompts," and states directly that "autonomy is a matter of degree, so calling a product an agent says little by itself about how much autonomy it actually has." The glossary's "Agent" entry backs this with a structural claim: the product around an agent "controls its tools, permissions, stopping rules, and opportunities for human review" — the agent label describes a capability, but the actual autonomy a given deployment has depends on choices made outside the model. The "Human-in-the-Loop" entry applies the same degree-not-label logic to oversight: it is "one common control for high-stakes automation," but "it counts as oversight only when the reviewer has the information, time, expertise, and authority to intervene." The glossary adds the failure case plainly: "A confirmation step users click through without reading adds delay instead."

The reasoning in both entries is the same move applied twice. Calling a system "an agent" or a workflow "human-in-the-loop" describes the presence of a mechanism, not its adequacy. An agent can choose "several steps and use tools toward a goal without being told each move," but how much latitude it actually has is set by permissions and stopping rules the product defines, not by the agent label. A human-in-the-loop step can exist on a flowchart while failing every one of the four conditions — information, time, expertise, authority — that would make the human's presence meaningful. In both cases, the glossary is converting a binary design vocabulary (is it an agent, is there a human in the loop) into a graded one that has to be checked against specifics.

This extends Does machine agency exist on a spectrum rather than binary?, which argues from Rammert's taxonomy that agency itself runs from passive to cooperative rather than splitting cleanly into human versus machine. NN/g's glossary arrives at the same non-binary conclusion from product-design vocabulary rather than social theory, and adds a concrete test — can the reviewer actually intervene — that the agency-spectrum literature frames more abstractly. It also sharpens How should AI interfaces handle the shift from doing to supervising?: the orchestration layer, where agents are supposed to negotiate and get consent for plans, is only doing real oversight work if it satisfies NN/g's four-part test rather than simply existing as a step in the interface.

The glossary names the conditions for genuine oversight but does not operationalize them: it gives no method for checking, in a shipped product, whether a reviewer actually has enough information, time, expertise, and authority at the moment a confirmation appears. It also does not say how much latitude an agent needs before the "agentic" label stops being misleading. Both entries describe what to ask rather than how to measure the answer, so applying either claim to a real system still requires a separate evaluation the glossary does not supply.