Robotic process automation (RPA) is a recording of a human clicking through a fixed path. It is brittle and legible: change the button and it breaks, and its rigid regularity makes it comparatively easy to spot. It never decides anything.

Agentic automation is different in kind. Given a goal, it chooses its own next action, adapts when the path changes, and produces behavior that looks far more like a person improvising. The advance is not speed. RPA was already fast. The advance is discretion.

• RPA is a script: same inputs, same steps, detectable by its regularity.
• Agentic is a policy: same goal, variable steps, detectable only by intent and cost.
• Detection therefore shifts from "does this match a known pattern?" to "does this behave with human-like purpose?"

That shift is why unsanctioned agentic automation inside business platforms is such a live problem. It is sandpaper friction against every UI-level defense, and it erodes those defenses precisely because it stopped being a script and started making choices.