An agent is not a chatbot and not a rule. It owns a goal — qualify this lead, in the channel they wrote in, in our voice — and adapts each step to the conversation in front of it, escalating to a human when confidence drops.
The useful boundary is simple: the agent handles the work between "a lead lands" and "the human matters." Everything before the human judgment call, it does end to end. The human reviews the agent's day in fifteen minutes, not the inbox.