The question is: what can it actually do in my systems? Not what it can say, and not what it can answer. If the honest answer is that it responds and then passes anything real to a human, that is a chatbot, possibly an excellent one, but nothing new. If it can update the CRM record, move the appointment, issue the refund, and knows where it should stop and fetch a person, that is something else.

Underneath the question sits a four-word definition worth stealing: model, tools, loop, goal. The model is the brain and its ceiling is that it can explain your return policy without ever processing the return. The tools are the hands. The loop is the actual differentiator, and the worked example makes it concrete, walking every step a real agent takes to move a booking from Wednesday to Friday. The goal matters because agentic does not mean self-directed, it means that once a human sets the goal, the agent pursues it independently. The turn comes late and it is the important part. When a chatbot gets confused it gives a wrong answer, which is annoying. When an agentic system gets confused it takes a wrong action, which is money, trust, and compliance. That failure mode is the reason Tars is built the way it is, with the agent free to decide what to look up, what to ask, and which tool to reach for, while the steps that write to your systems run as deterministic flows that execute the same way every time.

Watch this before your next vendor evaluation. It is the difference between an AI that responds and an AI that resolves.