Anna, who works in customer success at Tars, tells this one on herself. She ordered gym gloves, and on delivery day realized the shipping address was an old hostel where packages get left in a box outside with no call and no verification. The item was non-refundable and the parcel was already out for delivery. She went to the support chat to stop it. The bot told her not to worry, her package was out for delivery, which was precisely the problem, and then offered only template options. No free text, no way to explain the situation, no route to a person.

By the system's own metrics that was a successful conversation. She asked for delivery details and got the courier's number, the status, and the tracking number. The question was answered. The situation was not. The counter-example is a colleague whose regular food order arrived stale and who got an immediate refund, no questions and no photos, because she was a repeat customer who had never raised a complaint before and the system had the context to know it. The argument in between is that most agents are optimized for answering questions, while customers arrive with situations, and the usual measures cannot tell those apart: was the question answered, did retrieval return the right thing, did the conversation end without escalation.

What she asks for instead is three things a system has to understand. Context, meaning what is actually happening to this person. Intent, meaning what they are trying to solve. And escalation timing, meaning the moment the agent should stop trying and bring in a human. Watch this if your support reporting looks healthy and your customers do not sound it.