Shoppers have started to push back on automated service. The complaint is rarely that the chatbot is rude. It is that the chatbot cannot finish the job, and that getting to a person has become an obstacle course. Retailers reading this as a demand for better language models will miss the point. The failure is structural: a system built to answer questions has been deployed to resolve problems, and those are different tasks with different success criteria.
What shoppers actually complain about
Three complaints dominate. The first is repetition, where the assistant asks for information the shopper has already provided. The second is false confidence, where the assistant states a policy that turns out to be wrong. The third is the loop, where the only path to a human runs through a menu that never offers one.
Each complaint maps to a design decision rather than a model limitation. Repetition comes from systems that do not share context. False confidence comes from answering beyond the knowledge boundary. The loop comes from cost targets that treat escalation as failure.
The handoff problem
A handoff fails when the context does not travel with the customer. The shopper repeats the order number, the reason and the history, and each repetition raises the chance of abandonment. Escalation design, not model quality, decides whether automated service helps or hurts, and the evidence is that satisfaction drops sharply at the moment of transfer rather than during the conversation itself. Retailers that also route delivery and returns through assistants face the same test, as report shows.
Good handoff design has three properties. The transcript travels with the case. The human sees a summary rather than a raw log. The shopper is told what happens next and when. None of these requires a better model, and all of them require the service platform and the order system to share data.
Why the loop persists
Because it is measured as efficiency. A support organization that reports containment rate as its primary metric will always prefer to keep the shopper inside the bot, even when the outcome is a lost sale. The metric rewards the wrong behavior, and the behavior follows the metric.
The cost of a failed transfer
A failed transfer is expensive in a way that is rarely measured. The shopper who gives up does not usually complain; they simply stop trying, and the next contact is a chargeback or a public post. Support teams that only count inbound contacts will read the drop as success, while the cost moves to payments and to reputation, where it is much harder to attribute.
Why answers and service fail differently
An answer is judged on whether it is correct. A service interaction is judged on whether the problem is resolved, which often requires an action rather than a statement. An assistant that can describe a refund policy perfectly still fails if it cannot issue the refund, and shoppers do not distinguish between the two failures.
This is the same distinction that appears in search. A generative engine that summarizes product information accurately is useful, while an engine that promises availability it cannot verify is harmful. In both cases the boundary of authority matters more than the fluency of the output.
What a good transfer looks like
A good transfer is short. The shopper is told that a person will take over, given an estimate, and then sees the same case continue rather than start again. The human opens a summary that includes what was already tried. Nothing in that sequence is technically difficult, and almost none of it is standard practice, which is why the experience of automated service remains so inconsistent across retailers of similar size.
How to rebuild the answer layer
The first step is to define the boundary explicitly: which questions the system may answer, which it must route, and which it must refuse. The second is to instrument the boundary, so that every refusal and every transfer is logged with the reason. The third is to publish the rules internally, so that content, service and product teams write to the same standard.
The fourth step is the one most organizations skip: measuring resolution rather than containment. A bot that resolves sixty percent of cases and hands the rest over cleanly will outperform a bot that contains ninety percent and loses the customer, even though the second looks better on a dashboard.
None of this argues against automation. It argues against automating the wrong layer. Assistants are strong at retrieval, summarization and triage, and weak at accountability. Systems that keep humans accountable for outcomes and use automation to shorten the path will earn back the trust that the current generation of chatbots has spent.
