Most support deployments assume a clean handover: the assistant tries first, and a person takes over when it cannot cope. The handover is where these systems break, and the reasons are operational rather than conversational.
The bot does not know it is failing
Escalation requires detecting that the conversation is going badly, and the model producing confident wrong answers is the wrong judge of that.
Most systems therefore trigger on proxies: repeated questions, negative sentiment, an explicit request for a person, or a turn count.
Each proxy fires late. By the time a customer has asked the same thing three times, the escalation has already cost more goodwill than it saves.
The transcript arrives without the reasoning
An agent receiving the conversation sees what was said, not what the system consulted or why it answered as it did.
If the assistant quoted a policy incorrectly, the agent has to reconstruct where that came from before they can correct it.
Handovers that pass the retrieved documents alongside the transcript resolve far faster, and most products do not pass them.
The customer has to start again
Identity verification frequently sits in the human workflow rather than the bot workflow, so the first thing the agent asks for is information already given.
That repetition is the moment customers describe as the system wasting their time, and it is caused by two tools not sharing a session.
Where verification is completed before escalation and carried across, satisfaction after a handover approaches that of a direct call.
Staffing assumptions collapse
Automating the simple half of the queue leaves agents handling only difficult cases, back to back, with no easy calls in between.
Average handling time rises, which looks like a productivity decline in the reporting, and the team is judged against targets set before the change.
Attrition follows unless the metrics are rewritten, and the deployments that hold together are the ones where they were rewritten first.
The queue position is the real complaint
An escalated conversation usually joins the back of the human queue, so the customer waits the full time after already spending several minutes with the bot.
Placing escalated conversations at the front reverses the experience entirely, at the cost of the customers who went straight to a person.
That is a policy decision about who waits, and it is decided far more often by default configuration than by anyone choosing.