Companies measure assistant performance on resolution rates, but the interactions customers remember badly are the ones that should have reached a person sooner. Handoff is where deployments fail.
Recognizing failure is harder than producing an answer
An assistant generates a response with no reliable internal sense of whether it was correct. Confidence in generated text does not track accuracy in a usable way.
So the system cannot simply escalate when unsure, because the signal it would escalate on is unreliable. It has to infer difficulty from the conversation instead.
Common proxies are repeated rephrasing by the customer, explicit frustration, or several turns without progress, and each of these detects trouble only after it has already occurred.
Context rarely survives the transfer
The complaint that customers raise most often is having to repeat everything. This happens because the assistant and the agent desk are usually separate systems.
Passing a transcript is technically easy and practically insufficient, since an agent opening a call does not have time to read several minutes of conversation.
Useful handoff requires a summary the agent can absorb in seconds, which is a separate generation task most deployments never implemented.
Staffing math works against fast escalation
Assistants are deployed to reduce contact volume, and every escalation returns a contact to the queue. The financial case rewards containment.
Teams frequently tune thresholds toward keeping conversations inside the assistant, which is measurable, while the cost of a poor experience is not.
The result is systems that escalate late, which is the pattern customers describe when they say the assistant would not let them reach anyone.
Some categories should never be attempted
The strongest deployments route certain topics to a person immediately, regardless of whether the assistant could produce a plausible answer.
Typical categories in American consumer operations include billing disputes, account security, bereavement, and anything involving a regulated disclosure.
The reasoning is that a wrong answer in these areas costs far more than the contact saved, so the decision is made in routing rather than by the model.
The return path matters as much as the escalation
After a person resolves an issue, the conversation frequently continues later, and the assistant handling that follow-up needs to know what happened.
Where the agent's resolution notes are not available to the assistant, the customer encounters a system that has forgotten a call it participated in.
Closing that loop is unglamorous integration work, and it does more for the perceived quality of an assistant than improvements to the answers themselves.