Sentiment analysis runs on the conversation in real time, classifying the caller's emotional state and tracking how it changes turn by turn. The output is usually a scalar (positive to negative) plus tags like 'frustrated,' 'confused,' or 'urgent.'
The signal is real but only useful when wired into action. A dashboard that says 'this call has negative sentiment' after the call is over has no operational value. A system that triggers an escalation when sentiment crosses a threshold mid-call, hands off to a supervisor, and surfaces the trigger in the post-call review, that is sentiment doing real work.
For AI agents specifically, sentiment is a useful escalation signal: a sustained negative-sentiment turn sequence routes to a human before the customer hangs up frustrated. The customer never knows the sentiment signal exists; they just notice the experience getting better at the right moment.
Vorel uses sentiment as one input to the escalation policy. Sustained negative sentiment routes to a human, with the conversation context and the sentiment trace attached.