Every turn,
measured.
We surface per-turn sentiment, prosody, and frustration in real time — then put it to work for you: driving escalation rules, coaching your agents, and flagging the calls that didn’t go well.
Not just positive
or negative.
Six independent dimensions, fused into a single sentiment score — and exposed individually for the cases where the dimension matters more than the headline.
Valence
The pos / neg dimension. The number behind the headline score.
Arousal
How activated the speaker is. High arousal can be excitement or anger.
Frustration
Specific to friction — sighs, “really?”, “are you serious”, curt clipping.
Volume delta
Caller getting louder turn-over-turn — strong escalation predictor.
Interrupt rate
How often the caller cuts the agent off. Often a leading indicator of churn.
Pause length
Long thoughtful pauses vs. resigned ones. Subtle but real.
Word polarity
Lexicon-based. Picks up “lawyer”, “refund”, “scam” the moment they appear.
Recovery delta
Sentiment(end) − Sentiment(start). Where the agent actually changes things.
When sentiment moves,
something happens.
Sentiment alone is data. Tied to a rule, it’s a workflow — and our team wires those rules into your agent for you, often alongside sentiment-based call transfer. Three we deploy most:
Escalate to human #1 trigger
Page the on-call retention rotation, attach the live transcript, warm-transfer the caller within 4 seconds.
Switch tone #2 trigger
Agent shifts to empathy mode — slower pace, shorter sentences, lead every reply with acknowledgement.
Pin for QA #3 trigger
Auto-pin to the QA queue, with the three weakest turns annotated, so our team can review and refine the agent.
−0.20
is the cliff.
In the calls we run, the chance of a clean resolution falls sharply once sentiment crosses −0.20 and stays there for two turns. That’s why we set it as a sensible default escalation trigger — and tune the threshold to your calls over time.
We turn sentiment on for every Convexa agent we run for you. No flag needed.