What is sentiment analysis?

A classifier reads the message and outputs a score, typically between negative 1 and positive 1, or a discrete label (positive, neutral, negative). Modern systems use transformer embeddings or direct LLM prompting. Both perform well on clearly emotional text and poorly on sarcasm, idioms, and short utterances.

Three patterns: prioritize the queue (sort by most-negative first so frustrated customers get faster attention), alert on tone shift (route to a supervisor if sentiment drops sharply mid-conversation), and report trends (track average sentiment per product area, per agent, per channel).

Sentiment is correlated with satisfaction but not the same as it. A polite customer with a serious complaint may register as neutral. A casual customer venting may register as very negative without intending to escalate. Always pair sentiment with CSAT for the ground truth.