Customer sentiment vs agent performance: not the same

Short answer

Customer sentiment is a label that says whether the customer's words in a call read as positive, neutral or negative. Agent quality is whether the agent acted according to the company's standard. The two often overlap, but they are not the same: a customer arrives already angry, the agent does everything right, and the customer still leaves unhappy — that is not a bad call.

So sentiment should be used as a context signal, not as the agent's score: to choose which calls deserve a closer look and to see which topics customers are unhappy about. What the agent did is measured by behavioural criteria.

What sentiment measures

Sentiment analysis classifies the emotional direction of expressions in text. In a call it is usually applied to the customer's speech. Platforms do this at different levels:

• One overall label for the whole call: positive, neutral, negative
• A label per conversational turn plus a trend line across the call — AWS's Contact Lens documentation, for example, presents customer sentiment trend and distribution this way
• Separate readings for the start and end of the call — "how the customer came in, how they left"

In every case what is measured is a classification of words. It is not a precise measure of the customer's inner feelings and should not be presented as a tool that "reads" a person's mood.

What sentiment does not measure

1. 1The agent's behaviourA customer describing a problem uses negative words — "my money is gone", "I've been waiting three days". That is a description of the problem, not the agent's fault.
2. 2The causeNegative sentiment can come from the product, the price, delivery, a previous call or the agent. The label does not separate them.
3. 3Sarcasm and cultural style"Great, just great service" can be said sarcastically. A direct speaking style can also read as displeasure.
4. 4Tone of voice — when based on textSentiment derived from a transcript knows nothing about the voice's tone, pace or intonation. That needs separate audio analysis and validation.

Illustrative example: four quadrants

The four cases below are an illustrative example. Each shows a different combination of sentiment and the agent's behaviour score, and each calls for a different step.

1. 1Negative sentiment, high behaviour scoreThe customer is angry about a late order. The agent acknowledged it, clarified the cause and gave a new date. Step: the agent did well; the problem is in delivery and is passed there.
2. 2Negative sentiment, low behaviour scoreThe customer started calm, but the agent did not answer the question and interrupted twice. Step: review this call first and take it into a feedback conversation.
3. 3Positive sentiment, low behaviour scoreThe customer left happy, but the agent skipped mandatory information or promised an unapproved discount. Step: the most dangerous quadrant — a report that looks at sentiment will never surface this call.
4. 4Positive sentiment, high behaviour scoreAll is well. Step: it can go into training material as a good example.

Ways to use sentiment well

• Selection for review: among negative-sentiment calls, listen first to those with a low behaviour score
• Topic analysis: which contact types collect negative-sentiment calls — that is information for product and process owners
• Change within a call: if the platform provides it, how the customer was at the start and at the end — an indirect hint of de-escalation, not proof
• Trend: a sudden week-on-week rise in the share of negative sentiment can signal a new problem, such as a system outage

Misusing sentiment

• Adding sentiment to the agent's score with a weight — the agent taking upset customers is systematically penalised
• Ranking agents by their share of negative-sentiment calls — whoever works the complaints queue always comes last
• Using sentiment instead of CSAT (customer satisfaction survey) — one classifies words, the other is the customer's own answer
• Basing reward or discipline decisions on sentiment
• Presenting the label as a measure of the customer's real feelings

Sentiment, CSAT and behaviour score: which is which

1. 1SentimentAn automatic classification of the words in the call. Available for every call, but indirect and error-prone.
2. 2CSATThe customer's own answer to a survey. Direct, but reflects only those who respond and rates the product as well as the agent.
3. 3Behaviour scoreWhether the agent carried out the steps of the company standard. Measures what the agent controls, but depends on the quality of the criteria.

Merging the three into one number is tempting, but each answers a different question. Showing them side by side is more honest.

How to show sentiment in a report

When a sentiment figure lands in a report, the manager reading it often treats it as a quality metric. The report's structure can prevent that.

1. 1In a separate blockSentiment is not shown in the same table or on the same colour scale as the agent's score. It goes in a separate "customer signals" section.
2. 2By topicNot by agent, but by contact type: the share of negative sentiment on "delivery", the share on "refunds". That is information that finds its owner.
3. 3Crossed with the behaviour scoreThe most useful table is the count of calls in each of the four quadrants. The third quadrant — positive sentiment, low behaviour score — gets its own row.
4. 4With a method noteOne sentence under the report says how sentiment is computed, and in which language and recording quality it is reliable.

That structure turns the manager's question from "who is bad?" into "what are customers unhappy about, and who is quietly breaking the rules?"

Talking to agents about sentiment

Telling an agent "your calls have a lot of negative sentiment" is useless: they do not choose their customers. A useful conversation starts from a specific call: when, and after which sentence, did the customer become more unhappy? If it was after the agent's answer, that answer is what gets discussed. If the displeasure was there from the start and did not change, the question is different: did the agent acknowledge it and explain what was possible?

Limits

• A sentiment label is not a precise measure of the customer's inner feelings
• Sentiment taken from text does not reliably recognise tone of voice or sarcasm
• Speakers separated wrongly can attribute the customer's words to the agent, or the reverse
• One platform's sentiment metric cannot be compared directly with another's
• Sentiment is not the same as agent score, CSAT or FCR

What Vexvon Audio Analyzer offers

• Each call shows one overall sentiment label: positive, neutral or negative
• That label is not part of the call's 0–100 score: the score is computed only from the met, partial, missed or not-applicable statuses of your call standard's steps
• Each step comes with a comment and transcript lines — so the evidence that separates the fourth quadrant from the third is right there
• Each transcript line is attributed to the agent, the customer or "unknown"

More: Vexvon Audio Analyzer. The behavioural criteria for a call with an upset customer are in the de-escalation call scorecard.

First step

Take 10 negative-sentiment calls from last week and review each with behavioural criteria. In how many did the agent act correctly? Then review 10 positive-sentiment calls the same way — looking for the third quadrant. This article belongs to the support quality and coaching section. To try it on your own recordings, get in touch.

Further reading on this topic: audio analytics vs transcript analytics, sales call quality vs lead quality.