Why 52% more valuable has nothing to do with satisfaction.

Emotion is not a score · Prescriptive AI
Prescriptive AI  ·  Editorial Journal  ·  Perspective
Perspective · Emotional Dynamics

Emotion is not a score.

Satisfaction can be scored. Connection has to be read. Emotional Dynamics reads the shape of a feeling across 24 dyads, and Prescriptive AI turns that read into the next best move.

The number that does not move customers

In 2015, Harvard Business Review reported that customers who are fully emotionally connected to a brand are 52% more valuable, on average, than customers who are merely highly satisfied (Magids, Zorfas and Leemon). The authors went further: they estimated that moving a customer from highly satisfied to fully connected can yield about three times the return of moving them from unconnected to highly satisfied.

Satisfaction is what dashboards measure. Connection is what customers feel — and it is built, or lost, in individual moments: a hesitation before a question about personal data, a flash of irritation at a delay, relief when someone finally takes ownership. A sentiment score averages those moments away.

Where the score runs out

Two customers send messages that a sentiment tool scores identically: moderately negative. One is afraid of what will happen to their data. The other is angry about a second failed delivery. The score is the same. The right next sentence is not.

Fear needs orientation and a clear account of what is known. Anger needs acknowledgment and ownership before any explanation. Get it wrong and a recoverable moment becomes a lost customer. The problem is not that the score is inaccurate. It is that a score is a single axis, and feelings have structure.

Satisfaction can be scored. Connection has to be read.

The structure of a feeling

Robert Plutchik, whose psychoevolutionary theory Emotional Dynamics builds on, proposed eight primary emotions arranged in four opposing pairs: joy and sadness, trust and disgust, fear and anger, surprise and anticipation. Each varies in intensity — annoyance, anger, rage — and neighbouring emotions blend into dyads. Fear plus Surprise is Awe. Anticipation plus Fear is Anxiety. Surprise plus Anger is Outrage (Plutchik, 1980; 2001).

A second family of models describes emotion by position on dimensions rather than by category, most famously valence (pleasant to unpleasant) and arousal (calm to activated) in Russell’s circumplex (1980). Emotional Dynamics reads both: the dyad, and the valence and arousal around it. Twenty-four dyads, with intensity, direction and energy, is a far richer instrument than a plus or a minus.

Six dyads, six different next moves

These are six of the dyads that customer conversations meet every day, with the response principle each one calls for. The principles are illustrative.

Awe
Fear + Surprise
Disorientation. Orient before you inform: say plainly what the system can and cannot do, then offer detail.
Anxiety
Anticipation + Fear
Waiting for something that could go wrong. Give a concrete next step and a time.
Outrage
Surprise + Anger
A promise broken without warning. Acknowledge and own it first; explain second.
Disbelief
Surprise + Disgust
Reassurance has already failed. Offer evidence the person can check for themselves.
Despair
Fear + Sadness
Loss of hope. Hand over to a person, quickly and warmly.
Hope
Anticipation + Trust
Ready to proceed. Confirm, then make the next step effortless.

From describing to prescribing

Most analytics describes what happened. Better analytics predicts what will. The step that changes outcomes is the third: deciding what to do about it, in the moment, for the person in front of you.

The ladder Prescriptive AI climbs

  • Descriptive. What the customer felt. Most emotion analytics stops here.
  • Predictive. What is likely to happen next, such as churn risk or escalation.
  • Prescriptive. What to do and what to say, for the feeling in front of you. ConsentPlaceAgent reads the dyad and the confidence behind it, then shapes the next turn of the conversation accordingly. This is where ConsentPlace is shipping.

The honest objection

Emotion AI has serious critics, and they deserve a serious answer. In recital 44, the EU AI Act itself records “serious concerns about the scientific basis of AI systems aiming to identify or infer emotions”, citing limited reliability, poor specificity and weak generalisability across cultures, situations and individuals.

The scientific core of that concern was set out in 2019 by Lisa Feldman Barrett and colleagues. Reviewing the evidence on reading emotion from facial movements, they concluded that such movements are not reliable or specific enough, across contexts, individuals and cultures, to be treated as diagnostic of an emotional state. They called on technology developers, among others, not to treat facial configurations as reliable evidence of specific emotions (Barrett et al., 2019).

We take the lesson seriously and build on it. Four design choices follow from it.

Four design responses

  • Conversation, not faces. The critique concerns facial movements. ConsentPlace’s reads come from what people say, in the context of the conversation they are having.
  • Structure, not labels. Dyads, valence and arousal describe gradients instead of forcing a feeling into one box. That is consistent with Cowen and Keltner (2017), who found 27 distinct categories of emotion linked by smooth gradients.
  • Confidence, not certainty. Every read carries a confidence score, and lower-confidence reads can be handled more conservatively. Limits on reliability are reported, not hidden.
  • Customer-facing, transparent use. Article 5(1)(f) bans inferring emotions in the workplace and in education. ConsentPlace is designed for customer-facing conversations in which people are told what is happening.
A note on classification

Whether a system is an “emotion recognition system” is a deployment question.

The Act defines an emotion recognition system by reference to biometric data (Article 3(39)), and Article 50(3) requires deployers of such systems to inform the people exposed to them. How a given deployment is classified depends on what it reads and for whom, and should be settled deployment by deployment, with counsel.

What does not change is the principle: people should know, and the system should say how sure it is.

See it in the open

Two earlier sessions in our EU AI Act series show the read itself. In Episode 4, a user asks what the AI knows about them: the read is Fear, dyad Awe, at 58% confidence with a high topic-risk flag, and the response orients before it informs. In Episode 5, a user asks how bias is avoided: the read is Neutral with Curiosity, at 60% confidence, and the system does not project concern onto the question.

Both are simulated sessions, labeled as such, with the confidence shown rather than hidden.

What to do with this

Four moves for a customer-experience leader

  • Report dyads, not only scores. Track the share of conversations in Anxiety, Outrage or Despair, and what happens next.
  • Write a response principle per dyad before you automate anything. That is the prescription.
  • Set confidence thresholds for human handoff. A low-confidence read on a high-risk topic belongs with a person.
  • Tell people. Say when emotional reading is in use, and what it is for.

See a feeling read in real time.

Try the live demo and watch Emotional Dynamics turn a conversation into its next best move.

Try the live demo →

Sources

  1. Magids, S., Zorfas, A. & Leemon, D. (2015). “The New Science of Customer Emotions.” Harvard Business Review, November 2015.
  2. Plutchik, R. (1980). Emotion: A psychoevolutionary synthesis. Harper & Row.
  3. Plutchik, R. (2001). “The Nature of Emotions.” American Scientist, 89(4).
  4. Russell, J. A. (1980). “A circumplex model of affect.” Journal of Personality and Social Psychology, 39(6).
  5. Cowen, A. S. & Keltner, D. (2017). “Self-report captures 27 distinct categories of emotion bridged by continuous gradients.” PNAS, 114(38).
  6. Barrett, L. F., Adolphs, R., Marsella, S., Martinez, A. M. & Pollak, S. D. (2019). “Emotional Expressions Reconsidered: Challenges to Inferring Emotion From Human Facial Movements.” Psychological Science in the Public Interest, 20(1). pmc.ncbi.nlm.nih.gov
  7. EU AI Act (Regulation (EU) 2024/1689): recital 44; Article 3(39); Article 5(1)(f); Article 50(3). Official Journal of the European Union, 2024. eur-lex.europa.eu
  8. ConsentPlace simulated sessions, EU AI Act × Prescriptive AI Journal, Episodes 4 and 5 — simulated interactions run against ConsentPlace’s production system; no real customer or customer data is involved.

Leave a comment

Your email address will not be published. Required fields are marked *