Conversation to decision
The first closed-loop EQ decision model.
Emotion AI watches. Prism learns how people decide.
Prism coaches conversations in real time. It suggests the next move, checks whether it was made, and reads the reaction in voice, face, and words. Across thousands of coached conversations, those outcomes become statistical patterns of how people decide, linked to the emotional signals that came before.
REAL-TIME COACHING FEEDS EVERY OUTCOME BACK INTO THE MODEL
READINESS
0.34
- Decision trials in every conversation
- 100s
- Emotion metrics underneath
- 40+
- Voice, face, and language in one model
- 3 → 1
- Languages, live and post-call
- 100+
Fig. 02 · Spectrum
Emotion in.
Decisions out.
Prism fuses voice, face, and words into one read of where a person stands in a decision, and the move most likely to change it.
IN · 3 CHANNELS
OUT · DECISION CONTEXT
PRISM
Fig. 03 · The trial
Every conversation is an experiment.
Emotion AI learns by watching. Prism learns by coaching: it suggests a move, the move is made, the response is measured, and the outcome is recorded. A single coached conversation runs hundreds of these trials.
Δt < 300 MS · LIVE READS
01 COACHED MOVE
“So what made you look at this now?”
02 RESPONSE
VOICE
FACE
WORDS
Energy up
Leans in
“Honestly, we’re losing deals”
03 DECISION SHIFT
Problem urgency is high
04 FEEDBACK · OUTCOME LOGGED, PATTERN UPDATED
TRIAL 120 · 04:12
- 01 Coached move“So what made you look at this now?”
- 02 ResponseEnergy up · Leans in · “Honestly, we’re losing deals”
- 03 Decision shiftProblem urgency is high · readiness 0.34
- 04 FeedbackOutcome logged, pattern updated
Fig. 03B · Where the feedback comes from
Real-time coaching turns every conversation into data.
Watching alone can’t tell you what a frown leads to. Coaching can. Each suggestion Prism makes is a small test, each outcome is a label, and the EQ signals around it become the evidence.
- 01 · ReadEQ models read the moment.Voice, face, and words, every few hundred milliseconds.hesitation 0.74 · AU24 · 1.4 s pause
- 02 · CoachPrism suggests the next move.Live, on the rep’s screen or in the AI agent’s next turn.→ “Ask what’s driving the concern”
- 03 · ObserveIt measures what happened.Was the move made? How did the person react? Did the decision shift?move_taken ✓ · readiness 0.22 → 0.41
- 04 · CorrelatePatterns of how people decide.Across thousands of coached conversations, signal + move + shift become statistical decision patterns.hesitation at price + “ask why” → +18% readiness
↺ 04 → 02 · Every new pattern sharpens the next coached moveClosed loop
Fig. 04 · Micro-feedback
Micro-feedback, hundreds of times a conversation.
Learning from outcomes
1 label per deal: won or lost, weeks later
Learning from the conversation
240 labels, one every few seconds
- toward
- against
- neutral
240×
More feedback per conversation
~8 s
Between decision signals
Seconds
From move to label, not weeks
Fig. 05 · Method
- A recorded clip
- A rater guesses the emotion
- One label
- Something is said
- Voice, face, and words react
- The decision moves
Fig. 06 · Results
Feelings don’t predict decisions. Prism does.
Emotion on its own says little about what someone will decide. Add what led up to the moment and how the person responded to each move, and the prediction changes completely.
Bars start at chance (0.5). Same conversations, same moments; the only difference is what each model is allowed to see.
Two “okays.” Same pause, same lip press. Opposite outcomes.
- + Asked six questions earlier
- + Put a cost on the problem
- + Mentioned a Q3 budget cycle
The hesitation is about timing, not value.
Prism: likely yes, once timing is solved- − Asked no questions
- − Short answers since minute 10
- − Never named a cost of the problem
The hesitation is about value. The yes was never there.
Prism: unlikely, address perceived value firstFig. 06B · Components
Inside the model.
- Fig. 06B.1ModelFusion
Refract: three channels, one read.
The fusion model that reads voice, face, and words together, aligned word by word, with a confidence score on every read.
voice · face · words → 1Explore ↗ - Fig. 06B.2SignalChannel conflict
Tells: when the words and the face disagree.
Flags the hesitant yes, the polite no, and the pause before the price: unspoken concerns read from voice and face.
channel_conflict · concernsExplore ↗ - Fig. 06B.3EngineReal-time
Loopback: coaching that learns.
Suggests the next move live, checks whether it was made, and logs the reaction and the decision shift that followed.
move → reaction → shiftExplore ↗ - Fig. 06B.4PatternsLearned
Atlas: how people decide, mapped.
Statistical decision patterns from thousands of coached conversations. It ranks every possible next move by expected lift.
signal + move → liftExplore ↗
Fig. 07 · Beyond emotion
Ask what no emotion model can answer.
“Okay… that’s roughly what we expected.”
What should we say?
Ask what is driving the concern
Best next move. Expected lift: +18%.
Fig. 08 · Interface
Use the EQ models on their own. Deploy them live to learn decisions.
Start with emotion reads on recordings or streams. Turn on real-time coaching and every conversation also trains the decision model, so the Decision Context API keeps getting sharper on your own data.
{
"trial": 120,
"coach": {
"suggested_move": "So what made you look at this now?",
"delivered_to": "rep_screen",
"move_taken": true
},
"reaction": {
"voice": "Energy up",
"face": "Leans in",
"words": "Honestly, we’re losing deals"
},
"decision_context": {
"stage": "exploring",
"readiness_before": 0.78,
"readiness_after": 0.34,
"shift": "against"
},
"learning": {
"outcome_logged": true,
"pattern_updated": true
}
}Fig. 08B · Use cases
For conversations where decisions get made.
For sales & revenue teams
Know where every deal stands.
See where each buyer stands, what is holding them back, and the move most likely to change it, while the call is still live.
- Stage and stance per buyerRefract
- Unspoken concernsTells
- Live best-next-move coachingLoopback
“Okay… that’s roughly what we expected.”
→Ask what’s driving the concern. Expected lift +18%.
ILLUSTRATIVE WORKFLOWQuestions, answered.
It goes one layer past emotion: where a person is in a decision, what is driving it, what is blocking it, and the move most likely to change it.
Emotion models learn from clips someone labeled later. Prism also sees the move that caused each reaction and how the decision shifted afterwards, so it learns what an expression leads to.
Real-time coaching. Prism suggests a move during a live conversation, checks whether it was made, and measures the reaction and the decision shift that followed. Those outcomes, linked to the EQ signals around them, are what the decision model learns from.
Yes. The Emotion API works on its own, live or on recordings. Turning on real-time coaching adds the learning loop that powers the Decision Context API.
Yes. Prism re-weights to voice and language and shows which channels each read is based on, so a phone call is never a blind spot.
Both. Stream audio and video for live reads, or send recordings in batch. The output schema is the same either way.
Yes. Prism can be deployed on-premise, so audio, video, and every signal stay inside your environment.