Giving Dynamics 365 Contact Center a richer view of customer emotion than positive-vs-negative sentiment can offer — built as a Microsoft hackathon prototype on top of Hume AI's Expression Measurement models.
A two-minute walkthrough: the problem with flat sentiment, the prototype on top of Dynamics 365 Contact Center, and a live call where multi-dimensional emotion would have changed how the agent responded.
Customer service is, at its core, about understanding and managing customer emotions. Yet in today's contact centers, that emotion is flattened into a two-dimensional scale — negative or positive — and called sentiment.
Human emotion isn't two-dimensional. It's layered, contradictory, and often hidden underneath the words a customer chooses. A caller who sounds upset may be perfectly calm; a polite voice may be masking deep frustration. For an industry hyper-focused on customer emotion, measuring it this coarsely means we are, in effect, blind to the very thing we claim to manage.
Emotion analysis is a budding field in AI, and a new generation of models can extract multi-dimensional emotion measurements directly from voice. Hume Vision is a prototype that wires Hume AI's Expression Measurement models into Microsoft Dynamics 365 Contact Center, surfacing live emotion signals to the customer service representative as a call is happening — without replacing the existing sentiment pipeline, but enriching it.
The premise: by opening our ears to how customers feel, not just whether they feel good or bad, we can empathize more accurately, route more intelligently, and escalate before frustration boils over.
Five components, arranged so that nothing in the existing Dynamics 365 codebase had to change. Each addition sits beside the existing sentiment pipeline rather than replacing it — a deliberate choice that kept the prototype shippable inside a hackathon window.
In flow order: caller audio moves through the Dynamics 365 Speech Orchestrator, which a custom API layer (built rapidly with Copilot-assisted code generation) taps and streams to Hume AI's Expression Measurement models. Multi-dimensional emotion scores return and are persisted to Dataverse alongside existing call telemetry. A Tampermonkey userscript then injects the live emotion signal into the agent's dashboard, surfacing it next to the existing sentiment indicator — no changes to the underlying Dynamics UI required.
The build followed a hackathon-tight loop: identify the gap (sentiment is too flat) → find a vendor model that fills it (Hume's Expression Measurement) → prototype the thinnest possible integration (API + Dataverse + Tampermonkey) → demo on a live call.
The demo itself was the proof. A test caller opens with: "I have to be honest with you, I'm kind of disappointed with the service…" — language a sentiment model reads as clearly negative. But Hume's emotion signal told a different story: the caller was calm, not escalated. With that richer read, the agent could see no undue escalation was warranted, and could meet the customer where they actually were.
From that single signal, a set of downstream product opportunities opens up:
Affective computing was deemed an important qualifier in the hackathon judging. However, Microsoft preferred leveraging in-house emotion-detection capabilities where applicable, and any path forward needed to abide by Responsible AI policies. The project was not pursued further as Emotion AI vendor onboarding and feature prioritization did not align with broader roadmap decisions.