Pioneering AI & ML | Mary Salvatera

Overview

I led Comcast’s first generative-AI assistant, and made it feel human.

As Sr. Director of Product Design & UX, I set the strategy and design for Xfinity Assistant (XA), a generative-AI experience trained on millions of support interactions. I shaped XA’s persona and voice to speak in the first person, so understanding a customer’s intent felt less like a menu and more like a conversation, self-service that resolves the issue instantly.

The challenge

Support was fragmented across voice, chat, SMS, and social, and painful for customers at every turn. We needed one intelligent, consistent experience that scaled to a national customer base and deflected cost, without ever losing the human touch that keeps people loyal.

Generative AI

A virtual assistant that understands

Natural-language conversations powered by generative AI. XA reads context and intent, then answers instantly, handling 43% of Tier-1 inquiries on its own.

Machine learning

Proactive, not reactive

Real-time telemetry and ML notify customers about relevant events before they call, cutting incoming support volume by 20%.

Design system

One voice, every surface

A first-person persona and a shared conversational design system kept XA consistent across voice, chat, SMS, and social.

How the intelligence works

XA pairs generative AI for understanding with machine learning for anticipation, wrapped in a design system that keeps the experience human. Four layers work together to resolve issues before an agent is ever needed.

Intent engineGenerative AI interprets millions of phrasings and maps them to the real underlying need.

Proactive MLTelemetry and models surface the right nudge at the right moment, deflecting contacts.

Self-serviceInstant, personalized answers resolve common issues end to end, no queue.

Human handoffWhen it matters, XA routes to a live agent with full context intact.

The outcome

XA became a flagship AI experience across Comcast: $150M in value with $18M in first-year savings from deflected contacts, a 62% solve rate on the inquiries it handled, and an 89% NPS on assistant interactions, higher than human agents. It set the conversational standard other teams built on.

$150M

Value from deflected support

89%

Assistant NPS, beating agents

43%

Tier-1 inquiries auto-resolved

20%

Lower incoming support volume

Support, turned into a strategic advantage.

Machine learning met empathy and creativity, and set a new gold standard for customer experience at national scale.