Agentic tool for 28M customers to reduce support costs
A conversational, telemetry-driven repair experience designed to reduce repeat support journeys.
Minimum viable experience for Xfinity Assistant Web chat's repair and troubleshooting agent.
Overview
This project focuses on Xfinity Assistant’s web-chat Repair and Troubleshooting Agent. I contributed to the experience over time and led the creation of a demo for senior executives. The agent launched to 5% of customers by July.
1. Context & stakes
Comcast’s shrinking subscriber base made retention a priority. Repeat support presented a $16.2M monthly cost opportunity.
Business and User Impact
About 482K high-effort customers daily struggled with step-by-step troubleshooting based on moment-in-time telemetry, spending 10.6× more time in support and increasing costs and churn risk.
Problem and bet
We bet conversational GenAI using real-time, historical and contextual telemetry could reduce repeat contacts across channels for an intended audience of 28M users.
My Role: Design the agentic interaction and prototype a production-ready, telemetry-driven experience for repair and troubleshooting agent.
4. Defining the experience before the technology was ready
I defined the PoC and MVP before technical readiness, using Xfinity components for low- to mid-fidelity wireframes that secured cross-functional buy-in. Legacy systems, budget limits and competing ownership slowed delivery; live event and telemetry integration with Google Cloud/Vertex AI remained unresolved.
5. Designing for reliability and trust
I prioritized prompt weaknesses affecting reliability: context retention, telemetry use, repetition, turn-taking and graceful escalation. I also held firm on setting expectations before automated diagnosis, keeping the customer experience central to every decision.
Outcome & next steps
Launched to 5% of customers by July. Next: measure call-volume and repeat-contact reduction; turn the GitHub prompt framework into skills and parameterize schemas for reuse.
