GenAI Support Agent

Final Project: Speculative concept extending my work on Dell.com/Support

Certification Program: Product School

Timeline: October 2024

Role: Product Manager — Final project PRD, functional prototype development

The Problem

Finding product support for your Dell devices can be a daunting process. Most users aren’t sure where to begin or how to find the correct solution. 

Manually creating service requests can be time-consuming, prone to errors, and often leads to miscommunication for both consumer and enterprise customers. This slows down the resolution process, particularly when diagnosing complex enterprise hardware and creates a frustrating user experience.

Strengthen Moat
(Defense)

  • Data generation through customer interactions

  • Personalization based on devices, account, warranty status, etc.

Monetize Functionality
(Offense)

  • Faster resolution creates a better customer experience

  • Improved efficiency in accurate diagnostics, ordering parts, dispatching service, etc.

  • Reduced operational costs with a reduction in support center call volume

The Opportunity

By leveraging GenAI to help diagnose technical issues and automate the creation of Service Requests, Dell Technologies can enhance customer experience, reduce operational costs, and improve efficiency. 

This feature taps into multiple data sources, including SupportAssist™, knowledge base articles, drivers and downloads, manuals, and community discussions, to streamline customer support

PRD Abstract

The Dell Support Agent feature will automate diagnosis and issue resolution for both consumer and enterprise customers.

Using GenAI, this agent gathers diagnostic information from pre-installed apps like SupportAssist, the support knowledge base, drivers and downloads, and community forums. It guides users through a step-by-step troubleshooting flow, generates detailed service requests, and dispatches parts or technicians. 

Customers can check the status of their requests at any time and receive updates through their preferred communication channels (SMS, email, WhatsApp, etc). If the agent cannot resolve the issue, it seamlessly escalates to a human representative and passes along all relevant details.

OKRs

We’ll know we’re successful when we see the following:

GOAL
METRIC
Increase Customer Satisfaction by improving the overall support experience. Increase CSAT by 5% QOQ
Improve customer effort score for support sessions Improve CES by 10% QOQ
Improve the accuracy of the support model and troubleshooting recos Increase the accuracy of AI-driven troubleshooting to > 95%
Increase efficiency in Support Operations Reduce the average time to resolution by 20% compared to the current baseline by EOY
Achieve a 5% reduction in overall call center volume* for issues that can be resolved by the AI assistant by EOY
*Break-even point is approximately 2%

User Flow

Model Requirements

SPECIFICATION REQUIREMENT RATIONALE
Open vs Closed Source Closed Source Performance and scalability are crucial due to high support demand (~3MM visits daily).
Context Window 128K+ tokens A large context window is necessary for lengthy conversations and multiple troubleshooting attempts.
Modalities Multimodal Model should support image inputs to support Service Tag detection.
Fine-Tuning Capability Fine-tuning required The support app requires deep product knowledge and specialized diagnostic flows
Latency Low Latency (100ms-300ms) Since this is a real-time customer support application, low-latency models are critical for maintaining a smooth user experience.
Size Larger models (10B+ parameters) Larger models (10B+ parameters) for handling nuanced, complex support queries while balancing cost and performance.

Data Sources

RETRIEVAL AUGMENTATION 

  • Dell Customer User Profile (authenticated)

    • Device info

    • Warranty status

  • Dell SupportAssist app for real-time diagnostics (on eligible devices)

  • Support Knowledge Base

  • Drivers & Downloads for the latest software and security patches.

  • Dell Manuals and Documentation for product-specific information.

  • Dell Support Community for common user-reported issues.

Prompt Requirements

Risks & Mitigations

RISKS MITIGATIONS
Model Hallucination and Incorrect Answers Continuously update the AI model with new service requests, issues, and solutions. Regularly validate the model’s performance against technician reports to ensure accuracy.
Frustration with AI Responses Use sentiment analysis to detect customer frustration and trigger human escalation when needed. Ensure that the AI promptly offers the option to speak with a live agent.
Data Privacy Implement strict encryption and data privacy protocols, particularly for enterprise customers dealing with sensitive system data. Ensure compliance with GDPR, CCPA, and other relevant regulations.
High Operational Costs Optimize Token Usage: Optimize interactions to minimize token usage by avoiding redundant or verbose responses from the model.
Use batch APIs (where available) to reduce operational costs by processing multiple requests simultaneously.
Overload During Peak Times Deploy the AI system on scalable cloud infrastructure that can automatically adjust based on traffic demands.