Enhancing a Product Availability Voice Experience with Generative AI
As part of Lowe's Conversational AI team, I led the transformation of a legacy product lookup system into a modern, agentic conversational AI experience powered by a generative LLM, reducing rigid call flows and enabling dynamic, intent-driven dialogue.
The goal was to modernize the rigid product lookup experience into a dynamic, natural, intent-driven dialogue ultimately improving customer containment and satisfaction while increasing business value and revenue.
Timeline
1 Month
Ongoing - Analyzation & Optimization
Project Type
IVA - Intelligent Voice Agent
Roles
AI Product Designer & Prompt Experience Designer & Data Analyst
Problem 1: The existing product availability IVR experience provided structured responses when customers gave an item number. When customers gave a product description, the system offered only vague responses such as "I found a few items that meet your description," followed by an SMS to a generic product landing page. This limited interaction created a poor customer experience and failed to drive value for more open-ended queries.
Problem 2: Due to technical constraints, the IVR was limited to what kind of input it could understand at each prompt limiting the types of questions customers could ask resulting in high transfers to store associates.
Unless tens to hundreds of utterances were built into each node, the IVR was consistently restricted to yes/no answers and occasionally other phrases.
The Problem
The Opportunity
With advancements in LLM technology and agentic systems, there were opportunities to:
Understand natural language descriptions more effectively and inform the caller about the most relevant products
Generate informative, persuasive product summaries and answer any follow-up questions
Improve overall containment by guiding more calls to completion without agent transfer
My Role & Design Approach
Pre-release: I owned the agent experience design and prompt architecture for the new generative AI experience, focusing on the combination of how the platform would handle the logic for agent’s response & how the agent would guide the caller through the shopping experience.
Post-release: I owned analyzing production data - finding bugs that weren’t discovered through testing & uncovering pain points - and fine-tuning the instructions to enhance the experience.
Prompt Tuning & Testing
I led the prompt design for the enhancement for:
The backend input analysis (signal quality, intent clarity)
The customer-facing product summary (tone, structure, information hierarchy)
We piloted the experience in one store, then expanded to 10 stores based on initial performance.
Iterative Improvements
Production conversations revealed the need to:
handle different error codes returned from product lookup API
refine the amount of product details the agent needed before searching
Results
Containment:
Legacy: 15%
New: 43%
Nationwide rollout to all Lowe's stores after successful pilot resulting in ~123K labor hours saved annually.
What’s Next
After returning to the project, I did further analysis to dig into why the remaining 57% of pa calls weren’t being contained. I found that:
customers were using “product availability” language at the main menu as an umbrella term when in reality they could have been at any point in the “product” journey
the shopping experience needed to be enhanced with OOS handling, fulfillment type filtering, and product comparison buying guides
Impact
This project exemplifies how thoughtful LLM integration, strategic design, and prompt engineering can transform a legacy IVR experience into a smarter, revenue-driving system that aligns both user needs and business outcomes.