Hackathon, 2nd Place
Conversational AI agent for phone support
High call volumes led to excessive hold times and frustrated users. Traditional phone trees can force customers to listen to irrelevant options before reaching a human.
Overview
Goal
To deploy a Voice UI (VUI) agent capable of identifying customer intent, resolving low-level queries and intelligently route issues to the correct internal teams.
Role
As the sole designer, I defined conversational guardrails for various user personas and conducted Red Testing to expose where the conversational logic failed.
Team
9 people (including myself) across the organization from Design, Product, Engineering, Customer Support and Sales assisted in this project.
Outcome
Out of 7 teams that participated in the even, our team earned 1st place in the annual 2025 Hackathon.
Tools
Google Gemini, Notebook LM, 11 Labs, n8n, Google Sheets, Slack
Definitions
Conversation Guardrails
We trained the VUI with common and edge-case user personas such as a distressed user and non-native speaker.
Red Testing
A testing process to intentionally probe the AI system to force it to misbehave, bypass it's safety guardrails or reveal hidden vulnerabilities before bad actors do.
The Problem
WP Engine is known for the award winning customer service.
However, as the business scales and the customer base grows, support demand increases creating longer wait times.
Our human agents needed assistance routing and escalating support calls.
Instead of a menu-based hierarchy (e.g., "Press 1 for Sales"), we implemented an open-ended conversational trigger ("How can I help you today?").
Users that showed signs of frustration e.g., speaking over the bot, or repeated "I need a human" responses. This triggered immediate escalation to a live agent.
Limited VUI to handle low-level scenarios such as checking their balance.
With a global customer base, we also needed to ensure the VUI could handle multiple languages.
Lessons learned
The Red Testing process highlighted that the quality of the conversation is a business outcome, not just a design feature.
Minimizing robotic apologies like "I’m sorry, I didn’t get that" with "I need a bit more detail to get you to the right team." added a layer of intentional design.
Ensuring the agent could pull account-level data (e.g., "I see you're on an Enterprise plan") to adjust the tone and speed of the routing process dynamically prioritized high-value clients.
Outcomes
Following the internal pilot of the VUI agent, we determined that our customers prioritized human-in-the-loop support for complex technical issues.
As a result, the project served as a foundational research initiative, directly informing the architecture of our subsequent, highly successful AI-chat implementation.
WP ENGINE
Dynamic downgrade flows save customer support 81 hours a month
Automated a key workflow that enabled customers to downgrade their plan on their own, cutting support tickets and hours by 65% a month.
View case study
WP ENGINE
Plans page redesign brought in 8% increase in upgrades
This redesign simplified plan modification decisions, resulting in 49% retention rate, a 7.7% increase in upgrades, and 20% more add-on purchases.
View case study
