Project Overview
Aviary AI is an enterprise software for insurance companies, banks, and other financial institutions to assist with account activation, recovery, and general outreach. The software provides a way for these institutions to produce their own AI that can navigate complex conversations over the phone.
The scope of work outlined in this case study is the voice agent creation flow that requires users to provide the software with information to allow the AI to speak on their behalf efficiently and accurately.
Project Outcomes
Average onboarding time decreased from 45+ minutes to ~20 minutes
78% of clients created an agent without help from the Onboarding Team (up from 0%, fully supported onboarding)
Reduction in support requests for prompt changes from 27 per week down to 6.5 per week
Number of agents created increased 3x from before updates
Understanding the Current Process
When I began working on this project, the voice agent part of Aviary was in a pilot phase, meaning a designated Aviary representative was personally helping each new client onboard their AI agents. Each onboarding call averaged over 45 minutes and gave clients very little insight on how the product actually worked.
My first steps were to:
Listen to the existing calls between our Onboarding Team and Aviary clients
Speak directly with our Onboarding Team to understand the biggest obstacles that currently exist and the intricacies of the voice agent setup process
Speak directly with Aviary clients to get a better understanding of their technical knowledge, current confusions, and general needs of the product
Here's What We Found:
Some clients did not know what a "prompt" was or why it mattered.
Clients could not make small changes after launch without making requests to our implementation team.
Key details like voicemail behavior were often forgotten until the agent was already live.
Exploring Approaches
Fully Open AI Chat
Initially, we built a fully open AI chat with general guiding questions. Conversations got too long and users couldn't easily leave and come back. Editing the prompt later was also nearly impossible, because the information was buried in the chat history and not organized in any way.
Assisted Prompting
After realizing that a fully open AI wasn't giving enough insight and control, we shifted to an assisted prompting method. This allowed users to write all of their own prompts, with a checklist-style assistant to flag errors or gaps in requirements.
For some users, this worked well, but it didn't solve for the fact that many users were very unfamiliar with prompting. During testing, assistance was requested far more during the Assisted Prompting exploration compared to the Fully Open AI Chat.
Solution: The Wizard
We began working on solutions to find a middle ground of the previous two iterations. Our goal was to find a way to allow users with high AI-maturity to write their own prompts, while giving others the ability to chat with the AI. In order to fix the original problem with the open AI chat, we broke the conversation down into shorter, categorized sections.
Project Outcomes
(Same as Shown at Top of Page)
Average onboarding time decreased from 45+ minutes to ~20 minutes
78% of clients created an agent without help from the Onboarding Team (up from 0%, fully supported onboarding)
Reduction in support requests for prompt changes from 27 per week down to 6.5 per week
Number of agents created increased 3x from before updates
A Note on Figma Organization:
As a designer who prides myself on design ops, design system upkeep and maintenance, and developer handoff efficiency, I work hard to ensure my Figma files stay tidy and organized. Here are a couple of screenshots of the project in Figma.
I'm always willing to walk through design files to showcase this organization, as I believe it is an extremely important part of design operations.
Let's connect
Looking to have a conversation? Reach out!
Whether you're a hiring manager looking to chat, a junior designer looking for mentorship, or an event coordinator interested in me speaking… drop a line!