INTRODUCING ALEA
Airbnb's Conversational AI Voice Agent
My Role
UX Designer
design Team
Sachin Dabas
Joy Mukherjee
Angie Li
Tools
FigJam
Lottie
VoiceFlow
Google's PALM 2 LLM
DISCIPLINEs
UX
Conversational AI
Voice Prototyping
Paper Modeling
Timeline
3 weeks · 2022
Description
A concept about a voice based conversational agent to reimagine the Hosting experience for Airbnb users.
contribution
I took ownership of the Hosting experience, while Angie worked on Guest experience. Joy took the lead in prototyping and deployment. Majorly I assisted in created scripts, voice prototype in voiceflow, data collection for the demo and creating the entire physical prototype for robot.
OUTCOME
I posted the project's video on Linkedin and received more than 30k impressions and engagement with comments like "this is awesome, let's make this happen" — Braden Ream, CEO, VoiceFlow and "this is really cool" — Krijn Janse, Sr. Conversational Consultant @CGI.
Linkedin Post
TL;DR
New Airbnb Hosting Experience.
Hey Alea,
any guest requests?
When’s my next check-in?
how to improve my listing?
Alea
Contains Audio. For best experience, I created an experiential concept video demonstrating the experience, and established a roadmap for future development.
Challenge
Design generative ai experience for Airbnb users.
Generative AI’s ability to respond to virtually any request makes it difficult to predict where a user journey may wind up. The journey may not be linear and certainly not predictable.
I was provided with following needs and constraints:
Identify user problems and find out of box solutions.
Identify opportunity areas for business growth.
Design the solution, test and validate.
constraints
Create a Conversational User Interface in 3 weeks.
Working in a team of 2 designers and 1 engineer, we were constrained to only use Conversational UI for the project. In a way, we already knew the solution without knowing the exact problem. An upside down design space — sounded weird at first, but isn't most AI products built this way nowadays? I had first encountered this in a design talk by Yuhki yam — confessions of modern design. Anyways, we approached the problem by looking at the user journey map.
Work
Home
About
Contact
JOURNEY MAPPING
Identifying gaps in the hosting experience.
When we started, the space was too big, and I was interested in looking at business side, so I decided to focus on Hosting experience, while Angie decided to take ownership of the Guest experience. This helped us to critically look at each other's work in critiques, and collaborate better!
Figma File — Interactive document
key User PROBLEMS
Identifying user problems in different stages of the experience.
focus area
Empowering hosts using AI
Key problems and hypothesis to solve each problem using Conversational UI.
Testing and building a conversational agent
Identifying gaps in the hosting experience using happy paths.
When we started, the space was too big, and I was interested in looking at business side, so I decided to focus on Hosting experience, while Angie decided to take ownership of the Guest experience. This helped us to critically look at each other's work in critiques, and collaborate better!
Figma File — Interactive document
building an agent
Creating a conversational user interface
Building the AGENT
Virtual agent Demo using Google Bard’s LLM.
Alea is trained on general AirBnB and travel resources with data scraped from web. It’s designed to seamlessly offer personalized and instant assistance to hosts, equipped with an extensive database of booking and listing records from thousands of users, including valuable insights derived from performance and booking data.
While VoiceFlow was great for prototyping and in-person user testing, we recognized the need for a sharable and experiential prototype for handoff. Hence we created a virtual agent using Google Bard’s LLM.
Unused Ideas · PRototyping
Airbnb AI-powered robots
A physical robot adds an in-person presence, making the interaction feel more natural and human-like. Guests may feel more welcome and taken care of with a “live-in concierge.” The robot could identify and resolve issues before they become guest complaints.
Concept of Robotic agent for Airbnb Homes — Interactive embeded doc
CONCLUSION & RESULTS
Users valued seamless, human-like responses for simple inquiries but struggled when navigating complex scenarios requiring contextual understanding.
07/12 users from our study rated the conversational user interface useful, showcasing the effectiveness of the conversational results.
SIGNIFICANT SETBACKS
Ambiguity in user queries!!
We tested our conversational interface for a happy path, but it still required understanding and deciphering accents, since in many instances, our voiceflow prototype could not understand what the user was saying. In cases when users paraphrase their requests, the system couldn’t easily interpret, leading to frustration and a breakdown in communication. This was particularly challenging since it was difficult to include so many inputs in the prototype.
To address this, we implemented an iterative process of refining the scripts by continuously testing against real user conversations. Through user feedback, we identified patterns in the way users interacted, allowing us to improve intent recognition and response generation.
Critique sessions with Angie and Joy. (Paper model on the table)
LEARNINGS & TAKEAWAYS
It is significant to handle errors with care and politeness.
Implementing robust and polite error handling ensures that when the conversation goes off track, the system can recover gracefully without frustrating the user. For example, if a user asks for a feature that doesn’t exist, the Conversational User Interface could suggest alternatives rather than merely saying no.
Upon given more time, I'll be addressing more edge cases, refining the conversational flows, and explore visual and auditory modalities of the conversational agent.