Conversational AI Design
Case Study
Expedia
As Senior UX Manager for Expedia’s Conversational Design Team, we were tasked with AI and Chatbot experiences for both Expedia’s customer and agent experiences. Expedia had just launched their AI Agent tool called Voyager Next. As a new member to the team, I quickly noticed two significant opportunities:
First, the skills/flows, components and content were not reusable between Agent and Customer experiences. And No one had entertained the vast re-use aspects shared between agents and customers and how to map out those overlaps.
Secondly, designers and content strategists had not considered the efficiency of an agent when using a conversational model. Typing or deep clicking via a conversational flow is much less efficient or intuitive for an agent.
In the Knowledge Integration project below, these considerations were addressed and moved our conversation design platform towards a user-focused, scalable and intelligence-based approach.
Categories
Design Thinking
Systems Thinking
Information Architecture
Content Strategy
Conversational Design
UX Process
UX Wires
UX Design
User Testing
UX Process
For our conversation platform, a clear vision, strategy and UX process set the stage for a user-centered, scalable, systems thinking approach.
Understanding What Our Agents Need
To get a better understanding, I engaged our Business Intelligence group and Customer Support/Operations team to identify top problems agents were encountering with Voyager Next. I also traveled to El Salvador to interview and double jack with agents to get their first-hand perspectives.
Defining A Vision
First, our Vision (the why) as inspired from our company-wide vision. Second, Principles or Themes in support of the Vision. Third, the Guidelines/Examples in support of Principles/Themes. And lastly, Strategic Priorities to get us there.
Project Definition
So as I mentioned before, Voyager Next as a conversational agent tool was still in its early stages when I joined the team.
For this project-and several others, we were retiring an existing traditional desktop app and moving it within the Voyager Next conversational tool.
Knowledge base is a key component to resolving customer issues since it provides deep information on fare travel rules and restrictions, for example. Additionally, the following past challenges with Knowledge base need to be addressed: 1. No Integration with Voyager Next. 2. No Context. 3. No Intelligence. 4. Outdated Information Architecture. 5. Content Too Verbose. 6. Lacks Conversational Tone / Reuse for customer.
Ultimately, we are saving the company millions in licensing fees and operations efficiencies when we introduce help topics in a richer, more contextual experience.
Competitive Analysis
We looked at several knowledge base solutions and highlighted our findings: Zendesk, Salesforce and ServiceNow.
Conversation Challenges
The development team wanted to provide a strictly text-based solution since that is the fastest way to deliver a solution; however, from our research and looking at a long-term solution, there are many challenges to consider:
- Inflexible and arbitrary solution for when/where Agents access help topics.
- Requires programming and maintenance for each help topic insertion for each skill/flow.
- Requires agent to work on a serial skill flow rather than multiple and/or parallel skills.
- Embedded help topics, forms, emails, etc. make conversations longer to scroll and more difficult to navigate.
Conversation Opportunities
Rather than a text-only approach that requires typing and reading through a scrolling pane for knowledge answers, we introduced a one-click intelligent response with a modal that includes the following advantages:
- Let an agent decide when/where to access help topics, separately and simultaneously from primary skills/intents.
- No maintenance to hardcode help topics within various skills/conversation flow(s).
- Machine learning from agent interactions provide future intelligence for inline context suggestions.
- Use common UI/UX standards for common search and/or form interactions.
Tested Design Concepts
As mentioned earlier, all parts of the organization were geared towards speed over quality, and a purely conversational approach would be the fastest to implement. Anticipating push-back from engineering, I wanted to make sure that we tested designs with our agents that spanned, 1) purely conversational, 2) traditional app-like interaction…and 3) perhaps a hybrid of the two.
Final Design
Based on all of the previously tested designs and feedback, a final design was created using the combined benefits from Concept S and B.
This design illustrates omni-present entry points for non-contextual knowledge access.
Principle
~Clear, Fast and Easy
Guideline
~Shortcut to common actions
Once entry point is clicked, pop-out Help modal with level of available context displayed.
Principles
~Clear, Fast and Easy
~Empowered
Guidelines
~Use UI/UX standards & common interaction components
~Better defaults & recommendations
~Guidance & help in context
At article level, provide tab navigation to browse category list and provide inline navigation access to sub-topics.
Principle
~Clear, Fast and Easy
Guideline
~Use UI/UX standards & common interaction components
Lessons Learned
Through user testing and data analysis, moved leadership away from pure conversation model and created modular and multi-modality experiences for both knowledge base and history index projects to improve Voyager Next adoption and decrease agent time on intent resolution.
Overall satisfaction
Improved Adoption
Decreased Agent Time
Scalable & Consistent Interface
Influenced leadership to consider redesign of Voyager Next layout with 3rd pane with interaction model that is better equipped for multi-stepped processes, browsable/searchable content and frequently referenced materials (Next Step: Test Concept with Agents).
Design System & Research Hub
Lead diverse group of UX Designers to create robust documentation site to improve component reuse between customer facing and agent facing applications as well as partner acclimation/learning for conversation platform. Plus, built out a Research hub with current tests/results/summaries and a future research roadmap that led to acquisition of dedicated team researcher.
