Underwriting IA - Allstate Virtual Assistant
As UX moved in-house for Allstate’s virtual assistant for agents, we learned that the existing information architecture of knowledge was creating gaps and conflicting user paths. We needed to zoom out and take another look at how we delivered information to the user.
One key problem area was underwriting content. The experience had been designed in silos, which made it difficult to account for the whole picture.
As UX lead, I’m responsible for guiding the strategic approach to information architecture and conversation design. I redesigned parts of the experience to create a more holistic journey, and worked with a user researcher to understand and recommend necessary updates to the roadmap and intent architecture.
In a conversational UI, the experience relies almost exclusively on successful architecture and built-in navigational cues. The user can’t rely on visual guideposts or click to another page if they landed in the wrong place.
The entire team is new to AI, so it’s especially important to learn fast and apply new insights to our strategy.
In the case of underwriting, ongoing user research and prototype testing shaped and re-shaped our approach several times over.
PROCESS, ROUND 1
Our first iteration on the underwriting experience was based on a combination of user research and usability updates. When I tested the flows, I found that the existing structure forced users to answer technical questions that could be addressed through back-end logic. We also learned from our business partners that the designs were clashing, so the system was often unable to figure out which of our two main use cases was appropriate for a user’s question.
I simplified the designs and created an underwriting disambiguator to make it easier for both the system and the user to navigate the conversation. We ran a card sort of Allstate’s lines of business for underwriting, and we learned that users think about the LOBs differently than Allstate does. Our business partners felt strongly that we should present the lines the way they were defined, but the insights we pulled from the card sort helped convince them to try something new.
Process, round 2
Through additional user research, we learned that the existing mental model that informed the IA was incomplete.
The virtual assistant was designed based on delivering two types of documentation for two respective use cases. The research showed us that each documentation type mapped to multiple use cases; plus, there were additional documentation types unaccounted for.
Due to technical constraints, we had to wait until after our pilot release to completely rearchitect the underwriting knowledge. In the meantime, we needed to find a way to accommodate the additional use cases and documentation types.
I collaborated with the intent team to better understand how design might impact the intent model, and vice versa. We adjusted the use cases to focus on tasks, rather than phases in the journey, to provide a more comprehensive system of support.