A User-Centric Approach to Working with AI

By Dar Min & Sydney Choi

We’re constantly thinking about user experience, and our AI work is no exception. As we explore how to apply advances in AI to our tech stack — and in healthcare more broadly — we continue to evaluate potential applications with our users.

Here is a look at how user research helped inform our experimentation in two areas: automated messaging with members and clinical documentation for primary care providers.

Study highlight: AI-powered automated member support

Automated assistants, or chatbots, have gotten a bad rap for trapping customers in unnecessary loops or blocking access to human help when necessary. Despite their reputation, they are seen as useful by both companies and consumers, particularly when it comes to improving response time.

We explored how integrating an automated assistant into our messaging platform might help members get answers more quickly. Making sure we kept the high bar for our customer service experience, we ran a mixed-method research study with our members to help us test and evaluate our direction.

Our study kicked off with a set of interviews intended to gauge general sentiment around AI-supported chatbots in past customer service interactions. Next, we usability tested a prototype of our messaging experience that mimicked the behavior of an AI-powered chatbot, including realistic conversational flows, interaction patterns, and even delays.

Participant reactions varied, but one thing was clear: it didn’t matter so much if a response appeared automated or written by a person. What mattered most was if the response was quick, accurate, and trustworthy.

While participants expected chatbots to be able to handle simple questions such as “What does the word ‘co-insurance’ mean?”, they preferred talking to a real person about more critical questions, like “How will my benefits apply to this procedure?” and “what will my out-of-pocket cost be?”

Participants also shared stories about getting stuck in unhelpful back and forth with the chatbots, but they did value certain affordances offered that improve their customer support journey, such as suggestions chips and real-time responses.

This user research helped us understand what members were looking for in their interactions with Oscar. We leveraged these learnings to launch the first iteration of our chatbot for members earlier in the year, and we’ll continue to come back to our insights to inform future iterations.

Study highlight: AI-supported clinical documentation

We’re also experimenting with the most effective ways to reduce appointment documentation time for Oscar Medical Group providers. For background, Oscar Medical Group (OMG) is our team of 120+ providers, who offer virtual and urgent primary care for our members. OMG operates on top of our in-house technology stack, including our internally-built Electronic Health Record (EHR) system.

We started by focusing our experimentation on primary care visits via phone and video. These visits are real-time clinical interactions with patients, so we knew we wanted to test and de-risk potential AI-supported workflows before proceeding with implementation.

We worked with five primary care providers, their respective medical assistants, and two registered nurses (who acted as patients). We set up mock video visits and tested two potential workflows: one that utilized the visit transcript as an input to the AI-generated note and one that relied solely on the provider’s notes. By simulating real patient visits, we were able to gauge provider sentiment with the workflows during the visit and understand the potential impact to the provider’s daily routines.

We also sought to determine what level of latency would be acceptable to providers in a workflow that relies on visit transcription. Did it need to be real-time, or would some amount of delay fit naturally into post visit note finalization? We quickly understood through our mock visits that a delay longer than a few minutes would be disruptive. Providers needed to move on to their next visits and could not reasonably be expected to remember visit details when they later circled back.

We also learned that striking the right balance between AI and provider written documentation is key. We noticed that if the AI-drafted note deviated too far from what the provider wrote during the visit, providers then vetted the draft with extra rigor, which ultimately added to overall documentation time.

Through this study, we were able to understand the impact of this potential change without high engineering lift or major disruption to providers’ workflows. We will build upon these learnings to continue iterating on AI-driven efficiency improvements that unburden providers while keeping patient safety top of mind and delivering the best patient experience possible.

Conclusion

By acknowledging the skepticism many have to AI powered experiences, we can take steps to design the right solutions, and help our users build confidence and comfort with new technologies. We’ve learned that when it comes to AI work at Oscar, realistic simulations are a valuable part of our research repertoire, informing the work that needs to happen to maintain the trust we have with our users. By enabling members or providers to react to the speed, tone, and accuracy of real AI-generated content, we’re able to catch pain points early, address them, and launch AI-powered features that are truly valuable — for our members, doctors, and beyond.