Affirm AI chat
Consumer / Mar - Dec 2024
Product vision
Strategy
Interaction design
Design systems
Animation
Content strategy
Company
Affirm
Status
Phase 1 launched
In collaboration with Sarah Min, Ryan Eberle, Katie Taylor, Chris Stucchio, Krishna Guruswamy, Wingchi Wong, Cindy Ryoo, Drew Gill, Wendly Saintil, and Gurkirat Bahl.
Background
Context and Strategy
After a period of explosive growth, Affirm’s agent-based customer service model was too costly and inefficient to scale up.
Users were left with a slow servicing experience. This meant that we were failing to quickly answer simple questions and taking way too long to resolve complex user issues. Not only does this leave a bad impression, but it also can also lead to financial consequences for our users.
Establish an AI-powered customer service chat as our primary contact channel.
With AI chatbots becoming more effective and common practice, we sought to create our own. This allowed user queries to be diverted from our agents, leaving them more time to focus on more complex and time-sensitive user issues. While the initial cost to build and maintain this were substantial, the goal is to create long-term savings by improving operational efficiency.
Phase 1 (released)
MVP informational chatbot
Expanding Affirm's design system
Chat required all new styling tokens and components to be contributed to the design system. The following is a sampling of my design system contributions.
Safeguarded generative AI to identify the most common questions
To start with the lowest hanging fruit, we crafted a simple chatbot that would be able to answer questions based on our pre-existing help content. This could address a lot of common questions around product comprehension, policies, and more.
Our partnership with OpenAI enabled us to interpret user input more effectively. For the MVP of this product, we chose to not incorporate user data into the chat responses, thus only using AI on the back-end (and ensuring users' data security).
A sample conversation
Walkthrough of feedbacK flow
Improving our data model through user feedback
A key mechanism to our success was gathering direct feedback about our responses.
The data collected allowed the team to get very granular when monitoring specific case types.
Outcome: Improved efficiency, costs, and data
With the first chat launch, we saw very high adoption. This resulted in significant deflection from live agents, which paves the way for cost savings at scale and improved resolution time for the more complex queries that require human intervention.
Through our feedback feature, we were also able to identify opportunities to improve user comprehension across the various product teams. Product owners were now able to iterate and test with precision and speed.
Phase 2 (not released)
Integrating user data
New components + interaction design
With the chat connected to a user's database of loans and payments, a user needed to correctly identify which object to ask about. I redesigned the user data modules from across the app to fit within the chat experience to maintain a consistent visual language.
payment selection
Loan selection + Overflow
Walkthrough of a loan servicing issue
Breaking ground on resolving user servicing issues
The integration of a user's loan and payment data is the key step to servicing the more substantial user problems. While I was unfortunately laid off before this was able to be built, the expectation is that this release contributes to significant increases in cost savings and time to resolution.