Square
Risk investigation tool
Overview
At Square, internal tools are a critical part in delivering the best Seller (how Square refers to users) experience, especially for the Risk team. Our internal tools are what risk analysts use to review suspicious accounts, apply risk controls, and help resolve issues for our sellers. Even though these internal tools are so critical, as Square has grown, the investment has not kept pace. On average risk analysts depend on about seven different tools to complete their daily tasks and many of these tools have years of bloat as different teams have built into it with no central guidelines to follow.
My role
After working on a few more consumer focused products at Square, I decided to be part of a newly spun up team focused on improving the internal tools used by Risk analysts. I joined as the lead product designer and the rest of the team was made up of a service designer, product managers, engineers, operations, and machine learning data science (MLDS) team members.
Discovery
The service designer and I did a deep dive into what the experience was like for analysts and looked to identify how we could make the biggest impact quickly.
As part of the early explorations, I ideated a few concepts around how we could better organize the tool into these modular components called work surfaces. These work surfaces would provide dedicated areas for focused tasks, to help better streamline the overwhelming amounts of information analysts were currently exposed to.
Our discovery research revealed that analysts use multiple tools primarily because each tool contained unique data and importantly analytical capabilities. Based on this, we explored seeing if we could use LLM powered experiences to help centralize this into one tool.
Design sprint
I helped lead a one week design sprint to land the team on what we will prioritize working on. The end goal of this sprint was to present our team's roadmap and early concepts to the Square leadership team to secure buy-in and continued funding for the team.
At the end of the sprint, we intentionally presented lower fidelity designs to help focus the conversation around the concepts rather than UI details. The concepts we presented focused on using LLMs to improve efficiency, by allowing multiple analysts to work a single case by providing case summaries, offer recommended actions trained from policy documentation, and direct feedback loops where risk analysts actions directly refine Square’s model detection capabilities.
Based on Square leadership team's feedback, we landed on building our MVP based on the concept of direct feedback loops to improve Square’s risk models.
MVP
We focused the initial version to help improve upon how risk analysts reviewed active models that worked to prevent fraud on Square’s platform.
Risk analysts using this tool review active running models and verify whether flagged accounts are truly fraudulent. This verification process helps determine each model's precision rate, allowing us to optimize how analysts allocate their review time across different precision rates.
Auto actioning
For models with high precision, we eliminated the need for analyst review entirely, allowing flagged accounts to be automatically actioned based on the model's high precision.
Sampling
For models that have medium precision, the machine learning team developed a sampling approach requiring analysts to review only a subset of flagged accounts. This methodology enabled analysts to effectively evaluate thousands of accounts simultaneously while maintaining statistical confidence in the accuracy.
For the sampled accounts requiring analyst review, we leveraged our understanding of why these accounts were flagged and their common characteristics. This allowed us to design a streamlined review experience showing analysts only the most relevant information needed to accurately determine if an account was fraudulent.
Low precision
For models that had low precision, analysts would need to review all the accounts. But to help make their jobs easier, I designed a few UI elements focused on improving their efficiency, including an “action bar” to help analysts create grouping of accounts to enable bulk review and actioning.
I also worked on creating consolidated account views that brought together critical information previously scattered across multiple tools, providing analysts with a comprehensive overview without needing to switch between different interfaces.
Impact
The MVP helped improve the risk analysts efficiency by a drastic amount, going from being able to review 5-10 accounts per session to over 10k+. We are now currently working on scaling this experience to other workflows and bringing in more features like rule based access and custom views.