Designing an intelligent assistant for Freedom Mortgage.
How can generative AI transform compliance-heavy financial workflows? I led research and design for an AI assistant that cut the time loan agents spend on documentation-heavy questions by 82% — and saved every agent 2.3 hours a week.
Hours lost to repetitive questions.
Freedom Mortgage loan agents spent hours answering repetitive questions and digging through documentation — slowing customer service and frustrating agents and borrowers alike. The goal was an AI-powered assistant that could automate FAQs, cut paperwork, and free agents for higher-value work. My role was to lead user research, then translate the findings into prototypes and final designs.
Shadowing agents on the call-center floor.
We ran field research at a call center, shadowing loan agents through their daily workflow. Two core personas emerged — the Loan Advisor, focused on answering borrower questions quickly and accurately, and the Manager, focused on performance, efficiency, and compliance. Three findings shaped everything that followed:
Workflow breakdowns cause escalations. Agents follow rigid click-paths and stop entirely when the flow breaks — even minor errors escalate to managers, causing delays.
Tool fragmentation wastes time. Four-plus disconnected systems force agents to toggle between tools instead of serving borrowers.
High cognitive load and compliance risk. Poor system feedback and compliance gaps raise agent stress, reduce borrower trust, and expose the business to risk.
Jordan and Rachel.
From the research I built two personas. Jordan, the Loan Advisor, follows rigid workflows and escalates the moment the process breaks. Rachel, the Manager, is overwhelmed by repeat issues and constant troubleshooting. Both told us the same thing: the assistant had to reduce breakdowns, not add one more tool to toggle between.
What the FinTech leaders got right.
We analyzed Chase, Rocket Mortgage, and Intercom — studying their AI solutions for interaction models, animation, and tone of voice, and used ChatGPT to synthesize patterns across them. That gave us concrete recommendations for button placement, chatbot copy, and interaction style that shaped the first prototype.
Sketch first, then a system-locked prototype.
I started with low-fidelity sketches to explore ideas fast and align the team — testing multiple layouts in minutes, iterating cheaply, and avoiding costly mistakes in code. As I remind teams, it’s orders of magnitude cheaper to fix a sketch than to fix code. Using our design system and the competitive recommendations, I built the first interactive Figma prototype for early user testing and stakeholder walkthroughs.
Users wanted to type, not tap.
The first round of testing surfaced a clear signal: users disliked buttons in chat and overwhelmingly preferred natural-language input. For both usability and MVP scope, we pivoted to an NLP-only assistant — simpler to build, and aligned with what people already expect from AI. The final prototype folded in feedback from testing, marketing, and branding, and I worked with developers on a markup style guide so the build matched design intent.
Less time on paperwork, more time with borrowers.
Reducing workflow breakdowns cut unnecessary escalations, integrating tools saved agents hours each week, and clearer feedback loops lowered compliance risk — faster, more accurate loan processing that built borrower trust.