Designing an AI assistant for Freedom Mortgage.
Freedom Mortgage loan agents lost time searching documentation and escalating routine questions. I led research and design for an AI assistant that reduced time spent on documentation-heavy questions by 82% and saved each agent 2.3 hours a week.
Routine questions were slowing down the floor.
Loan agents spent hours answering repeat questions, searching documentation, and moving between disconnected systems. When a workflow broke, they often had to stop and ask a manager for help.
The goal was to give agents one place to find reliable answers without adding another tool to manage. I led the user research, then turned what we learned into prototypes and final designs.
I started on the call-center floor.
We shadowed loan agents through their daily work and watched where the process slowed down. Three patterns kept appearing.
Broken workflows caused escalations. Agents followed rigid paths and stopped when something unexpected happened. Even small errors could require a manager.
Disconnected tools wasted time. Agents moved across more than four systems while trying to answer a single borrower question.
Unclear feedback increased risk. Agents had to remember complex rules without enough guidance from the product, which raised stress and created room for compliance mistakes.
The assistant had to prevent breakdowns, not create another one.
I turned the research into two personas. Jordan, a Loan Advisor, needed fast answers without leaving the task at hand. Rachel, a Manager, needed fewer repeat questions and more confidence that agents were following the right process.
Both pointed to the same design requirement: the assistant had to fit the existing workflow and help agents recover when it broke.
I studied how mature assistants handled the details.
We compared assistants from Chase, Rocket Mortgage, and Intercom, looking closely at interaction patterns, motion, and tone of voice. I used ChatGPT to help organize the patterns, then translated them into recommendations for button placement, chatbot copy, and response behavior.
I kept the first ideas cheap to change.
I sketched several layouts before moving into Figma. That let the team compare directions quickly while the structure was still easy to change.
I built the strongest direction with the existing design system, then turned it into an interactive prototype for user testing and stakeholder walkthroughs.
People wanted to type, not tap.
The first test produced a clear signal. People disliked choosing from buttons inside the chat and preferred asking questions in their own words.
We removed the buttons and focused the MVP on natural-language input. I folded feedback from users, marketing, and brand into the final prototype, then worked with developers on a markup guide so the build matched the design.
Less time on paperwork, more time with borrowers.
The assistant gave agents a faster path to reliable answers inside their workflow. That meant less time searching documentation and more time helping borrowers.