Origence AI Agent: Match Assist
Project overview
CUDL is Origence's platform connecting 1,100 credit unions with auto dealers. Matching is still manual. Dealers compare PDF rate sheets by hand, hedge by submitting to several credit unions at once, and call it shotgunning. Origence built a four-module AI initiative to fix this. I led design on the matching module, Match Assist. The brief I got was a chatbot. Six dealer sessions killed it. I watched one dealer scan four PDFs in ninety seconds, muttering credit union tiers and score cutoffs under his breath, and realized the bottleneck isn't asking, it's seeing. Match Assist is what came out of that. An AI that lives inside the dealer's existing application form, narrows from 18 credit unions down to 4 as fields fill in, and recommends without ever picking for the dealer.
Matching Over Asking
The brief was a chatbot. After six dealer sessions I came back proposing the opposite. One dealer told me, "I would rather see than ask," and when I went back through every recording, manual comparison took four times more time than any chat-style behavior. Asking takes seconds. Comparing is where the ninety-second decisions actually happen. I pitched pivoting the whole module from chatbot to matching. My PM pushed back, fairly. So I went back, timed every action across the recordings, pulled platform data on submissions, and the 4x held up. The data did the convincing for me. Chat got demoted to FAQ. Matching became the core.
Shotgunning Isn't Laziness
Dealers routinely submit applications to multiple credit unions at once. I assumed this was sloppy work until research showed the opposite. Outcomes are genuinely unpredictable. Five credit unions might technically approve the loan, but you don't know which one actually will, or which will process fast enough. Submitting to all of them is rational hedging. That reframe changed the product. Match Assist can't just tell dealers to stop shotgunning. It has to give them enough confidence to bet on one. Information alone isn't the answer. Confidence is.
AI Inside the Form, Not Beside It
I held the AI back on purpose. When the dealer first opens the form, the AI panel shows 18 possible credit unions. As key fields fill in, that number narrows to 4. No recommendation appears before there's enough customer data to actually back one up. Engineering pushed back on recomputing matches on every keystroke and proposed triggering after key fields instead. I took it. The constraint sharpened the design. Confidence arrives when there's enough data to be confident, not before. That ended up being a better principle than the one I started with.
Holding AI Back Until There's Signal
I held the AI back on purpose. When the dealer first opens the form, the AI panel shows 18 possible credit unions. As key fields fill in, that number narrows to 4. No recommendation appears before there's enough customer data to actually back one up. Engineering pushed back on recomputing matches on every keystroke and proposed triggering after key fields instead. I took it. The constraint sharpened the design. Confidence arrives when there's enough data to be confident, not before. That ended up being a better principle than the one I started with.