1 of 14

CASE STUDY

Match

Assist.

Designing how an AI should integrate into expert workflows,

without replacing the expert's judgment.

ROLE

Lead Product Designer & Research Lead

TIMELINE

Ongoing

TEAM

1 Designer, 1 PM, 3 Engineers

PLATFORM

CUDL Dealer Portal Desktop Web

01 / 13

2 of 14

"Design a chatbot for auto dealers."

That was the brief.

Research

killed it.

Six dealer sessions later, I came back with a different product.

02 / 13

3 of 14

Turning Point

Session 3 changed everything.

Six dealers in. One scene rewrote the brief.

I don’t want to ask. I want to see.

DEALER · 8 YEARS EXPERIENCE

4

PDFs open simultaneously

90 sec

to scan and decide

more time than any Q&A behavior (across all 6 sessions)

He muttered CU tiers, score cutoffs, ratio thresholds. Out loud. While scanning. The pattern wasn’t one dealer. It was the job.

03 / 13

4 of 14

The setup

CUDL connects 1,100+ credit unions

with auto dealers. Matching is still manual.

Dealers compare PDF rate sheets by hand. Most hedge by submitting to several CUs at once. We call it “shotgunning.”

Scope

The AI initiative has four modules. This case focuses on Module 2.

04 / 13

01

Chat & Virtual Assistant

Re-scoped to FAQ · Supporting

02

Dynamic Rate & Program Matching

Where I spent my time.

PRIMARY FOCUS

03

Deal Structuring & Optimization

Future phase

04

Cross-sell Recommendations

Supporting

5 of 14

Problem statement

How should an AI integrate into the dealer's workflow to reduce manual effort and improve decision quality, without undermining the dealer's expertise?

05 / 13

6 of 14

Research

Understanding how dealers actually work.

Four methods, triangulating two things: how dealers behave in the moment, and what credit unions consider a high-quality submission.

Method 1

Dealer workflow walkthroughs

Remote sessions. Dealers shared their screens and walked through how they currently find and compare rates for a real customer scenario.

N = 6 Screenshare, 60 min

Method 2

Credit union RM interviews

Focused on submission quality, common dealer mistakes, and what makes a good vs. bad application.

N = 3 Semi-structured

Method 3

Platform data analysis

Submission patterns, rejection rates, multi-submission frequency to quantify the "shotgunning" behavior.

90-day window

Method 4

Competitive heuristic evaluation

Evaluated RouteOne, DealerTrack, and other lending platforms to understand existing patterns and identify gaps.

3 platforms 12 flows

06 / 13

7 of 14

Findings, continued

Four insights.

Each with a so what.

01

Shotgunning is rational, not lazy.

Dealers submit to multiple CUs because outcomes are unpredictable. It's risk reduction, but it floods CUs with low-quality apps.

Give dealers confidence, not just information.

02

Two completely different workflows.

Experienced dealers match score to CU in seconds from mental models. New dealers have no mental model. Same task, opposite workflows.

Serve both without forcing either into the other's path.

03

Finding rates isn't the problem. Comparing them is.

Every CU uses different formats, conditions, and update schedules. Standardizing is where the time goes.

The value is making comparison possible, not picking a winner.

04

Dealers treat their judgment as their edge.

Relationships with CUs and rate intuition make dealers feel indispensable. A tool that looks like it de-skills them will be resisted.

AI must augment expertise, not replace it.

07 / 13

8 of 14

Synthesis

The four insights collapse into two non-negotiables. Everything downstream snaps to these.

PRINCIPLE 01

Compare, don’t ask.

Dealers don’t need answers. They need confidence. The product makes structured comparison possible across CUs. Visibility first, AI recommendation second.

from Insights 01 + 03

PRINCIPLE 02

Augment, never replace.

Experts scan and act in seconds. Novices need progressive guidance. AI surfaces what dealers might miss. The final call stays with the dealer.

from Insights 02 + 04

08 / 13

From four insights to

Two design principles

9 of 14

Exploration

Three directions, three interaction models.

The question wasn't "how much AI." It was how does the AI live in the dealer's workflow?

DIRECTION A

Conversational

Dealer describes the customer in chat. AI returns recommendations as messages.

Lowest learning curve. Single interface.

Comparing 5+ lenders in chat is inefficient.

Typing is slower than scanning under pressure.

DIRECTION B

Structured dashboard

Filterable comparison table with AI signals. Chat in a persistent sidebar.

Highest information density. Expert-friendly.

Three interaction zones compete for attention.

Cross-sell banner interrupts decision-making.

DIRECTION C

Workflow-embedded

SELECTED

AI embeds into the existing application flow. Matching updates as key fields are filled. Chat as a floating overlay.

Zero learning curve. AI appears where dealers already work.

Progressive disclosure. Info arrives as it becomes relevant.

Cross-sell appears post-decision, not during.

09 / 13

10 of 14

Mid-Fi · Core flow

AI arrives when there’s enough data to be useful.

As the dealer fills in key fields, the AI panel narrows from 18 possible credit unions down to 4 specific matches. No AI before there's signal. No delay after.

Fig. 12a · Early state. 18 possible matches.

Fig. 12b · Fields complete. 18 → 4 matches.

10 / 13

Footnote — Progressive matching: triggered after key fields, not on every keystroke. Design constraint from engineering.

11 of 14

Mid-Fi · Comparison

Standardize first, recommend second.

All CUs normalized into a consistent card format. Approval likelihood, monthly payment, and fees compared side by side. AI recommendation is badged, never imposed.

Every option stays visible and selectable. The AI explains why it recommends Pacific Federal, but the dealer can choose any lender in two clicks.

11 / 13

12 of 14

Mid-Fi · Post-decision

Cross-sell, only after the decision is made.

Submission confirmation. Cross-sell appears here once, after the lender has been chosen. Not in the decision flow.

Conversation 3 · where this placement came from

Eng wanted cross-sell at the comparison screen, where the slot already existed. I wanted it post-submission, because research said mid-decision promotion erodes trust.

We didn't settle it in a meeting. We agreed v1 ships his version, v2 tests mine, and dealer behavior decides.

12 / 13

13 of 14

Validation

How we'll know if this works.

v1 hasn't shipped. Here’s the metrics framework we'll use once it does.

POST-LAUNCH METRICS FRAMEWORK

PRIMARY

Multi-submission rate (apps to 3+ CUs)

PRIMARY

Funding success rate

PRIMARY

Time-to-submit

SECONDARY

AI recommendation acceptance

40 to 60%

SECONDARY

Cross-sell engagement

QUALITATIVE

Dealer sentiment. Trust in AI, via interviews and CSAT.

Interviews +CSAT

13 / 13

14 of 14

Thank you

The best AI tool is one

the user doesn't have to

go looking for.

It arrives when it's useful.

It explains when it recommends.

It steps back when the expert already knows what to do.

DESIGNER

Kevin

CASE

Match Assist

YEAR

2026