Explaining loan denials cut support calls by 32%

Company

LendingClub is a U.S. digital lending platform and marketplace bank with over 5M members, offering personal loans, auto refinancing, small business, medical, and education financing alongside full-service checking and savings.

Team

Senior Product Designer (Me)
2x Product Designers (Mid-Level)
2x UX Researchers
Product Manager
2x AI/ML Engineers
Compliance/Legal Advisor

Skills

XAI UX (Explainable AI)
Service Design
Data Ethics
Regulatory/Compliance UX
Design Systems

Results

−32%

[[bld]]"Why was I denied?" support contacts. [[/bld]] After launch (12-week pilot), inbound complaints dropped, showing the explanations reduced confusion.

+57%

[[bld]]Second-chance approvals within 90 days.[[/bld]] Denied users who acted on feedback were more likely to be approved when reapplying.

+29%

[[bld]]Application completion rate.[[/bld]] More applicants finished the flow instead of dropping out mid-process, a strong indicator for increased trust.

−41%

[[bld]]Loan‑officer handle time on escalations.[[/bld]] Clear explanations meant officers spent less time re-justifying decisions, cutting effort on low-value cases.

Challenge

The platform used an AI model to underwrite loans and credit cards. Customers got a simple approved/denied answer with little context. Denied applicants felt the process was a "black box". Trust fell. Legal risk rose because lending laws require specific reasons for adverse actions. Loan officers also struggled to understand or defend model outputs.

The golden question:
How might we design AI credit decisions that are transparent, fair, compliant and helping users act, even when the answer is "no"?

Research

We did mixed‑methods research to learn what "fair and clear" means to real people, because we needed the explanations to teach, not just tell.

Data & operations review

9 months of call reasons, denial emails, and dispute tickets. [[bld]]Top patterns:[[/bld]] Confusing wording, no "next steps", and uncertainty about credit‑pull type.

User interviews

28 applicants (approved, declined, and "approved with conditions"). We tested early formats: bullet lists vs. charts vs. narratives. [[bld]]People preferred 3–5 ranked reasons with brief plain text and one clear action per reason.[[/bld]]

Internal shadowing

10 loan‑officer sessions to see where they struggle to explain the AI results. Officers relied on ad-hoc language, showing the need for shared explanations that work for staff and customers alike.

Compliance input

Turned [[bld]]regulatory must‑haves into UX checks[[/bld]] (e.g., soft vs. hard pull clarity; specific adverse‑action reasons; data rights). This reframed compliance as a design tool, making fairness and trust part of the product instead of an afterthought.

Workshop

I ran a two-day alignment workshop to bring legal, risk, engineering, support, and design into the same room. Each team looked at the problem from a different angle, so involving everyone was key. In a highly regulated industry, this kind of alignment is essential. Design choices can affect not only user trust, but also compliance with strict laws and the platform’s risk exposure. By working together, we built a shared picture of what "safe transparency" should mean in practice. [[bld]]What we did[[/bld]] [[bld]]1. Journey mapping:[[/bld]] This surfaced where transparency breaks down today, and where applicants, officers, or regulators might get lost. [[bld]]2. Decision principles:[[/bld]] We co-created a set of five guiding rules. These principles became a shared "north star", so every discipline had the same criteria for judging designs. [[lst]] Transparent by default (plain‑language reasons and thresholds) Helpful, not scolding (next steps, not blame) Human agency (easy path to a review) Consistent across products (personal loan → auto → education, etc.) Inclusive and accessible (grade‑8 reading level, WCAG AA) [[/lst]] [[bld]]3. Risk ledger with Legal:[[/bld]] We drafted a mini heatmap of regulatory hot spots. This gave the team a living checklist of legal risks, helping engineers and designers know exactly where extra scrutiny was required. [[bld]] 4. KPI alignment:[[/bld]] We defined success metrics (see "How we measured impact"). Shared KPIs tied design work directly to business and compliance goals, making trade-offs visible and measurable. [[bld]]Outcome[[/bld]] The workshop produced a shared blueprint for "safe transparency". Every discipline left owning the design direction: legal confidence in compliance, risk comfort with fairness checks, engineering sign-off on feasibility, support alignment on messaging, and a design pattern library to guide consistent decisions across loan products.

Design

In redesigning the lending experience, my focus was to create a system that balanced user trust, regulatory compliance, and scalability across loan products. Every design choice was guided by a single principle: make AI-driven decisions clear, consistent, and actionable for both customers and internal teams. From the start, I moved the team’s focus away from polishing single screens and into defining shared rules. We agreed that explanations needed to follow one simple structure: [[bld]]reason, metric, and tip[[/bld]]. This gave us a pattern that worked everywhere, from personal loans to small business loans, and saved us from reinventing the wheel for each product. It also created confidence inside the company, since everyone could rely on the same framework. [[bld]]Clarity on soft vs. hard pulls:[[/bld]] upfront distinctions ease fears of "score hits", reducing drop-off and building confidence.

Each decision had to work for three very different groups: customers, loan officers, and compliance. Customers needed plain words and [[bld]]clear next steps[[/bld]]. Loan officers needed factors they could stand behind when speaking with applicants. Compliance needed every explanation to map back to a valid regulatory reason. My role was to hold all three perspectives together and make sure the design respected them equally. [[bld]]Transparency as trust:[[/bld]] trust grows when people understand why. A short list of top drivers with coaching turns a denial into a learning moment. [[bld]]Human review respected:[[/bld]] lending is high-stakes. AI should augment, not replace people. A clear review path respects that.

Early on, engineers leaned toward exposing raw model scores and legal pushed for generic denial codes. I reframed the problem around fairness. What people needed most was to understand why they got the outcome and what to do next. By showing prototypes and test results, I won support for this approach at the leadership level. That alignment turned explainability from a compliance checkbox into a shared vision: every AI decision should teach, not just judge. [[bld]]One explanation pattern across all products[[/bld]], reducing learning time and ensuring consistency. [[bld]]Copy as the product:[[/bld]] in credit, words are the product. The right sentence can change sentiment and compliance posture.

User Testing

We tested for comprehension, perceived fairness, and actionability.

Method

We conducted moderated sessions with 24 participants across credit tiers, unmoderated tests for copy, and A/B on confirmation screens and "Ask for human review" placement. We used teach-back, where participants explained the decision in their own words, to measure clarity. We performed a reading level check to keep content at grade-8 and improve Flesch reading ease.

Key findings

Users understood the reason for the decision twice as often when using the new report, showing a clear improvement in comprehension. People consistently preferred short and specific reasons over long lists. The option for human review played an important role in reducing both anger and churn intent, since it reassured users that they had a meaningful way to contest the decision if needed.

Iteration

[[bld]]1. Numbers to narrative[[/bld]] We kept numbers when helpful (e.g., "Your DTI is ~44%"), but led with plain insight and what to do. [[bld]]2. Reduced cognitive load[[/bld]] Only the top 3 drivers are visible; a "See all factors" drawer holds the rest. [[bld]]3. Better review path[[/bld]] Moved "Ask for human review" above the fold and confirmed in testing that it didn’t create support noise, it redirected people from chat to the right workflow. [[bld]]4. Reason codes[[/bld]] We aligned UI reasons to adverse‑action code sets so letters and logs are consistent.

How we measured impact

We measured success by looking at how applicants behaved from start to finish. At the beginning we tracked whether people [[bld]]clicked on a loan offer and went on to complete an application[[/bld]] (application completion rate). This told us if the flow was clear enough to encourage follow-through. When people reached a decision, the strongest signal was [[bld]]how many contacted support to ask why they were denied[[/bld]] (support inquiries). If explanations were clear, those calls should disappear. We also looked at [[bld]]how many denied applicants not only tried again within 90 days but were actually approved on their second attempt[[/bld]] (second-chance approvals). That showed whether the coaching element was effective at turning a rejection into a positive outcome. Finally, we measured the impact on internal staff. We tracked how long loan officers [[bld]]spent handling escalations[[/bld]] (handle time). If applicants already understood their decision, officers would no longer need to re-justify it, saving valuable time.