Explaining AI loan decisions

01 / Define

The problem to explore

A hypothetical digital lender uses an automated model to approve or reject small business loans. Applicants who are rejected receive no clear reason or next step. Explore how to explain decisions honestly, show what can be improved, and offer a human review.

This is an original UXCharan practice scenario, not a commissioned client project or a validated research finding. Treat the problem as a hypothesis. Adapt it when your research points elsewhere.

Who to learn from

Small business owners applying for a loan through a fully digital lender. Recruit consenting adults with relevant experience where possible; document recruitment limits and avoid claiming your sample represents everyone.

02 / Discover

Research before choosing a solution

Questions to investigate

How do applicants react to automated rejections?
What level of explanation is useful without being overwhelming?

Start with a small set of interviews or contextual walkthroughs that fits your access and consent arrangements. Compare two existing approaches to the same task. Separate direct observations, participant statements and your own assumptions. If primary research is not possible, clearly label desk research and list what remains untested.

Map the experience

Trace the task from its trigger through completion and follow-up. Identify where information, responsibility or confidence is missing. Choose one opportunity you can justify with evidence.

03 / Design

A manageable project scope

See the decision and main reasons
understand what could change the outcome
request a human review

Choose one primary device and a realistic context of use. Explore at least two approaches before refining a flow. Build only the screens and states needed to demonstrate the selected task; a complete production system is outside this brief.

Include the difficult moments

Decision partly based on missing data
applicant disagrees with the data
appeal outcome is the same

Document loading, empty, error and success states where relevant. Consider keyboard order, text enlargement, contrast, meaningful labels and alternatives to colour-only status. Use synthetic data and a fictional brand; do not collect real medical records, credentials or payment information.

Project-specific artefact

An explanation screen with plain-language reasons and actionable next steps.

04 / Test & improve

What would make the design work?

Do applicants understand why they were rejected and what they can realistically do next?

Turn this question into realistic prototype tasks. Record whether participants complete each task, where they hesitate, what they misunderstand and how much help they need. Choose success criteria before testing; report observed counts with sample size, rather than claiming statistically proven improvements.

Prioritise findings by impact on the task. Revise the flow and retest the most consequential change where possible. If you cannot retest, explain that limitation.

05 / Tell the story

Your submission checklist

Problem framing, intended users, scope and your role.
Research approach, anonymised evidence and assumptions that changed.
Journey or flow, alternatives explored and reasons for your decisions.
An explanation screen with plain-language reasons and actionable next steps.
Viewable Figma prototype or a shareable document, including key edge cases.
Usability findings, before-and-after revisions and remaining limitations.
A concise portfolio narrative that labels this as a self-initiated concept.

Show what you learned and changed. Do not invent testimonials, interview quotes, launch metrics, employment relationships or business outcomes. Disclose material AI assistance and verify generated content.

Share CS-59 for feedback →

Useful methods for this project

These sources inform the learning process; they do not validate the fictional scenario or endorse UXCharan. Explore more tools in the resource library.