← All case studies

Business analysis · Retail lending · Responsible AI

From requirements to evidence.

Evaluating whether an AI lending accelerator delivers what its users need—and explaining what must improve before wider use.

500applications
analysed
15user stories
evaluated

The initial analysis covered loans and mortgages. The supplied prototype supported personal loans. Improvement targets were modelled; no live bank deployment or realised benefit is claimed.

The challenge

Speed matters. So does the quality of the decision.

An 8.4-day baseline, repeated information requests and uncertain hand-offs created a case for change. My work explored where AI could help while keeping people accountable.

01 / Investigate

Understand the journey

Analysed application data, mapped the current process and considered the customer experience alongside operational delays.

02 / Define

Make needs testable

Reviewed workshop material and recordings, developed user stories and assessed where AI assistance was appropriate.

03 / Design

Turn needs into screens

Created Balsamiq wireframes for applicant and underwriter journeys, considering status visibility, support and decision controls.

04 / Evaluate

Check the actual product

Used applicant and underwriter views, compared the build with my designs and traced findings back to requirements in the final report.

Design versus delivery

Follow the journey beyond the submit button.

I checked the implemented experience against the intent of my wireframes, including what happened after an underwriter confirmed a decision.

Applicant

Apply and return

The application included document upload, a sample application and a save-and-resume option. These helped customers understand the journey and complete it at their own pace.

Underwriter

Review and decide

The review showed an AI recommendation against bank criteria, an underwriter name and a required reason when changing the recommendation. Human judgement remained part of the decision.

Follow-through

Keep the outcome visible

A customer status screen and end-of-journey support were present. My design placed support earlier; final customer outcomes, completed-case retrieval and senior referral progress were not demonstrated.

Requirements traceability

A prototype can work and still have gaps.

I assessed 15 user stories as met, partially met or not met. This describes observed prototype coverage, rather than proving production readiness.

4

Met

Delivered the requirement as assessed.

6

Partially met

Present, but incomplete or different from the requirement.

5

Not met

Required functionality was not available in the prototype.

01 / Case continuity

A recorded decision must remain retrievable.

After confirmation, the application disappeared from the review queue and only a short “decision recorded” message remained. I could not find a completed-case search or inspect the final decision history.

Recommendation: retain searchable completed cases with decision, rationale, timestamp and reviewer identity in a persistent audit trail.

02 / Referral ownership

Escalation needs a visible next step.

Senior referral progress and ownership were not visible. An underwriter needs to know who is handling a case and whether it is moving forward.

Recommendation: add a referral tracker showing the assigned senior underwriter, status, waiting time and outcome.

03 / Affordability context

Income alone is an incomplete picture.

I could not find a monthly expenses section. That limits the information available to support an affordability review.

Recommendation: capture expenditure and relevant financial context, validate the inputs and show how they inform the recommendation.

04 / Customer communication

A promise of updates needs an end-to-end test.

The customer screen promised messages at each step, but final approval, decline explanations and senior referral updates were not demonstrated. Human support was available at the end of the journey.

Recommendation: test outcome notifications and status changes, and make support available from the start of the application.

05 / Responsible AI

A recommendation needs evidence and controls.

Bank-criteria checks and mandatory override reasons were useful foundations. Enforced role access, fairness monitoring and final decision audit history could not be verified in the prototype.

Recommendation: validate those controls before wider use, and link document-analysis findings to the evidence supporting each recommendation. Keep a human accountable for the final lending decision.

These are observations from a supplied demonstration prototype. Unverified controls are not assumed to be absent from a future production product. The figures summarise my project assessment; full internal artefacts and applicant data are not published here.

What I learned

Writing a requirement is only the beginning. Following it through the actual user journey is where the gaps become clear.

This project strengthened my requirements elicitation, process mapping, wireframing and critical evaluation. Revisiting evidence also changed my conclusions: I corrected initial findings after discovering save-and-resume, sample applications and the named underwriter.

Next time, I would take structured notes earlier, remove duplicate requirements, map more alternative journeys and plan the work in daily stages. My view of AI is practical: use it to analyse evidence more deeply and explain recommendations clearly, with humans retaining the final say.

Discuss my project ↗