Overview
Check fraud is a rampant problem — banks and credit unions lost up to $24 billion to it through 2023. We tackled it by improving a highly manual alert-investigation process, while laying the foundation for an AI-driven version of it.
I discovered gaps in the check-image comparison flow, and championed an image comparison tool that improved investigation efficiency by 80%.
Problem
Agents monitor alerted transactions and investigate where necessary. A single investigation could take over 2 hours, mostly because information was scattered across tools and required an unnecessary number of clicks to find.
The biggest UX issue was in check comparison — agents comparing a check image against previous images to spot signs of fraud. I focused in on just that portion of the flow.
Needs assessment
- User need: minimize screen transitions during an investigation.
- User need: keep new features inside the core flow so they're actually discoverable.
- Business need: Unit21 didn't offer a dedicated check-fraud solution — there was a clear opportunity to build an end-to-end one.
How are alerts investigated?
Each bank sets its own guidelines for what to investigate — transaction activity, entity relationships, and comparing checks against past checks. To find pain points, I interviewed 10 users across 5 banks, from fraud specialists to VPs of fraud.
The biggest issue: agents constantly jumping between platforms to find the information they needed.
How agents and managers felt
What functionality was actually required?
As part of user research, I conducted user walkthroughs, competitive analysis, and ethnographic research. The following user requirements emerged as a result.
As a fraud analyst, I want to be able to view check images in Unit21.
As a fraud analyst, I need to compare the given check image with previous check images in a single view.
As a fraud analyst, I'd like to be able to zoom into check images, since I'd like to find inconsistencies and analyze signatures as they are usually small.
As a fraud analyst, I would like to enhance check images to identify signs of alterations and correct digital artifacts that occurred during an upload.
As a fraud analyst, I would like the ability to conduct quick analysis on an entity's past transactions.
As a fraud analyst, I want to overlay old checks on the check I am viewing to identify differences in signature, check stock, handwriting, etc.
Automating visual comparison could save 5–10 minutes per investigation — but fraud managers surfaced real risk: a risk-averse industry, the threat of automation bias, and no manual baseline to check AI's work against. Most agents were also 40+ and not ready to trust AI outright.
I designed the manual comparison first, architected so automation could layer on later — trust before efficiency.
Making it discoverable
I explored three approaches and reviewed trade-offs with engineering and product:
Agents compare against recent or flagged images directly from the check view.
Agents filter and drill into transactions, then select matching check images to compare.
A stepper lets agents walk through transactions to pick check images to compare.
.png)
We chose option 2 for flexibility — option 1 wasn't flexible enough, and option 3 required rebuilding a complex transaction table in a modal that would be too costly to implement.
Testing with users
I ran usability tests with 7 users. Two issues stood out: 4 of 5 participants struggled to find the trigger point for comparing images, and scrolling to compare checks broke recognition in favor of recall — forcing people to remember details across screens instead of seeing them side by side.
Testing showed the "flexible" path failed novices. I re-introduced option 1 alongside option 2 — serving both novice and expert users — and switched the comparison UI to side-by-side, favoring recognition over recall.
Final iteration
.png)
Learnings
Designing enterprise SaaS is tricky when the people you talk to aren't always the people who'll use the product. This project taught me to prioritize talking to agents over management — including finding them through unconventional means like LinkedIn outreach — while still balancing user needs against business goals.