Research
I followed up on customer feedback with qualitative interviews at several banks.
The problem was clear:
Users rarely trusted automated matching suggestions and manually verified each one.
The reason was not automation itself. Users simply couldn’t understand why the system considered two transactions a match.
Understanding the matching logic
Before changing the interface, I wanted to understand what actually makes a good match.
Together with our domain experts, I analyzed hundreds of matched transactions and validated the recurring patterns with banking users.
Three signals consistently mattered most:
- matching transactions usually balance exactly
- shared identifiers are a strong indicator
- matching transactions usually occur within a limited time window
Product model
I translated these findings into a simple decision model based on the three signals.
Instead of treating matching as one opaque confidence score, each signal could be evaluated independently. Hard mismatches could prevent a suggestion, while stronger signals increased its confidence.
Together with engineering, I used this model to refine the existing matching logic and define what the product needed to communicate to users.
Designing for explainability
My goal was not to tell users how confident the system was, but to let them understand why it suggested a match.
I explored both approaches: a single confidence score and an explanation based on the underlying signals.
User testing showed that the score was too abstract. Users preferred seeing the evidence behind the suggestion, especially the matching parts of the transaction reference.
Final product
Showing every signal at once worked for a single suggestion, but became overwhelming across lists of transactions.
I therefore moved the detailed explanation into progressive disclosure.
Users can understand confidence at a glance and expand uncertain matches to inspect the relevant signals. Only the evidence that actually needs attention is highlighted.
This keeps the list efficient while giving users enough information to verify the system whenever they need to.
Impact
After release, production data showed more transactions being matched across banks.
Follow-up interviews confirmed the shift: users were more willing to accept suggestions because they could understand the reasoning behind them instead of manually verifying every match.
The solution didn’t ask users to trust the automation more.
It gave them enough evidence to decide for themselves.