Sophisticated technology, a legacy product, a team of two
Feedzai is a Portuguese startup operating in enterprise fraud and risk management, powered by machine learning, AI predictive models and analytics across two main use cases: banking transactions and e-commerce fraud. The technology was genuinely sophisticated. The product layer through which analysts accessed it was a legacy Alert Manager that didn't scale and couldn't handle where the business was going.
I joined as the second designer, with a mandate focused on building and leading the practice: design and research. Low UX integration in the product development lifecycle and complex ML workflows meant risk analysts in banks and payment companies were making hundreds of judgment calls a day with real financial and reputational consequences.
Design the product. Establish the practice.
- Design Case Manager from scratch: a new risk and case management system with rich data visualisation and self-service configuration
- Define the design team's vision and grow the team from two people to eight
- Implement a UX and research practice inside the company: from workshops and personas through to recurrent usability testing
- Make ML model output legible and actionable for analysts without a data science background
Meeting the analysts who'd use it
Before a single screen was designed, I led the team in workshops to identify key use cases and the people behind them. Understanding how fraud analysts triage alerts, weigh evidence, and document decisions for regulatory audit was the prerequisite to designing anything useful. That work fed 40+ design interviews with analysts, compliance officers, and risk managers across financial institutions, and converged into three working personas: the manager, the intern, the detective.
The research revealed something counterintuitive: the biggest friction was not alert volume. It was context switching. Analysts were losing time not in the decisions themselves, but in assembling the information needed to make them.
Structure before screens
Case Manager needed a structure the analyst could hold in their head: nine sections, from Search and Overview through to Rules and Admin, mapped out before a single screen was drawn. Paper sketches and mid-fidelity wireframes tested that structure against real flows, from account recovery to self-service rule configuration and its audit trail, before anything reached engineering.
Reporting was its own problem alongside it. Compliance leadership needed to see patterns no individual analyst could: aggregated risk signals, approval rates, team performance, in a form that didn't require a data science background to read.
From legacy alert queue to decision workspace
With the structure validated, high-fidelity screens replaced the legacy Alert Manager one workflow at a time. Each was validated through recurrent usability testing: moderated sessions, SUS evaluations, and attractdiff questionnaires with a predefined internal cohort, giving quantitative and qualitative signal throughout the build, not just at the end.
The interface was built around the analyst's decision process, not around the data model. Information was structured in the order the analyst needed it. Insights translated ML output into visual forms that compliance leadership could read and trust: specific enough to be meaningful, honest enough not to overstate certainty.
Case Manager: evaluative research
Problem. After the formative stage of ideation and exploration, it was paramount to evaluate the product's usability.
Solution. A recurrent set of usability tests and surveys with a predefined internal cohort, evaluating system experience, readability and ease of use: scripted and moderated interviews, task success metrics, SUS evaluations and attractdiff questionnaires.
Usability testing sessions (July 2016 + January 2017)
Three calls that shaped the outcome
Design the AI as a collaborator, not an authority
The model score was visible but never the headline. Analysts saw the evidence first, then the model's assessment. This preserved analyst judgment and improved accuracy. Analysts caught model errors that a more automated interface would have missed.
Audit trail as a design constraint, not an afterthought
Every decision in the product had regulatory implications. Building the audit trail in from the start, not as a reporting add-on, shaped how interactions were structured across the whole product and gave compliance teams the evidence they needed.
Research as a product input, not a UX deliverable
The 40+ interviews weren't background for a presentation. They fed directly into prioritisation decisions, IA choices, and interaction patterns. Establishing that practice early meant the product was grounded in how analysts actually worked, not how we imagined they did.
Scaling the team meant documenting the system too: a UI kit and, working with the technical writing team, a Feedzai Academy training platform kept a design team that grew from two to eight, and the customers who used the product, working from the same reference.
Components, documentation and academy
Products that analysts trusted
In a domain where bad decisions carry real financial consequences, trust was the only measure that mattered. Case Manager and Insights became defining products in Feedzai's commercial expansion.
What this changed
Feedzai was where I first seriously engaged with the question of how to design alongside machine learning, not just in front of it. The challenge of making opaque model outputs legible and actionable without eroding human judgment is one I have carried into every AI-adjacent project since. That question is more relevant now than it was then.
The model score was never the headline. Evidence first, then the model's assessment.