A tax rate changes. Nobody notices until the money is due.

The challenge

Changes to tax rates, rules and configurations can affect transactions long before a tax team discovers the impact during reporting, reconciliation or an audit. By then, the problem may have affected many transactions and created an unexpected financial exposure.

The data team had demonstrated that an AI model could detect relevant anomalies. The design challenge was turning those detections into timely, meaningful alerts. People needed to define what to monitor, recognize when an issue required attention, and understand its severity and evidence well enough to decide what to do next.

My contribution

I led design in an 8-person team of data scientists, developers and analysts. The team had shown that AI could detect potential tax problems. My role was to connect that capability to the decisions people needed to make: what to monitor, which findings deserved attention and how to respond.

The information model came first. I defined a shared vocabulary for the condition being monitored, the alert a person creates, where it applies and how they are notified. This became a simple setup sequence: what should the system watch, where, and who should it tell? A live, plain-language summary let people check their choices. We chose to keep each alert focused on one concern, trading some flexibility for a clearer meaning when it fired.

Five parts, one structure. That model connected alert setup, the dashboard, activity history, notifications and detailed findings. The dashboard showed what was being monitored and what needed attention; the activity feed gave people a record they could revisit. I prototyped both a full-page history and a drawer beside the dashboard, so the team could see the trade-off between more room for detail and staying in context.

An alert's risk is not a finding's severity. Communicating urgency meant separating two things that were easy to blur: the general risk of an alert type, and the severity of a particular finding, which depended on transaction volume and the money potentially at risk. Notifications carried that distinction into email and Teams, where people might first meet an issue. Labels and action language set critical issues apart from informational updates without relying on color alone.

Explanations people could check. The detail view answered three questions: what happened, why, and what someone could do next. AI explanations were clearly labeled and linked to the relevant rules, dates, transaction details and configuration changes, so people had evidence to inspect before accepting an explanation or acting on it.

Resolution stayed under human control. The design offered an AI-assisted path with reviewable progress, and a manual option available throughout. We developed it far enough to test the approach and left the deeper resolution workflow for future exploration. I brought the connected experience into a Figma Make prototype, working within the existing component library to keep it practical to build.

Methods

• Information modeling and shared vocabulary.
• Alert configuration with a live, plain-language summary.
• Dashboard, activity-feed and notification design.
• Severity classification by alert type and individual detection.
• Explanation and content design grounded in data-science outputs.
• Interactive prototyping in Figma Make and validation through a proof of concept using real EU VAT data.