Hand-coded HTML · Software simulation
Fraud Signals: AI Alert Triage
A software simulation of a bank's fraud-alert queue, where analysts read AI risk signals, add customer context and choose a proportionate action.
The challenge
AI fraud models flag far more payments than are actually fraud. New analysts tend to either trust the score blindly, blocking genuine customers, or wave through real fraud when the score is low. They needed practice reading the signals, not another policy deck.
My approach
- A realistic alert workspace: queue, risk gauge, signal bars with explanations, and a customer profile.
- A 3-minute briefing that can be read aloud by the browser, as an audio read-out for learners who prefer to listen.
- Two coached practice alerts, then a scored shift of 10 alerts covering card testing, account takeover, mule accounts, scams and false positives.
- Every alert asks the same three questions, so the shift report can break performance down by skill.
Objectives and how each is assessed
Every assessment item was written before the content, and maps to exactly one objective.
| Learning objective | Assessment |
|---|---|
| Interpret the signals on an AI fraud alert. | For each of 10 alerts: name the pattern the signals point to (10 points) |
| Use customer context to separate false positives from real risk. | For each alert: pick the customer fact that matters most (10 points) |
| Choose the proportionate action: clear, hold and verify, or escalate. | For each alert: choose the action; pass mark 24 of 30 across all three |
Results and reflection
Add pilot results, reviewer quotes or what you would do differently next time.
Design artefacts
- Signal guide (printable job aid)
- Storyboard (available on request)