Silvia Da RosaPROTOTYPE · CONVERSATIONAL FRAUD PROTECTION
Prototype · simulated data Español

Demo controls

María, 74, tries to transfer $45,000 to a new account after receiving a message from an unknown number claiming to be her grandson.

What the agent detects

Risk: low
  • Unusual amount for the customer's history
  • New destination account, no previous transfers
  • Urgent language and emotional pressure
  • Recent contact from an unknown number
  • Matches the "grandparent scam" pattern (family member + urgency + third-party account)

Every signal is explainable and auditable. The agent never decides on its own: it pauses, asks, and verifies. The final decision always belongs to the person or to a human at the bank.

Design principles applied

  • Transparency: the agent always identifies itself as automated; it never poses as a human.
  • Autonomy without lockouts: a reversible 30-minute pause; it never freezes funds or decides for the customer.
  • Human escalation: a "talk to a person" option visible at every step of the conversation.
  • Data minimization: it never asks for passwords, PINs, or documents over chat; if the customer sends them, it discards them and raises an alert.
  • No blame: clear, respectful language that never patronizes or shames the victim.
  • Accessibility: adjustable large text, AA contrast, 48px buttons, button-based replies (no typing required), screen reader compatible.
  • Consent: a family member is notified only if María authorizes it at that moment.