Toward a Detection Framework for Agentic Commerce Fraud: A Research Scope and Taxonomy Proposal

AI shopping agents are increasingly browsing, comparing, and completing purchases on behalf of human users, shifting a meaningful share of e-commerce transactions away from traditional human-initiated behavior. This change undermines core assumptions built into existing fraud detection systems, which rely on signals such as browsing speed, session length, device fingerprints, and purchase velocity to separate legitimate activity from fraud. Agent-initiated transactions often lack these signals entirely, or produce them in patterns that legacy models misclassify as suspicious, while genuinely fraudulent agent activity can appear clean and low-risk by the same measures. This work proposes a structured taxonomy distinguishing agent-driven purchase behavior from human behavior, identifies specific points at which current fraud detection heuristics fail against this taxonomy, and outlines a detection framework incorporating agent identity verification, mandate-based authorization metadata, and behavioral baselining. The framework is evaluated directionally using synthetic and public transaction data rather than proprietary industry data. This document represents an early-stage research scope and taxonomy proposal, intended as a foundation for a full technical paper.