Lessons About the Fragile Edges of Agentic Commerce: A Retrospective of Anthropic's Project Vend

This IDC Perspective reviews Anthropic's Project Vend, an experiment that tested whether a midtier agentic AI model could run a small vending operation with real inventory and access to digital tools. The AI agent misunderstood customer intent, misjudged pricing, hallucinated financial and physical capabilities, and often lost track of its commercial goals. These failures exposed how fragile agentic AI systems remain when they must apply judgment rather than follow predictable instructions.This document highlights why agentic AI works best in narrow and controlled domains. Commerce requires ongoing interpretation of socio-commercial engagements, which current agentic models cannot reliably navigate. IDC concludes that fully autonomous commerce is still limited, and buyers should evaluate vendor claims carefully to assess how an AI agent behaves when confronted with uncertainty."Agentic AI performs confidently in deterministic environments yet can falter as soon as it encounters fluid human behavior, bad faith actors, and contextual ambiguity. Anthropic's Project Vend illustrates that commerce requires continuous interpretation rather than fixed logic, and today's models are not yet equipped to bridge that gap ... at least not yet," says Heather Hershey, research director, Worldwide Digital Commerce at IDC. "Just remember that, in the context of agentic AI, 'yet' is a very powerful word."