Commerce Test Shows Where AI Agents Break Down
· Source: pymnts.com
In short
Alibaba.com released CommerceAgentBench, a public GitHub testing toolkit that graded 13 AI model families on 107 real commerce tasks, with the best model, Claude Opus 5, completing only 61.7% of them. Agents consistently failed on multi-step work such as landed costs, return disputes, payment fraud detection and multi-leg shipping routes, and scores also shifted depending on the harness used. Merchants remain cautious, with 46% unwilling to hand pricing to agents, even as roughly 132 million U.S. adults have already made AI-assisted retail purchases.
Read Original Article — Read the full article on the publisher’s site.
In the issue of week 38, 2026: AI Moves Closer to Supply Chain Decisions, but Data and Returns Still Lag
Tags: agents, logistics, ecommerce, benchmark, fulfillment, automation