# AI agents fail at complex commerce tasks

> Alibaba tested 13 AI models on real business work and found even the best stumble on multi-step jobs.
- Title: AI agents fail at complex commerce tasks · Pymnts | Shortify
- Summary: Alibaba tested 13 AI models on real business work and found even the best stumble on multi-step jobs. AI agents work well for simple tasks but break down…
- Keywords: alibaba, artificial intelligence, ai agents, commerce, ecommerce, claude opus 5, ai benchmarking, logistics, multi-step workflows, ai limitations, business, finance
- Source: Pymnts — https://pymnts.com/news/artificial-intelligence/2026/commerce-test-shows-where-ai-agents-break-down
- Author: PYMNTS
- Published: 2026-09-18T19:36:53+00:00
- Read time: 1 min
- Topics: alibaba, artificial intelligence, ai agents, commerce, ecommerce, claude opus 5, ai benchmarking, logistics, multi-step workflows, ai limitations, business, finance
## Alibaba tests AI agents on real work
Alibaba.com released CommerceAgentBench, a public testing toolkit that measured how 13 AI models handled 107 real commerce tasks from its 10 million small business users.
## Best model only passed 61.7% of tasks
Claude Opus 5 scored highest at 61.7%, but Alibaba.com President Kuo Zhang called it high enough to be useful yet low enough to be a warning.

> high enough to be useful and low enough to be a warning
## Multi-step jobs consistently broke agents
Agents failed when tasks required tracking information across many steps, like calculating landed costs with multiple variables, resolving disputes with conflicting documents, or booking multi-leg shipping routes.
## Payment fraud hid in email threads
One agent missed a payment anomaly buried in a supplier email thread with 300 messages, showing how agents lose track of details in unstructured data.
## Different harnesses change model scores
Claude Opus 5 passed 61.7% of tasks on Alibaba's Accio harness but only 56.1% on OpenClaw, showing that the software framework matters as much as the model itself.
## Task-level routing beats one-size-fits-all
No single model led across all three harnesses, suggesting that different AI models should handle different types of tasks rather than one general-purpose model doing everything.
## 132 million Americans already use AI shopping
About 132 million U.S. adults have bought retail products with AI's help, with Amazon capturing 59% of these AI-assisted purchases.
## Merchants restrict AI access to pricing
46% of merchants are unwilling to let agents handle pricing, while 42% restrict access to fraud detection, disputes and liability decisions.
## Key takeaway

AI agents work well for simple tasks but break down when they need to track information across many steps, like calculating shipping costs or resolving disputes.