Nothing in that chain was a bug. The model behaved exactly as designed. The failure was asking a probabilistic reasoner to own a real-time, real-money decision based on data that lied.
This is the exact scenario we break down in The AI Commerce Architecture Guide — along with a simple framework for knowing which AI applications are safe to build (analysis, content, internal tools) and which ones need a system engineered specifically to prevent this kind of cascade (bids, pricing, inventory-aware execution).
If you're experimenting with AI on your Amazon operations, this is the framework to have in hand before you connect anything to production data.
