Air Canada's chatbot invented a bereavement refund policy. A tribunal ruled the airline liable — $812 CAD in damages, and a precedent every e-commerce team should read twice. The principle the tribunal set is simple: you own every word your AI says to a customer. The bot is not a "separate legal entity." If your AI shopping assistant promises free shipping to Hawaii that doesn't exist, that's your liability — one Shopify merchant ate $8,000+ on exactly that. Here's the tension nobody flags in the demos. Shoppers who engage with AI convert at roughly 4x the rate of those who don't (12.3% vs 3.1% in 2026 benchmarks). The upside is real. But the losses are asymmetric: a single hallucinated spec, an invented return policy, an unsafe recommendation — and the gains reverse faster than they accrued. 46% of shoppers already don't trust AI recommendations. 89% verify before they buy. Every fabrication confirms the skeptic. In our research, almost every major e-commerce AI failure — Amazon Rufus citing the wrong Super Bowl city, Klarna walking back its 700-agent replacement, the dialect bias Cornell Tech documented — traces to the same architectural gap: there's no verification layer between the language model and production. No platform vendor sells that layer as middleware — every major AI shopping tool wires the model straight to the shopper, with nothing checking its claims against a ground-truth product knowledge graph. That layer is the work. Grounding product data, verifying claims before they reach the shopper, building the EU AI Act transparency requirements in before they apply in August 2026 — not bolting them on after a tribunal ruling. If you're deploying an AI shopping assistant, save this and ask your next vendor one question: what checks the model's claims before they reach the shopper? #EcommerceAI #AIShoppingAssistant #AIGovernance #RetailTech #AgenticCommerce