A tribunal ruled Air Canada had to honor a refund policy its chatbot completely made up. That should worry every retailer rolling out an AI shopping assistant. Your customers don't separate "the bot" from "the brand." When the AI invents a spec, an allergen claim, or a return policy, the brand absorbs all of it — legally and commercially. Shoppers who engage with AI convert at 4x the rate of those who don't — 12.3% vs 3.1%. But errors reverse those gains asymmetrically. One Shopify merchant lost over $8,000 when its assistant invented free shipping to Hawaii. Amazon's own Rufus told shoppers the Super Bowl was in the wrong city — and walked people through building a Molotov cocktail through ordinary product queries, no jailbreak. None of this is because the models are "dumb." It's structural. Most e-commerce AI has no verification layer between model and production — no secondary check against a ground-truth catalog. Retrieval pulls conflicting sources, the model picks the confident-sounding one, and nothing checks it. And the catalog is usually part of the problem: PIM systems run 30-40% attribute completeness on long-tail products, so the model fills gaps the data never gave it. The fix isn't a better prompt. It's the verification and data-grounding layer we build before any answer reaches a shopper. If you've deployed an AI assistant: what's the worst thing you've caught it inventing — and did anything stop it before the customer saw it? #EcommerceAI #AIReliability