Retail AI succeeds where it touches money mechanically — forecasting, recommendations, catalog operations — and embarrasses brands where it improvises with customers. Automate the shelf, supervise the conversation.
Why: Courts have already held companies to policies their chatbots invented. An improvised answer about a return window is a contract you didn't intend to sign.
Why: Unbounded repricers produce the classic failures: $0.01 listings, four-figure books, and margin-destroying loops with a competitor's bot.
Why: Catalog enrichment is retail's highest-ROI AI use — but a systematic error (wrong material, wrong size conversions) replicated across a category is a returns wave and a trust hit.
Why: Models beat humans on baseline seasonality; humans beat models on discontinuities (viral moments, supply shocks). The blend beats both.
Why: CTR-optimized recommenders converge on redundancy ('you bought a mattress, here are more mattresses'). Margin-and-retention objectives are what the business actually wants.
The reliable returns in retail AI are operational: demand forecasting, catalog enrichment, search relevance, and recommendations. They’re high-volume, mechanical, and tolerant of small errors caught by sampling. The risky surface is conversational: anything that speaks to a customer in your brand’s name can create commitments, and gets screenshot when it fails.
Hence the split discipline above — bulk automation with statistical QA on the operational side, strict grounding and human escalation on the conversational side, and guardrails anywhere an algorithm touches price.