Where does AI actually earn money in retail?
In the operating loops that run daily. Demand forecasting that blends sales history, seasonality, weather and promotions. Replenishment that turns those forecasts into orders and keeps them flowing through the warehouse, where AI in supply chain picks up. Dynamic pricing inside guardrails. Fraud screening on transactions. Each loop is boring alone; together they decide most of a retailer’s margin, which is why they were automated years before chatbots arrived.
What changed with generative AI and agents?
The customer-facing layer caught up. Service agents now resolve order status, returns and product questions end to end instead of deflecting to FAQ pages. Shopping assistance moved from keyword search to conversation, and virtual try-on moved the fitting room into the product page. And buying itself is becoming delegable: when a customer’s own agent discovers, negotiates and purchases, that is agentic commerce, and it changes who owns the customer relationship. A retailer’s content and data now have two audiences: people, and the agents shopping for them.
What does AI in retail look like for a mid-sized retailer?
Less exotic than vendor decks suggest. The accessible wins: a forecasting model on your own sales data, automated replenishment suggestions a buyer approves, a service agent grounded in your order system, and marketing copy generated inside brand rules. Pick the loop with the most expensive current errors, usually stockouts or overstock, and automate that one first against a measured baseline. The phone channel counts too: an AI receptionist answering store questions is retail AI at its simplest.
What goes wrong in retail AI projects?
Data before models: fragmented product data, inconsistent SKUs and untracked promotions quietly cap every model’s accuracy. And automation without bounds: pricing engines need floors, ceilings and competitive rules a human sets. Give every automated retail decision a boundary a merchant chose and a report a merchant reads. The failures are rarely the model being wrong; they are the business not noticing for six weeks.