Amazon's New AI Can Reorder Stock Before Shelves Go Empty

Amazon and Databricks just published a step-by-step blueprint for something retailers have struggled with for years: ordering the right amount of stock before it runs out. The problem was never spotting demand — it was the gap between a good prediction and an actual order being placed. Forecasts live in one system, supplier information in another, and the person who clicks "order" often can't log into either. This setup connects them.

The loop runs in four steps. First, Databricks' forecasting tool predicts seven-day demand for every product in a catalog — tens of thousands of items — without anyone tuning it item by item. Second, an AI agent flags only the products surging in demand, defined as selling at least 50% above their recent two-week average. Third, Amazon Quick checks live supplier availability and picks the cheapest supplier who can cover the gap. Fourth, it either places the purchase order automatically or, if no single supplier can handle it, opens a ticket for a human to review.

For a normal person, the payoff shows up at the store or in your delivery box. The fastest-selling items — the ones that vanish first — get reordered the moment demand shifts, instead of days later after someone notices. For retail staff, it means planners stop copying numbers between spreadsheets and start handling only the genuinely tricky cases, where supply is short or unusual.

The catch is that this isn't a product you switch on. It's a technical recipe for engineers, requiring paid Databricks and Amazon accounts, specific permissions, and setup commands run from a terminal. There's also real risk in letting software place orders unattended — a bad prediction or a stale supplier feed could mean ordering too much or too little. The design only escalates to humans when no rule fits, so the rules themselves need careful watching.