Databricks Genie and Amazon Quick automate replenishment orders

The walkthrough's framing is that forecasting stopped being the hard part. Foundation models can now predict demand across an entire catalog with no per-item tuning, so the bottleneck moved to the distance between a good forecast and an order placed on it. That gap exists because the forecast, supplier availability, and the ordering system each live somewhere different — and in many setups the person who places the order has a login for none of them.

The article's design choices follow from that. Rather than consolidating the systems, the loop coordinates them: the forecast, the supplier feed, and the Supplier Order API stay where they are and are joined on a shared product key only at decision time, so there is no ETL pipeline to break and no sync job to monitor. Because the supplier feed belongs to a system the retailer does not control, the authors present the pattern as one that travels beyond this example.

A second choice is repeatability. Genie regenerates SQL per question, which suits a person exploring data but not a scheduled job that needs the same answer every run — so the shipped Genie definition pins the surge query and states the numeric floor explicitly. The demo result is deliberately small: six surging SKUs, five routine orders, one exception ticket for the SKU the feed under-supplies on purpose, which exercises the human-review path. Whether this holds up unattended likely hinges on how often the surge query drifts and on the demo endpoint's default of accepting unauthenticated writes.