Inventory rebalancing with human-in-the-loop copilots
We built a planning copilot that pairs forecasts with operator judgment for a mid-market retailer.
The retailer had solid demand forecasts but weak follow-through at the store level. Transfers were slow, markdowns were ad hoc, and regional teams trusted their gut over the plan. We built an inventory rebalancing copilot that paired the model’s suggestions with clear levers for operators.
What we shipped
-
Signal map and confidence bands. We mapped the signals that should matter - demand deltas, lead times, promo calendars, and weather - and surfaced confidence bands on every recommendation so planners knew when to trust the model.
-
Two-step recommendations. The copilot first suggested transfers and markdowns with cited evidence, then generated the communication drafts for store managers. Managers could accept, edit, or request new evidence before execution.
-
Fast feedback loop. Accepted and rejected recommendations flowed into a small eval set. We replayed the set weekly after retrains and tracked precision/recall on “high-confidence” vs “needs review” suggestions.
-
Guardrails for margin protection. We added rules for minimum margin, stock-out risk, and store capacity. When a guardrail fired, the UI showed which constraint was hit and offered alternatives (smaller transfer, delayed markdown, or no action).
Results
- Pilot stores saw on-shelf availability lift by 5% while reducing emergency transfers.
- Store managers trusted the system because every recommendation had a “show receipts” panel with sources and last-updated dates.
- The planning team cleared stale suppressions weekly, which kept the guardrails from silently blocking good suggestions.