Out-of-stocks down 32%, slow-moving inventory down 18%, categories per buyer doubled

Replenishment for more than twenty thousand SKUs ran on buyer experience and simple formulas. Promotions, weather and holidays could not be factored in, so frequent stock-outs and end-of-season overstock existed side by side. New product listings and delistings had no data behind them, and stores and head office often disagreed.

The group wanted each buyer to manage more categories without adding headcount, and wanted the reasoning behind every decision on record.

Replenishment and assortment moved from gut judgment to Agent recommendation plus human confirmation, rolled out category by category.

• Replenishment recommendations: store-level orders generated from sales, stock, promotion plans, weather and holidays, with the reasoning attached
• Assortment review: new listings and delistings scored on sales, margin and basket affinity, with supporting evidence
• Buyers see recommendations in their existing system, can edit them and record why, building a decision log
• Weekly review of recommendations against actuals, with the model and rules recalibrated continuously

The forecasting model and the large language model have separate jobs: time-series forecasting produces the numbers, and the LLM explains them, folds in promotions and external factors, and writes the recommendation. Buyer edits flow back as feedback data.