Target's AI Inventory Ledger detects and auto-corrects unknown out-of-stocks across nearly 2,000 stores, recovering lost sales
Target built the Inventory Ledger — a machine learning ensemble that infers inventory inaccuracies from underlying data patterns and automatically corrects them — processing up to 360,000 inventory transactions per second across nearly 2,000 U.S. stores to ensure products are available and replenishment is triggered.
Background
Target's inventory records were systematically inaccurate: physical audits revealed that half of all out-of-stock events were invisible to its systems. The system thought inventory existed when shelves were empty, so replenishment was never triggered and guests encountered empty shelves with no fix in progress. With 100,000+ SKUs per store across nearly 2,000 stores, this was a problem of enormous scale and commercial consequence.
What Was Implemented
- Built the Inventory Ledger: a real-time, event-driven system recording every inventory change for every item at every store location, processing up to 360,000 transactions per second
- Developed an ensemble of thousands of specialized ML models (gradient boosted trees and neural networks) per product category, trained on millions of historical labeled out-of-stock examples
- Built an arbitration engine to select the single best correction signal when multiple models disagree
- Fully automated corrections: when confidence threshold is met, Inventory Ledger is updated and replenishment triggered without human review
- Explored and then retired IoT shelf sensors and camera-based computer vision — too costly at scale; production system relies on ML inference only
Results
Target confirmed substantial sales lift for products returned to availability that would otherwise have remained as unknown out-of-stocks. The primary source does not disclose a specific percentage lift figure. The system operates across "almost every product category" in all stores and processes up to 360,000 transactions per second . Half of all out-of-stocks that were previously invisible are now detected and auto-corrected.
Lessons
- Half of all out-of-stocks can be invisible to standard inventory systems; ML-based inference is necessary to detect them
- Category-specific ML models significantly outperform universal models across a diverse retail assortment
- Decoupling detection methods from the inventory correction system allows parallel testing without risk to core operations
- IoT sensors and cameras were effective in testing but not economically viable at 2,000-store scale; data-only ML inference won on ROI
- Independent measurement teams (separate from development teams) are essential for unbiased outcome measurement