Undisclosed national retailer

In large-format retail, inventory accuracy erodes silently across locations — and at sufficient scale, the revenue impact becomes structural rather than episodic. This leading national retailer operated thousands of stores under multiple brands, each generating daily inventory data that increasingly diverged from actual shelf reality. The gap had multiple causes: product expiration, theft, delivery discrepancies, and manual reporting errors all contributed to shrinkage that was difficult to isolate or attribute. Items listed as "in stock" in the system were routinely absent from shelves, creating invisible lost-sale events that only surfaced in aggregate revenue underperformance. Detecting and correcting these discrepancies through manual store audits was cost-prohibitive at this scale, leaving the problem largely unaddressed — and the revenue loss ongoing.

Factored designed and deployed a Probabilistic Perpetual Inventory Alert Program, integrating machine learning and predictive analytics across the retailer's full data estate. The foundation was a set of data pipelines consolidating three sources: daily inventory reports across all brands and store locations, shipment and delivery records, and SKU-level product metadata. ML models trained on historical sales patterns then identified abnormal velocity deviations — items where actual turnover rates were statistically inconsistent with system-reported stock levels, signaling a likely real-world discrepancy. When the models flagged an anomaly, the system generated a prioritized, actionable alert delivered to store devices each morning before customer traffic. Each alert included expected sales velocity, current stock count, and upcoming delivery dates, giving staff the context needed to investigate and resolve issues efficiently. The design prioritized signal quality over volume to drive adoption rather than alert fatigue.

Within six months of deployment, the system recovered $80M in total sales — establishing inventory accuracy as a directly measurable financial outcome rather than an operational metric. Key results:

• 5.77 million intelligent alerts generated across store locations
• 78% production alert accuracy, enabling staff to act on high-confidence signals
• $25M recovered directly from out-of-stock corrections alone
• 85% model accuracy target set, with active refinements underway to align with evolving business methodologies

Beyond the headline numbers, the implementation demonstrated measurable staff adoption: morning-timed, ranked, and contextualized alerts kept the system integrated into existing store workflows rather than operating as a parallel burden.

• SKU-level sales velocity modeling surfaces inventory discrepancies faster and cheaper than manual audits — the signal is already embedded in existing transaction data.
• Alert design is as important as model accuracy: staff adoption depends on prioritization, timing, and context. Volume without ranking creates noise that gets ignored.
• Framing inventory accuracy as a financial metric — with recovered revenue as the KPI — creates executive visibility and sustains investment in model refinement.
• Launching at 78% accuracy with a target of 85% is a viable production posture; iterative improvement post-deployment is expected, not a sign of failure.
• Data pipeline completeness (inventory reports, shipment records, SKU metadata) directly constrains model quality — integration breadth is a prerequisite, not an enhancement.