In 2020, as COVID disrupted global supply chains overnight, Levi's inventory planners found themselves flying blind — manually adjusting stock allocations across dozens of locations using forecasting models that had been built for a stable world. By the time they spotted a mismatch, the window to respond had already closed.

The Problem

Levi's relied on manual, rule-based inventory allocation models embedded in their SAP system. The models couldn't adapt to sudden demand shifts, which meant planners spent most of their day doing reactive corrections instead of proactive planning. Four problems compounded each other:

• Inaccurate forecasting — static models couldn't account for real-world variability, causing chronic overstock and understock.
• Manual reservation management — every allocation adjustment required manual intervention, slowing response to supply chain disruptions.
• No real-time visibility — planners lacked live dashboards, so issues surfaced only after stock had already moved in the wrong direction.
• Cognitive overload — the volume of daily manual tasks left no capacity for strategic thinking or proactive optimization.

My Role

SAP brought me in as customer engagement and design lead to partner directly with Levi's and design a solution. My mandate: take a complex machine learning capability and turn it into something inventory planners would actually use. I ran discovery, led solution design, drove prototyping, and worked cross-functionally with data scientists, engineers, and Levi's stakeholders — all within a three-month sprint.

What We Did

Research & Discovery

I ran workshops and interviews directly with Levi's inventory planners to map their daily workflows and identify exactly where time was being lost. We analyzed historical sales data and existing SAP allocation logic to understand the gap between model predictions and real-world outcomes. This gave us a clear picture of where ML could step in, and where human judgment needed to stay in the loop.

Solution Design

The core design question was: how do you make machine learning outputs feel trustworthy and actionable to someone who's been doing this work manually for years? We built the system around three principles:

• Predictive allocation engine — a machine learning model trained on historical sales, seasonality, and supply chain signals to dynamically adjust stock distribution.
• Real-time inventory dashboard — a live visualization layer giving planners immediate visibility into stock levels, accuracy metrics, and anomalies.
• Actionable alerts — surfaced recommendations that planners could act on in one click, rather than having to interpret raw ML outputs.

Critically, the system was designed to integrate into Levi's existing SAP infrastructure — not replace it — so planners didn't need to change platforms.

Prototyping & Validation

We built a pre-production prototype in three months and tested it with Levi's planners against real historical data. The prototype included dynamic accuracy metrics, real-time stock views, and a streamlined reservation adjustment flow that replaced multi-step manual processes with single-click actions.

Results

• 30.8% improvement in inventory allocation accuracy — measured against baseline forecasting models using historical data.
• Manual workload significantly reduced — planners shifted from reactive correction to proactive review, reclaiming time for strategic optimization.
• Faster supply chain response — real-time visibility allowed the team to catch and respond to disruptions before they caused downstream stock failures.
• Pre-production prototype delivered in 3 months — stakeholder engagement early in the process meant the final product was ready for deployment alignment without major redesign.
• Strengthened SAP's position as an AI-driven enterprise partner — the project became a reference case for cloud-native, ML-powered supply chain management.

What I Learned

• AI needs to earn trust, not demand it. Planners who'd been doing this work manually for years were skeptical of black-box recommendations. The design had to make the model's reasoning legible — showing why an allocation was suggested, not just what it was.
• Real-time visibility changes behavior more than any algorithm. The live dashboard shifted planners from reactive to proactive almost immediately — seeing the current state of inventory turned out to be more impactful than the predictive model alone.
• Speed of delivery creates momentum. A working prototype in three months gave stakeholders something concrete to react to. It converted skeptics faster than any roadmap presentation could.