Chewy replaced best-guess reordering with models that adjust to unreliable data.

Replenishment sounds simple: buy enough of each product so shelves don't run out. At Chewy it means coordinating over 100,000 SKUs across more than 20 fulfillment centers, buying from thousands of vendors, and executing over half a million customer orders daily, with demand fluctuating, vendors performing inconsistently and constraints changing.

The temptation is to trust the data and apply business rules. But the data is unreliable and the decisions are deeply interconnected: what you order at one center depends on what you hold and where you place it. Chewy built a science-driven suite of models on a custom engineering platform, supported by continuous analytics monitoring.

The system corrects unreliable data, accounts for supply uncertainty, and honours vendor constraints as it decides purchase orders, safety stock and inventory placement. This is the key move: rather than deny the messiness, the models absorb and correct it before optimising.

Verified through causal analysis, the solution improved inventory placement, reduced split shipments and shortened shipping distances, delivering financial impact in a thin-margin business. It made Chewy a 2026 Franz Edelman Award finalist.

Build models that correct bad data, then optimise.

Correcting bad data at the source beats optimising a decision built on noise.

Interconnected replenishment choices cannot be solved one warehouse at a time.

Fewer split shipments and shorter distances cut cost in a thin-margin category.

Continuous monitoring keeps the models honest as conditions drift.

When the input data is unreliable, spend effort making the model robust to it before trying to optimise the decision on top.

Chewy's replenishment platform was selected as a 2026 Franz Edelman Award finalist, and the case showed how a real-time, error-tolerant model suite could be built at e-commerce scale. The result — better inventory placement with fewer split shipments and shorter distances — is the kind of structural, recurring gain that matters in a category with very small margins.