The Model Was Never the Problem

Russell Allgor, Chief Supply Chain Scientist at Auger and Amazon’s Chief Scientist for 24 years, opened his plenary address at the 2026 INFORMS Annual Meeting in San Francisco with a question: If demand, supply, transportation, and inventory plans each work as designed, why can the end-to-end decision still fail?

Allgor pointed to handoffs as the culprit. When changes fail to reach the next team’s system, he said, people reenter data, call colleagues, and override recommendations they cannot evaluate.

Four Obstacles to Autonomous Decisions

Drawing on qualitative interviews from The Readiness Gap, a Global Supply Chain Institute white paper by Alan Amling and Shawn Bhimani that was based on interviews with practitioners and software providers, Allgor identified four recurring obstacles to reliable autonomous decisions being made.

• Decision rights. A transportation team may choose a cheaper, less reliable route, leaving downstream facilities to carry more safety stock. Someone must own that trade-off.
• Decision context. Planners draw on spreadsheets, conversations, and experience. They often record the outcome while leaving the information and reasoning outside the system.
• Data sufficiency. Allgor suggested testing plausible values for uncertain inputs. If the answer holds up, proceed while improving the data. If it changes materially, find out which inputs matter.
• Authority and write-back. The business defines the rules and what the system can do without approval. An authorized decision takes effect when the system of record accepts it and updates the records.

Allgor presented to two cases to demonstrate how these obstacles can play out in practice. 

Optimized Fulfilment

Allgor began with the hypothetical example of a retailer’s two fulfillment centers, each capable of handling five orders. Suppose, he said, each incoming order goes to the nearest center that still has capacity. Early orders can be fulfilled at one center even though the other could have handled them. Later orders from customers near the full center then have to be shipped from farther away. The example indicates that individually reasonable assignments can undermine the capacity plan.

Amazon faced the same problem across hundreds of its fulfillment centers. Orders were shipped from nearby fulfillment centers until capacity ran out, and then, in his words, from “wherever we still have room.” To address the issue, Amazon divided its U.S. fulfillment network into eight largely self-sufficient regions, matching capacity to expected demand. It changed its order-assignment rules to prioritize fulfillment within each region, so the capacity plan guided individual orders. 

The rollout required coordination among fulfillment centers, sort centers, delivery stations, and software teams. In the first half of 2023, Allgor said, the share of orders fulfilled entirely within their region rose from 62% to 76%. Shipping distance to customers dropped 15%, and middle-mile touchpoints declined 12%. Nearly 16 million driving miles were avoided.

Fanatics Case Study

Open sales-order lines for customized apparel had to be matched to finished goods or selected for specific decoration, Allgor explained. Both options required inventory, and decoration required production capacity as well. In 2025, planners manually created 59% of work orders. The rules processed orders individually by priority, checking facilities in the same sequence and locking in each assignment even if a later order needed the inventory more. 

A new approach considered all open orders together, using a mixed-integer program to find an allocation and an improvement heuristic to refine it. The new approach increased the share of demand allocated by 150 basis points, or 1.5 percentage points.

Even with the improved allocation, though, planners still had to weigh what trade-off the model did not yet capture: whether to ship the full order late or only part of it on time. Sales had authority over the dates and quantities promised to customers. To test whether the data could support these decisions, Allgor said, the team doubled the supplier lead-time buffer from 7 to 14 days. Among items requiring a purchase order, only about 3% more would miss their promised dates. The team judged the data sufficient to proceed.

Sales leaders then delegated some of that authority to the system, setting limits on delivery delays and quantity reductions. The system operated within those limits and sent larger sales changes for approval, explaining what would allow the order to proceed. Fanatics updated its order system to record authorized date or quantity changes before processing the associated work orders and pick tickets. The system optimized roughly one million units each night, with more than 90% of that allocation running without manual intervention.

Looking Ahead

Allgor said he envisions every company having an end-to-end operating plan it could quickly adapt to disruptions or business changes and carry through to execution. A good answer can lose its value if conditions change before it arrives, said Allgor, and coordinating simpler models can help capture essential business rules while keeping the system manageable.