Fulfillment floors don't fail dramatically. They fail through a thousand small mis-allocations — a picker idle while a lane backs up, a supervisor radioing three people to cover one gap. That coordination overhead was the real cost, and it scaled linearly with volume.
The problem
Task allocation on the floor was a human coordination problem solved by a thick supervisory layer. As volume grew to counter wage inflation, that layer grew with it — the opposite of leverage. Each supervisor held a partial, stale picture of the floor, and decisions were reactive.
The approach
Rather than build a better dashboard for supervisors, I reframed allocation as an agentic control loop: software agents that observe floor state, assign the next-best task, and re-sequence continuously as conditions change. The design principle was to automate the coordination, not the judgment — humans stayed in the loop for exceptions and escalations, where their sense actually adds value.
- A live state model of the floor: who is where, what's queued, what's aging.
- Allocation agents that optimize against throughput and SLA constraints.
- An exception path that surfaces only the decisions worth a human's attention.
What shipped
The engine moved allocation from a supervised, reactive activity to an autonomous, anticipatory one. The supervisory layer shrank by 80%, and the roles that remained shifted from coordination to exception handling and continuous improvement.
Cutting the headcount was the easy part. Earning the floor's trust in an invisible allocator was the real work.
Why it connects
The instinct here is the same one from a decade earlier — see Bayesian forecasting for concrete delivery, where the goal was also to predict and pre-position rather than react. The hardware side of this floor draws on the robotics and dynamic-bin-positioning patents. And the sequencing lesson — what to automate first — is the subject of a field note.