Amazon priced Locker space like airline seats, lifting throughput 9%

Amazon Lockers accepted packages first-come-first-served. The result: lockers filled with three-to-five-day packages, and next-day and two-day packages — the ones customers paid a premium for — were rejected at the door. In 2017, 40% of unjustified delivery rejections traced to bad ship-option capacity reservations.

The textbook fix, airline-style yield management, didn't apply: a package's dwell time in a locker is unknown when it's accepted, so you can't price the slot by duration. Amazon's team rebuilt the problem: machine learning predicted demand and dwell-time distributions per locker, and a linear program set capacity reservations for each shipping speed.

The system effectively rationed a scarce physical asset — locker compartments — in favor of fast-shipping customers, without building a single new locker.

Deployed for the 2018 holiday season, it lifted worldwide Locker throughput 9% year over year, affecting millions of customers, and was published in the INFORMS Journal on Applied Analytics in 2024.

Reserve locker space per shipping speed, like seats

It treated locker space as inventory to be allocated, not a queue to be endured.

ML handled the unknown dwell times that broke classic yield-management math.

No new hardware: the gain came purely from a better allocation policy.

First-come-first-served is a capacity allocation policy too — usually the wrong one. Reserving capacity for the traffic that matters most beats adding capacity.

The capacity-reservation system became the standard way Amazon manages Locker inventory and was documented in a 2024 INFORMS Journal on Applied Analytics paper, 'Amazon Locker Capacity Management.'