Inside Amazon's 'Tetromino': a delivery station designed so AI tackles the hardest human job

Amazon is designing a near-fully-automated delivery station that it believes could process packages at roughly 2.5x the rate of its current facilities, according to an internal Amazon planning document reviewed by Business Insider and reported by Eugene Kim. Internally codenamed "Tetromino," the concept would use AI and robotics to handle the final stage of a package's journey before it reaches a customer's door: organizing parcels and loading them into delivery vehicles, a step industry consensus still calls among the hardest to automate.

Last-mile delivery is the part of the Amazon network where humans still do the most manual work. Trucks arrive, drivers sort packages by route, and the loading process is slow, error-prone, and labor-heavy. Amazon's existing delivery stations, even with scanners and conveyor belts, still rely on workers walking aisles, picking parcels, and packing vans. Tetromino would replace most of that with robotic arms and AI vision systems, a design Amazon's own document says could lift throughput to about 2.5 packages processed for every one today.

The technology would not all be in-house. The document points to Boxbot, an AI-and-robotics startup, as a likely source. Boxbot's system moves packages onto trays and uses AI to retrieve and sequence them for loading; the company's website claims the approach can load delivery vehicles up to 10x faster than manual methods. That 10x figure is a vendor self-claim, not independently verified, and the kind of marketing number that gets harder to confirm in real warehouses. The dependency on a third-party vendor is itself a tell: Amazon is treating vehicle loading as a buyable problem, not just an in-house engineering one.

The document also sketches a financial path that Amazon itself is now disputing. Per the planning document, Tetromino calls for a $103 million pilot in 2028, five sites in 2029 at roughly $85 million each, and ten more in 2030, with total planned spending north of $530 million by 2029. An Amazon spokesperson told Business Insider the project is an "early-stage concept" and that the specific financial figures and road map are "inaccurate and don't reflect our current plans." The spokesperson also pointed to ongoing €700 million (around $770 million at recent exchange rates) European delivery-station pilots that use machines for unloading, sorting, and scanning as evidence of Amazon's existing automation direction.

Even at a more conservative throughput multiplier, replacing human loaders with robots changes the per-package cost curve at delivery stations in ways that compound across Amazon's network. Delivery stations are a high-volume, low-margin operation, and labor is the largest controllable input. If 2.5x is even approximately right, the math is the story: more packages per station per hour means fewer stations needed for a given volume, fewer workers per station, and faster route turnover for drivers. Amazon on its July earnings call already said it expects to more than double its fleet of robotic arms in 2026, and is separately deploying software that recommends worker reassignments based on real-time demand. The Tetromino leak has been picked up by Seeking Alpha and other re-reporters, who have not added mechanism beyond the Business Insider original.

Competitors are working on the same step. FedEx is expanding AI-powered Dexterity robots for trailer loading. UPS and DHL use robots for unloading. That all three are publicly investing in vehicle loading is the strongest evidence that delivery-station tasks remain the stubborn part of the automation stack, not because no one has tried, but because grasping irregular, varied-sized packages and packing them tightly into a van is mechanically harder than sorting boxes on a conveyor.

If Tetromino reaches even partial deployment, the human roles most at risk are the picker-and-packer jobs at delivery stations, the workers who currently do the organizing and loading that the design would automate. The human jobs that survive, at least under this concept, are upstream sorting, downstream driving, and machine supervision. Amazon's spokesperson did not address the workforce question directly, and the internal document frames the project in cost and throughput terms, not labor terms.

Three signals would move this from concept to procurement. A public Boxbot pilot, or a clearer technology partnership, would be the first. A 2028 pilot site showing up in local permits or earnings commentary would be the second. Amazon using the phrase "delivery station" alongside "robotic arm fleet" in its next 10-K would be the third, and the first corporate-filing signal that the document's ambition has survived internal review. Until one of those lands, Tetromino is a leaked ambition, not a project.