Amazon has spent years automating fulfillment centers, where robots move inventory, assist picking, sort products, and reduce the number of touches required to get an order out the door. Its next major automation challenge may be much closer to the customer.

Amazon is reportedly developing an advanced delivery-station concept known internally as Project Tetromino, designed to automate more of the work that occurs immediately before packages are loaded into delivery vehicles. According to reporting by Business Insider, an internal planning document suggested the concept could process packages at roughly 2.5 times the rate of Amazon’s existing delivery-station design.

The precise rollout plan is far from settled. Amazon told Business Insider that Tetromino remains an early-stage concept and that specific investment figures and timelines contained in the document no longer reflect the company’s current plans. But whether those particular numbers survive is almost beside the point.

The more important story is where Amazon is trying to automate next.

Moving Automation Downstream

Delivery stations occupy a particularly difficult position within parcel networks. They receive packages after fulfillment and middle-mile transportation, sort them into individual delivery routes, sequence and stage those shipments, and prepare them for delivery drivers.

Amazon says its European delivery stations can involve roughly 40 different operational processes as packages move from arrival to final dispatch. The company has committed more than €700 million to delivery-station technology in Europe, including systems for unloading, sorting, scanning, and reducing repetitive manual handling. Amazon has detailed that investment and its delivery-station automation strategy here.

Historically, many of these processes have required considerable manual handling because a delivery station deals with a constantly changing mixture of parcel sizes, destinations, routes, vehicles, and departure times.

That variability makes the operation considerably more difficult to automate than simply moving standardized totes around a fulfillment center.

Tetromino appears aimed directly at this problem.

The reported concept would use AI, robotics, automated storage, and sequencing technology to reduce manual package handling between induction into a delivery station and placement into the correct delivery flow.

That changes the automation question from “Can a robot move this package?” to something much more sophisticated:

Can an automated system determine where thousands of irregular packages should temporarily reside, continuously reorganize them, and release them in exactly the right sequence for hundreds of delivery routes?

That is a logistics orchestration problem as much as a robotics problem.

The Tetris Problem of Parcel Logistics

The Tetromino name is particularly appropriate. A tetromino is one of the geometric shapes used in the game Tetris, and last-mile logistics increasingly resembles a three-dimensional version of the same puzzle.

Packages arrive in different sizes and shapes. They belong to different routes. Routes leave at different times. Vehicle capacity is finite. And the sequence in which packages are loaded matters because the driver ultimately needs to retrieve them efficiently during the route.

One technology reportedly being evaluated in connection with Tetromino comes from logistics automation company Boxbot.

Boxbot’s system combines automated package storage with software that can sort, sequence, and retrieve parcels. The company describes applications specifically for last-mile delivery stations, including vehicle loading, order management, and wave dispatch, and advertises vehicle-loading performance improvements of as much as 10 times for its technology. That figure is Boxbot’s own performance claim rather than an Amazon projection.

The underlying concept is important regardless of the specific vendor.

Traditional parcel sortation answers the question: Where does this package go?

A more advanced system must answer several questions simultaneously: Where should this package be stored right now? When should it be retrieved? In what sequence should it reach the loading area? And how should that sequence change when the operating plan changes?

That requires the physical automation layer and the decision layer to work together.

Throughput Is Only Part of the Economics

The reported 2.5-times throughput target naturally attracts attention, but throughput alone may not be the biggest economic opportunity.

Amazon has repeatedly emphasized lowering its cost to serve by shortening transportation distances, reducing package touches, improving inventory placement, and increasing consolidation. The company continues to invest aggressively in robotics across its fulfillment and delivery operations. Amazon’s recent delivery-station robotics work provides a clear indication of the direction it is pursuing.

Delivery-station automation could extend the same operating philosophy into the final node of the network.

A highly automated station potentially provides several benefits simultaneously: greater throughput from the same building footprint, fewer manual package touches, more consistent route staging, shorter vehicle-loading windows, improved ergonomics, and better utilization of expensive last-mile assets.

The labor implications are more nuanced than simply saying robots will eliminate warehouse jobs. A highly automated facility should require less direct labor per package processed, particularly for repetitive sorting, staging, and handling activities. But Amazon has not publicly confirmed what Tetromino would mean for total staffing levels, and the company continues to describe its robotics strategy as one that complements employees.

Tetromino therefore looks less like an isolated robotics experiment and more like another step in a much longer automation roadmap.

The Industry Is Chasing the Same Problem

Amazon is not alone.

Parcel and transportation companies have increasingly turned their attention toward automating the irregular loading and unloading processes that historically resisted conventional robotics.

FedEx, for example, has worked with Dexterity AI on robots capable of loading trailers filled with packages of different sizes, shapes, materials, and weights. FedEx has described the use of AI-powered robotic systems as part of its broader effort to automate complex material-handling operations.

This represents a broader transition in warehouse automation.

The first generation of large-scale logistics robotics worked best when operations were redesigned around highly controlled environments. Goods-to-person systems, automated storage systems, conveyors, and autonomous mobile robots all benefited from structure.

AI-enabled robotics is increasingly being directed at the opposite problem: bringing automation into environments that cannot easily be standardized.

Parcel handling is one of the clearest examples.

Why Tetromino Matters

Project Tetromino may change substantially before Amazon builds a production facility. The reported capital plan — including an initial $103 million pilot and additional sites — should be treated cautiously because Amazon has explicitly said those figures and the associated roadmap do not represent its current plans.

But the direction is more significant than the timetable.

For much of the past decade, the automation race in e-commerce centered on fulfillment. The next competitive frontier is increasingly the movement between fulfillment and the customer’s door.

If Amazon can combine automated buffering, AI-based sequencing, robotic package handling, and tightly orchestrated vehicle loading, the delivery station becomes something different from today’s labor-intensive sort-and-stage operation.

It becomes an automated execution layer connecting the warehouse directly to the delivery route.

And that could ultimately be much more consequential than simply building another faster warehouse.

The post Amazon’s Project Tetromino Points to the Next Frontier in Last-Mile Automation appeared first on Logistics Viewpoints.

Descartes Systems Group has acquired Tai Software, an AI-powered transportation management system provider focused on freight brokers, for approximately $100 million. The transaction was funded with cash on hand.

The acquisition expands Descartes’ transportation management capabilities while bringing Tai’s transaction, carrier, and shipment-execution data into the Descartes Global Logistics Network. More broadly, it strengthens Descartes’ position in a freight brokerage market where transportation execution, visibility, compliance, carrier management, and automation are becoming increasingly interconnected.

California-headquartered Tai provides a TMS designed to manage the freight brokerage workflow across the shipment lifecycle. Its platform supports truckload, less-than-truckload, drayage, and cross-border transportation and brings together functions including quoting, carrier sourcing, load execution, billing, and customer engagement.

“The acquisition expands our transportation management capabilities for freight brokers and adds valuable transaction, carrier and shipment execution data to the Descartes Global Logistics Network,” said Andrew Wimer, associate general manager of transportation management at Descartes.

A Deeper Position in Freight Brokerage Execution

The strategic significance of the deal extends beyond the addition of another TMS product.

Freight brokers operate in an environment that requires increasingly tight coordination between pricing, carrier sourcing, transportation execution, compliance, visibility, exception management, and financial processes. Historically, many of those capabilities have been supported by separate applications or loosely connected workflows.

Tai gives Descartes a platform positioned directly inside that operational process.

Descartes already has capabilities in areas including carrier onboarding, compliance, fraud prevention, and real-time visibility. Connecting those services more closely with a broker-focused TMS creates the potential for a more integrated freight execution environment.

Descartes CEO Edward J. Ryan highlighted that complementary fit in announcing the transaction.

“Tai complements our strengths in carrier onboarding, compliance, fraud prevention, and real-time visibility,” Ryan said. “By combining our solutions, we see a significant opportunity to help freight brokers navigate change, streamline freight execution, improve operating margins, strengthen customer and carrier relationships, and support digital transformation.”

Those are management’s expected benefits rather than guaranteed outcomes, but the combination reflects a clear direction in transportation technology: the value of a TMS increasingly depends on how effectively it connects with the systems and services surrounding the transportation transaction.

The Data Dimension Matters

The acquisition also reinforces the growing strategic importance of logistics execution data.

A brokerage TMS captures information throughout the freight lifecycle: quotes, carrier interactions, shipment activity, execution events, and financial transactions. Descartes specifically identified the additional transaction, carrier, and shipment-execution data as a benefit of the acquisition.

That data can become increasingly important as transportation platforms incorporate more analytics, optimization, and AI.

AI-enabled freight execution requires more than a model layered on top of an application. Useful automation depends on timely operational context: what shipment is moving, which carrier is involved, what commitments have been made, what exception has occurred, what alternatives are available, and what action can safely be taken.

A TMS embedded directly in brokerage execution can provide much of that context.

This makes Tai valuable not only as an application but also as another operational data source within the broader Descartes network.

Descartes Continues to Build Through Acquisition

Tai is the latest addition to Descartes’ long-running acquisition strategy. The company has completed dozens of logistics technology acquisitions as it expands across transportation management, visibility, compliance, delivery, safety, and related execution capabilities.

The pace has continued in 2026.

Descartes acquired Latin American last-mile logistics technology provider Drivin in July for approximately $30 million upfront, with additional consideration possible based on performance.

In April, the company acquired Pittsburgh-based Idelic, a fleet safety technology provider, for approximately $28 million upfront, also with potential additional performance-based consideration.

Those businesses address different parts of logistics operations, but together they illustrate Descartes’ broader strategy: adding specialized capabilities that can become part of a larger logistics technology network.

Tai adds an especially important piece because the TMS sits at the center of the freight broker’s daily operating workflow.

Why It Matters

The Tai acquisition is another indication that transportation management technology is moving beyond the traditional concept of the TMS as a standalone system of record.

For freight brokers, execution increasingly depends on an interconnected technology environment capable of coordinating carrier selection, compliance, visibility, fraud prevention, pricing, exceptions, documentation, and financial processes.

At the same time, AI and automation are raising expectations for what those systems should do.

The next stage is not simply better visibility into a transportation problem. It is the ability to identify the problem, understand its operational context, recommend or initiate the appropriate response, and connect that decision back into execution.

That makes integration increasingly important.

Tai gives Descartes a stronger position inside the freight brokerage transaction itself, while Descartes gives Tai access to a broader set of transportation services, data, and network capabilities.

The transaction therefore fits a much larger evolution in logistics technology: transportation management is becoming part of a broader digital execution architecture.

For freight brokers evaluating technology, the competitive question is gradually shifting from Which TMS has the most features? to something more consequential:

Which platform can move most effectively from data, to decision, to execution?

Descartes is betting that a more tightly integrated Tai will help it answer that question.

The post Descartes Acquires Tai for $100 Million, Expanding Its Freight Brokerage Technology Stack appeared first on Logistics Viewpoints.

Every organization has decision rights, whether they are formally documented or embedded in daily practice. Managers approve spending at different thresholds, planners can override certain rules, warehouse supervisors can reprioritize work, and executives retain authority over decisions with strategic consequences. Agentic AI means companies will now have to extend that familiar management discipline to machines.

The question follows naturally from the principle of reversibility. Once an organization decides which decisions are low enough in consequence to automate, it still needs to specify exactly what the machine is permitted to do. That means designing machine decision rights rather than treating autonomy as an on-or-off feature.

Decision Rights Need to Become Explicit

Much of human authority depends on institutional knowledge. An experienced transportation manager knows when an expedite can be approved, which customers warrant exceptions, and when a service failure is more expensive than the freight. A machine needs those boundaries expressed in data, rules, objectives, and escalation logic.

This is one reason repeated conversational interaction is not a sufficient operating model. As I argued in Why Repeated Prompting Is a Weak Operating Model for Supply Chain Decision Support, durable operational AI requires structured context and repeatable workflows. Decision rights are part of that structure because the system needs to know not only what it should recommend but what it is authorized to execute.

A Practical Authority Spectrum

Machine authority can be thought of as a spectrum. At the lowest level, an agent observes and identifies relevant events; it can then progress to investigation, recommendation, preparation of an action, execution within guardrails, coordination across multiple systems, and eventually adaptive optimization inside defined limits. The value of this spectrum is that companies can assign different decision classes to different levels.

A shipment-risk agent may be allowed to investigate every exception, prepare a rebooking for routine freight, and execute automatically only when incremental cost is below a threshold and no regulatory or strategic-customer conditions are present. The same agent may escalate a high-value international shipment to a person. Authority becomes contextual rather than uniform.

Trust Should Be Earned by Decision Class

Executives often ask whether employees trust AI, but the question is too broad. A planner may trust an agent to identify inventory risk but not to reallocate supply, or trust it to reallocate ordinary inventory while retaining human approval for constrained strategic products. Trust develops through observed performance on a specific class of decisions.

This is where auditability matters. Every consequential automated decision should record what the agent observed, which data it used, what alternatives it considered, which rule authorized the action, what it executed, and what happened afterward. That record supports governance, but it also creates the learning loop that allows authority to expand when performance justifies it.

The Human Role Moves Up the Decision Stack

Delegating routine decisions does not eliminate management. It changes the work of management toward setting objectives, defining constraints, resolving conflicting goals, reviewing novel cases, and deciding where autonomy should expand or contract. People become responsible for the design and supervision of the decision system rather than personally touching every transaction.

This is consistent with the broader continuous-intelligence operating model that is beginning to emerge. Machines can monitor conditions continuously and act on bounded opportunities, while humans concentrate on ambiguity, strategy, relationships, and high-consequence judgment. The operating model becomes a deliberate allocation of decision work between people and software.

Governance and Architecture Converge

Machine decision rights cannot be separated from the execution architecture. They also depend on the practical foundations described in Five Requirements for Operational AI in Supply Chain Management, because authority is only useful when context, integration, workflow control, reliability, and governance are strong enough to support it. Authority has to be enforced technically through permissions, financial limits, workflow states, system access, and escalation paths. A policy document that says an agent may spend up to $5,000 is meaningless if the architecture cannot reliably enforce the threshold.

Likewise, the coordination premium depends on coherent decision rights. Several agents acting autonomously without a hierarchy of authority can create conflicts faster than people can resolve them. Governance is therefore not a brake on agentic AI; it is the architecture that makes coordinated autonomy possible.

Decision Rights Become a Management Discipline

Supply chain leaders should begin inventorying important decision classes in the same way they inventory processes and applications. They can classify decisions by frequency, value, reversibility, uncertainty, regulatory consequence, customer impact, and required judgment. From there, they can determine which decisions should remain human, which should be machine-assisted, and which can become autonomously executed within guardrails.

This is likely to become a new management discipline because it sits at the intersection of operations, technology, finance, risk, and organizational design. The companies that develop it well will be able to grant machines meaningful authority without surrendering control. That prepares the ground for the larger question behind this entire sequence: what does the supply chain operating model look like after AI becomes an active participant in work?

The post The New Management Discipline: Designing Decision Rights for Machines appeared first on Logistics Viewpoints.