We’ve spent the past year pushing ourselves to build first-class AI workflows for supply chain. The deeper we deployed our agents into real operations, the more we learned how difficult applying AI to supply chain really is.
Frontier models have never been more capable. The challenge is now turning that capability into AI systems that understand organizations, economic networks, and the physical world well enough to reliably participate in complex supply chain operations. The industry’s current practices and understanding of AI aren’t far enough along. Getting there takes research.
That’s why we’re launching Pallet Labs.
Much of this work won’t be commercialized directly, but all of it makes the products we build better.
Here’s what we’re exploring:
Context and tribal knowledge
A supply chain organization is, at its core, a specialized knowledge process running on people. Decades of expertise live in operators’ heads, outdated documents, and workarounds no one ever wrote down.
We believe representing that knowledge as a living model of relationships, decisions, and ripple effects will fundamentally change what AI can do for a supply chain, so we’re pursuing questions like:
- How should AI distinguish between written SOPs and the way operators do the work?
- How do small operational disruptions propagate through an entire supply chain?
- How should AI preserve tribal knowledge as companies grow, change, and experienced operators leave?
Capturing that knowledge is the first step toward building AI that can meaningfully participate in an organization.
Applied AI research
Modern AI models are remarkably capable, but they’re not fully understood. In supply chain, a mistaken customs declaration or a misrouted shipment has real consequences.
The limiting factor in how much of an operation you can hand to AI isn’t what models can do, but how well you understand what they’ll do.
The better we understand where models succeed and fail, the better we can match them to the right operational problems. It’s the same judgment you’d apply to any capable collaborator: learn what it’s good at and assign the right tasks.
We’re pursuing questions like:
- What combination of AI capabilities produces the most reliable customs document processing system?
- How should AI learn from exceptions without overfitting to them?
- How should AI resolve ambiguous shipment requests without unnecessary human intervention?
There’s more to getting AI right than accuracy, speed, and cost. It’s whether a system fails safely or silently, and whether it becomes more capable with every shipment it touches.
The economics of freight
Supply chains generate an extraordinary amount of operational data. Manufacturers, carriers, warehouses, brokers, retailers, and consumers form a network bound together by contracts, prices, and incentives. Understanding how that network behaves matters just as much as understanding the transactions flowing through it.
We’re studying both, asking questions like:
- What is the true marginal cost of a load once you account for dwell, missed appointments, and empty backhauls?
- What does coordination between carriers and brokers cost, and who captures the surplus when AI reduces it?
- How flexible is freight capacity during sudden market swings, like port congestion or an oil price spike, and who wins and loses when they hit?
The goal is to understand networks well enough to anticipate disruption in the chain, identify where margin is leaking, and uncover opportunities traditional software can’t see.
Physical intelligence
Robotics and computer vision are making warehouses, loading docks, and distribution centers increasingly observable by intelligent systems in entirely new ways. We want to know what it takes for AI to meaningfully participate in the physical world — to observe what’s happening on the floor, understand it, act on it, and close that loop into end-to-end autonomy.
That leads to questions like:
- How can AI identify delays from video surveillance before they become service failures?
- How should humanoid robots and self-driving forklifts coordinate picking, replenishment, and material movement inside a warehouse?
- What technologies are required to build a truly end-to-end autonomous warehouse?
The road ahead
Pallet Labs is our dedicated effort to study the hardest and most important problems in applying AI to one of the world’s most complex industries. We hope our work advances the field and informs the products we build.
We’re excited to share what we learn along the way.
Authors
