The supply chain is AI's next big test

If you live near a port, you’ve probably seen the photos. Cargo ships anchored in lines, waiting days or weeks for a berth. The Port of Los Angeles had over 100 container ships queued at one point. Each one carrying thousands of containers. Each container carrying goods that someone ordered weeks ago and is still waiting for.

The supply chain is broken. Not in one place. Everywhere. All at once.

Ships can’t unload because ports are congested. Containers can’t move because there aren’t enough trucks. Trucks can’t run because there aren’t enough drivers. Warehouses can’t process because they can’t hire enough workers. And the goods that do make it through arrive late, incomplete, and at inflated prices.

I’ve been thinking about this for months, and I keep arriving at the same question: we built AI that can write poetry, generate art, play chess at superhuman levels, and fold proteins. Can it optimize a supply chain in crisis?

The optimization problem

A global supply chain is, at its core, a massive combinatorial optimization problem. You have goods at point A. They need to reach point B. Between A and B are ships, ports, trucks, warehouses, rail lines, customs checkpoints, and a thousand variables that change daily.

Traditional supply chain management uses forecasting models, historical data, and experienced humans making judgment calls. It works when things are stable. When the system is within normal parameters, the forecasts are decent, the humans make good calls, and goods flow.

When things are not stable, when a pandemic disrupts labor, a Suez Canal blockage halts trade for a week, a chip shortage cascades into automotive shutdowns, and consumer demand spikes simultaneously, the traditional models break down. The variables are too many, the interactions too complex, the changes too fast for human judgment to keep up.

This is exactly the kind of problem AI should be good at. High dimensionality. Rapid change. Complex interdependencies. Optimization under constraints.

What’s being tried

Flexport is using machine learning to predict port congestion and reroute shipments before they arrive at a backed-up port. Instead of finding out your ship is stuck in a queue after it’s already there, AI models predict the queue weeks in advance and suggest alternative ports.

Project44 offers real-time supply chain visibility using AI to track shipments across modes (ship, truck, rail) and predict delays. Their models process data from millions of shipments to estimate arrival times with better accuracy than traditional tracking.

Freightos is using AI to price freight in real time, adjusting to demand the way airlines adjust ticket prices. The idea is that real-time pricing creates better allocation: if a route is congested, the price goes up, demand shifts to alternative routes, and the system rebalances.

These are real products being used right now, during the crisis. Not research papers. Not demos.

The gap

But here’s the honest part.

The supply chain crisis isn’t getting better. The AI tools are helping at the margins, optimizing individual decisions, improving visibility, catching problems earlier. But the systemic issues, the shortage of truck drivers, the limited port capacity, the chip production bottleneck, are physical constraints that software can’t wave away.

You can optimize the routing of a container from Shanghai to Chicago. You can predict which port will be less congested next Tuesday. You can dynamically price freight to shift demand. But you can’t create a truck driver who doesn’t exist. You can’t build a container crane in software. You can’t manufacture a semiconductor chip faster than the physics of lithography allows.

AI is good at optimization within constraints. The current crisis is a constraint problem, not an optimization problem. We don’t have enough of the physical things we need. No algorithm fixes that.

What I’m watching

I think this is a defining test for AI in industry. Not the fun kind of test (write a poem! generate art!). The hard kind. The kind where the answer isn’t “wow, impressive” but “did it actually help?”

If AI-driven supply chain tools can demonstrably reduce delays, cut costs, and improve throughput during the worst logistics crisis in decades, then the case for AI in operations becomes unassailable.

If they can’t, if the tools make marginal improvements but the crisis resolves only when the physical constraints ease (more drivers, more ships, more chips), then the case for AI in operations becomes more complicated. Useful in good times. Helpful in bad times. But not sufficient.

I’m genuinely uncertain which way this goes. And I’m watching closely.

The ships are still waiting off the coast. The algorithms are running. The answer is somewhere between them.

Related thinking: