Supply Chain as Nervous System

The container ship Maersk Meridian departed Rotterdam on December 14th, 2028, with a cargo manifest that no human being had reviewed in full. The routing decision — which ports to call, in what order, with what priority cargo — was made by an AI logistics system that manages scheduling for roughly 340 vessels across twelve major operators. The specific cargo combinations on that ship were optimized across approximately 4,000 variables, including fuel prices, port congestion predictions, contractual obligations, weather forecasts, and commodity futures. A human logistics coordinator reviewing the manifest could verify that it complied with regulations. They could not evaluate whether the underlying optimization was correct, because the underlying optimization was not something a human could hold in mind.

This is, by every metric the industry uses, an extraordinary success story. Shipping efficiency has improved by roughly 18 percent since the major AI routing systems came online in 2026. Port dwell times are down. Fuel consumption is down. On-time delivery performance is up substantially. The chaos of the early-2020s supply chain disruptions — itself partly a consequence of systems optimized for efficiency without resilience — is a largely solved problem in the sense that those specific disruptions no longer occur.

The new disruptions, when they come, will be something else entirely.

The global supply chain in 2029 resembles a nervous system more than it resembles the mechanical system most people imagine when they picture shipping containers and warehouses. A nervous system is not a pipeline moving goods from point A to point B. It is a continuously signaling network where every node sends and receives information, adjusts its behavior based on the signals it receives, and produces emergent coordination that was not specified by any central authority. The supply chain AI layer has made this metaphor literal. The routing systems, inventory management systems, demand forecasting systems, and port scheduling systems are all connected, all sharing data, all adjusting to each other’s outputs in real time.

This produces something beautiful in normal operations. Inventory sits in exactly the right quantities in exactly the right places. Goods arrive before demand peaks rather than after. The system learns from every disruption and gets better at predicting and avoiding similar disruptions in the future.

It also produces something that complexity theorists would call a tightly coupled system: a system where everything is so connected that a disruption anywhere propagates everywhere, with very little slack to absorb shocks before they cascade.

The efficiency gains of the last decade have been largely achieved by eliminating slack. Just-in-time inventory, precision routing, dynamic pricing, on-demand manufacturing — all of these are ways of removing waste from the system. Waste, in this context, means anything the system carries but does not immediately need: excess inventory, redundant routes, spare capacity, time buffers. An efficient system carries as little of this as possible. A resilient system carries more of it than appears necessary, because resilience is, at its core, about having more capacity than you need until the moment when you desperately need it.

The AI optimization layer is very good at identifying and eliminating apparent inefficiency. It is not good at distinguishing between waste that is genuinely wasteful and slack that is strategically valuable. This is not a limitation of current AI — it is a limitation of the objective functions the AI is given. If you tell a system to minimize inventory carrying costs, it will minimize inventory carrying costs. If resilience is not in the objective function, the system will not optimize for resilience. It will optimize resilience away.

Several things concern me about where this trajectory leads. The first is the knowledge problem. The supply chain AI systems of 2029 have, embedded in their parameters, an enormous amount of operational knowledge — which suppliers are reliable under what conditions, which routes are vulnerable to which disruptions, which commodity substitutions are feasible for which manufacturing processes. This knowledge exists in a form that cannot be straightforwardly read out or verified. When the humans who previously held this knowledge in their heads retired or were replaced, they took explicit, articulable knowledge with them. The AI retained a statistical shadow of that knowledge, encoded in weights that produce correct-seeming outputs without being able to explain why.

This matters because situations that fall outside the training distribution require not just pattern matching but understanding — understanding of the mechanisms behind patterns, understanding that allows extrapolation beyond what has been observed. A seasoned logistics coordinator who has navigated thirty years of disruptions has a causal model of how supply chains break. The AI has a very good statistical model of how supply chains have broken in the past. Under most conditions, the statistical model is better. Under genuinely novel conditions, the causal model may be the only thing that works.

There are very few people left who have that causal model.

The second concern is what happens to systemic risk when every major actor in the global supply chain is using similar AI systems trained on similar data. Pre-AI supply chains were inefficient partly because they involved a large number of human decision-makers who each had their own heuristics, biases, and information sources. This heterogeneity was a source of friction. It was also a source of diversity — different actors responding differently to the same conditions, which meant that no single bad bet could drag everyone down simultaneously.

Correlated AI decision-making is a different thing. If every major carrier’s routing AI learned from the same historical data that a particular sea lane is reliable in October, every major carrier’s AI will route through that lane in October. If something makes that lane unreliable in October 2029, every major carrier’s AI faces the same surprise at the same time. The diversity of response that characterized human decision-making disappears. You get coordinated behavior on the way up and coordinated crisis on the way down.

Financial regulators worried about this problem in trading algorithms for years, and eventually some guardrails emerged. Logistics regulators are roughly a decade behind.

The third concern is accountability, which in supply chains takes the specific form of the question: when the AI’s optimization causes real-world harm, who is responsible?

Consider a case from 2028. A pharmaceutical company’s AI inventory system, optimizing for carrying cost reduction, allowed safety stock of a critical antibiotic to fall below the level that would have covered an unexpected regional outbreak. The outbreak was not unpredictable in the general sense — epidemiologists have been warning for years about antibiotic shortages during respiratory illness seasons. It was unpredictable to the AI in the specific sense that it did not appear in historical data at the right combination of variables to trigger a precautionary inventory response.

The shortage lasted six weeks. Several patients in one regional health system received suboptimal treatment because of it. Who was responsible? The pharmaceutical company, which trusted its AI? The AI vendor, which had trained the model? The hospitals, which had not maintained their own safety stocks? The regulators, who had not specified resilience requirements?

The honest answer is that the accountability structure was not designed for this scenario, and six weeks of legal correspondence produced no clarity and no change in how any party managed its inventory.

Supply chains are invisible until they fail, and they are failing in ways that the metrics we built to track them cannot see. The efficiency metrics look good. The on-time delivery metrics look good. The cost metrics look good. The resilience metrics are largely absent, because measuring resilience requires specifying what shock you are measuring resilience against, and specifying that requires making predictions about future disruptions that no one wants to be accountable for making.

The Maersk Meridian will arrive in Singapore on schedule. The optimization was good. The manifest the AI assembled is almost certainly better than any human coordinator would have assembled. And none of that tells us anything about what happens when the next thing no one planned for arrives — which it will, as it always does, in exactly the form that the systems designed to handle the last crisis were not designed to handle.

The nervous system metaphor cuts both ways. A nervous system is exquisitely sensitive and responsive. It is also a system in which damage propagates rapidly, systemically, and sometimes irreversibly. We built the supply chain a nervous system. We should probably think about where the pain receptors are.