AI Won’t Fix a Broken Operating Model.

A model is not an operating decision

Rail is fertile ground for artificial intelligence: noisy demand, constrained assets, complex connections, intermittent failures, and many repeated decisions. Yet a recommendation becomes valuable only when someone is authorized to act on it and the surrounding workflow can absorb the change.

A model can predict that a block will miss interchange. It cannot decide whether the local railroad should hold other cars, call a crew, request a different slot, notify affected customers, or accept the miss. Those actions belong to the operating model.

Bad incentives produce bad intelligence

AI systems learn from recorded outcomes and organizational behavior. If teams avoid coding an exception because it harms a metric, the model learns from sanitized history. If estimated arrivals are overwritten without preserving the prior forecast, the organization loses the evidence needed to understand error.

The most dangerous implementation is a polished answer placed over contested inputs. It converts uncertainty into authority without doing the work required to earn trust.

“AI does not remove the need for an operating model. It forces the organization to reveal whether one actually exists.”

Build the decision envelope first

Before selecting a model, define the decision: who owns it, what options are allowed, what constraints are inviolable, what evidence is required, and when a human must intervene. This decision envelope turns AI from a general promise into a bounded operating component.

Start with decisions that are frequent, reversible, and measurable: prioritizing exceptions, proposing work lists, flagging likely misconnections, drafting customer explanations, or identifying assets for inspection. Measure whether the recommendation improved the decision, not whether users opened the tool.

The contrarian view

The best early AI projects may look unambitious. They will reduce time spent finding the next problem, make uncertainty explicit, and help experienced operators compare options. They will not claim to run the railroad.

That restraint is a strength. Reliable decision support creates the operating evidence and trust required for more autonomy later.