We are putting artificial intelligence into the systems people depend on most. Power grids. Water systems. Communications networks. Public safety. The conversation around this is dominated by one question: what can AI optimize? That is the wrong question to lead with. In infrastructure, the question that matters is not how capable the system is. It is how safe it is when it fails.
Everything I have learned in twenty-five years of building and securing mission-critical systems points to the same truth. The systems that earn trust are not the most advanced ones. They are the ones engineered for the bad day.
The question we keep skipping
Most AI conversations are about capability. Faster decisions, lower costs, fewer people needed in the loop. Those gains are real. But infrastructure is not an app. When a recommendation engine fails, someone sees the wrong movie. When infrastructure fails, the lights go out, the water stops, or a first responder loses the network in the moment they need it most. The stakes are categorically different, and they demand a categorically different question. Not "what does this deliver when it works?" but "what happens when it doesn't?"
AI adds a new failure surface
Here is what gets lost in the enthusiasm. Adding AI to a critical system does not just add intelligence. It adds a new way for that system to fail.
It adds opacity, because a model's reasoning is harder to inspect than a rule someone wrote. It adds drift, because a model that performed well last year may quietly degrade as conditions change. It adds automation bias, because humans tend to trust a confident machine and stop watching. And it adds dependency, because once a system relies on AI to function, you have to ask what happens when that layer is wrong, unavailable, or attacked. None of this means AI does not belong in infrastructure. It means AI raises the bar for how carefully infrastructure must be governed.
Resilience first, not capability first
So the principle I hold is simple. In critical infrastructure, resilience comes before capability. You design for the failure first, then you add the intelligence. A system should degrade gracefully, not collapse. There should be a human fallback that still works when the model does not. And someone should be accountable for the failure mode before the system ever goes live, not scrambling to explain it afterward.
This is not caution for its own sake. It is the difference between a system people can depend on and one that merely impresses until the day it doesn't.
What oversight actually requires
Governing AI in infrastructure comes down to questions any serious leader can ask, even without technical depth. What happens to the people who depend on this if it fails quietly? Can the system fail safe, or does it fail catastrophically? Is there a human path that still functions when the AI is down? And who owns the outcome when something goes wrong? If the answers are vague, the system is not ready, no matter how impressive the demonstration.
The age of AI in infrastructure is here, and it should be. But the standard cannot be how smart the system is. It has to be how safe it is when it breaks, because on the other side of that system are people who cannot afford for it to fail.