But one of the world’s biggest industries has largely bypassed the software revolution, and still runs, to a remarkable extent, on human know-how. In manufacturing, the real bottleneck is usually not the machine on the shop floor: it is the person running it, carrying years of hard-won knowledge in their head, earned one job at a time… and whose expertise is near-impossible to scale.

In the current US context, where tariffs are back at the center of industrial policy, this matters — significantly. In Washington, everyone suddenly wants more American industrial capacity. But while policy can change incentives, it cannot by itself create capability.

This is what too many discussions about reshoring still miss: a factory’s limits are not just physical. They are cognitive.

As well as buildings, machines, and customers, you need experts who know how to actually run things: to quote for new work, program the job, avoid scrap, and work around the particular foibles of the machines in the shop (and those operating them). Too much of that workflow still has to fit inside somebody’s head.

This is where manufacturing has always gone wrong. Those running factories almost always prefer to buy machines over software. But this is backwards. The machine is the overhead; the software that captures knowledge and joins everyone up is what determines how well the business runs.

I once saw the CTO of Ocado (the UK online grocery and technology company) walk up to a terminal in a warehouse that told him exactly what to do. He could work effectively almost immediately, even though he had never been in that part of the building before. The intelligence of the operation had been captured by the system.

Today, that’s absolutely not what happens in manufacturing, where everyone in every factory has a different way of doing things, and all that knowledge is siloed. But if we’re to keep up with society’s need for the things we want, that has to change.

The solution to this problem is AI. But in manufacturing, the need is not for a solution that helps us communicate better or create more; AI has to become a building block that supports the rest of the industry.

An AI Fit for Purpose

Factories need systems grounded in the real constraints of tools, machines, materials, tolerances, and physics. In manufacturing, wrong answers have physical consequences that break expensive machines and halt production.

So, an AI that’s actually useful in that context will be domain-specific, reliable, and embedded in real workflows. At my company, CloudNC, we’ve built an AI that thinks like a machinist, and it already accelerates CNC machine programming in hundreds of factories across the US. This wasn’t an overnight success — it took 10 years to build (so far) and may never be perfect or complete — but it shows that expert judgment can become software.

Once that happens, the effects are not theoretical. The hidden tax on manufacturing today is waiting for the one person who knows how to do the job, create a toolpath, or make a key decision. AI can free senior programmers from repetitive work so they can oversee more work, or give juniors a strong starting point instead of a blank screen. Done properly, applying domain-specific AI makes best practices become more consistent, and the scarcity of expert knowledge stops being the bottleneck.

This is why AI matters to reshoring and defense far more than most people realize. If the US wants a stronger industrial base, it will require more than tariffs, subsidies, or new facilities. It will need factories that can absorb more complexity and produce more with the skilled people they already have. That is especially true in defense, where the conversation has shifted decisively toward manufacturability, affordability, and speed.

This step change does not stop at CAM, or at any one workflow. The same pattern will play out across quoting, process planning, scheduling, setup, inspection, and the dozens of operational decisions that still live in tribal knowledge. Over time, AI becomes not just a tool inside the factory, but the way the factory is run.

That is when manufacturing starts to change shape. Uptime rises, lead times shrink, costs come down, and more production moves closer to where it is needed. And the pace of innovation in hardware starts, at last, to creep closer to the pace of innovation in software.

Every industrial revolution starts the same way: first as an advantage, then as a requirement. AI in manufacturing will follow the same path. It is currently an enabling technology, but it will quickly become a required one.

The factories of the future will still need great machinists, engineers, and operators. But they will no longer be constrained by how much critical knowledge can fit inside a few people’s heads. And once that shift takes hold, AI will stop looking like an optional tool for manufacturing: it will look like infrastructure.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.