Why AI in Production Is So Hard to Get Right

At Elmia Produktionsmässorna this week, Scandinavia’s largest manufacturing trade fair, one seminar title stopped people in their tracks: “Why AI in production almost always fails.”

Independently, a report published that same week by Manufacturing Dive covered the MIT Initiative for New Manufacturing (INM) Symposium held in Cambridge, Massachusetts in early May. Executives from Ford, GE Appliances, Amgen and ArcelorMittal had all pointed to the same root cause: data management and accessibility is the single biggest barrier to scaling AI in manufacturing. Not the algorithm. Not the model. Not the price of compute. The data.

This deserves a closer look.

Most AI projects in manufacturing follow the same arc: a team runs a pilot, gets promising results in a controlled setup, then hits a wall when they try to scale. The culprit is almost always the same: the data needed to make AI useful in production isn’t accessible, isn’t connected, and isn’t current.

That’s a structural problem.

Manufacturing data lives in fragments. The ERP holds order history. The MES tracks machine output. The operator keeps the rest in memory. These systems are seldom talking to each other and were designed for humans to manually enter data.

AI doesn’t work on batch data entered after the fact. It needs context: what job is running, on which machine, at what specification, for which customer, with what material batch. Preferably in real time and across systems.

Most factories don’t have that. Not because they haven’t invested in technology — many have spent significantly — but because their systems were built for a different model of work. The ERP was designed to record what happened. AI needs to understand what’s happening now, and what should happen next.

The factories succeeding with AI are the ones that solved the data problem first.

The good news is that this is a solvable problem. You don’t need a perfectly integrated tech stack before you start. You need to begin mapping your information flows: where does data live today, who touches it, and at what point does it become invisible to the rest of the operation?

The factories that will lead with AI in the next five years aren’t waiting for a perfect system. They’re starting now, one process at a time, building the context layer that makes any model actually work.

That foundation is entirely within your control to build.