AI in Manufacturing: Where Manufacturing Companies Should Start

In almost every manufacturing plant I've walked into, someone had already tried something with AI. Usually it was started by someone in maintenance or production planning, because that is where the problems are most concrete.

And usually it stalled at the same place. Not at the model, but at the question of how the data gets off the machine in any form you can actually work with.

That is the real starting line for AI in manufacturing. Not picking the technology.

Why manufacturing is a special case

In the back office, data sits in systems that were built to hand data out. Email, ERP, CRM, file storage. You can get to it.

On the shop floor it's different. There's a machine from 2009 sending values to a controller that talks to a piece of software nobody has had a support contact for in years. Next to it, a new machine that logs everything imaginable, but only into its own portal. And between them, a measuring station where someone writes values on a clipboard and types them into Excel at the end of the shift.

That's not negligence. It's the normal outcome of twenty years of investment decisions that were each correct at the time. It just means the question "can we use AI for quality inspection?" is a data question first and an AI question second.

What you actually need to get started

The good news: considerably less than most people assume. There are three prerequisites, and none of them is "clean data across all systems".

One: You can get to the data for a single use case. Not all the data. The data for one process step, over a limited period, exported by hand if necessary. A CSV file is enough. If someone can export those values manually once, the prerequisite is met.

Two: There is someone who really knows the process. Not the process description in the quality manual, but the person who knows why batch 4 always gets re-measured and that the line behaves differently after a weekend. Without that knowledge you build something that misses reality.

Three: It's clear what a good result would be. "Less scrap" isn't enough. "The defect gets caught at loading bay instead of at assembly" is enough. The difference is that you can check the second one.

What you explicitly do not need: A cleaned-up data landscape, an MES project as a precondition, a data warehouse, a cloud strategy or a data scientist on staff. All of that may become necessary, but later. Not as a prerequisite for the first step because that's the most reliable way of never starting.

The mistake that costs the most time

The most common mistake in manufacturing companies isn't naivety. It's the opposite: starting too big and too thoroughly.

A typical project might look like this. Someone decides to do it properly. So first an inventory of every machine and data source. Then an assessment of which lines can be connected. Then a concept for a unified data platform. Then quotes for the platform. Nine months later there is a very good presentation and not one single insight into whether AI does anything for this company at all.

The alternative is neat, it works, and it's cheap to run. Take one use case, export the data for it manually if you have to, prototype something in two days that actually runs, and see whether the thing holds up. If it does, you now know exactly what a proper data connection would be worth. If it doesn't, it cost two days instead of nine months.

Where it's worth looking

The best first use cases in manufacturing are rarely in machine control. They sit at the handovers: wherever information moves from one area into the next and somebody reshapes it by hand along the way.

• Inspection reports someone retypes so they make it into the system
• Complaints someone reads in order to assign them a root cause
• Maintenance reports written on paper and never evaluated
• Customer enquiries about spare parts where someone goes hunting for old drawings
• Shift reports copied together from three sources every morning

These use cases have one decisive advantage over anything that reaches into machine control directly: if the output is wrong, somebody corrects it. No production line stops. That keeps the risk small enough to simply try and learn.

I've collected concrete examples with numbers from real projects in a separate article on AI in manufacturing.

If you want to work out which workflow is worth it in your own area, there's a free 30-minute video with a template for you to fill in that walks through exactly that type of assesment.

Watch the free 30-minute video

Why the people from the shop floor have to be in the room

There's a reason AI projects in manufacturing fail on acceptance so often: the workforce has lived through several waves of digitalisation over the past decades, and not every one of them made the work better.

When a system built by an outside entity and announces that it will be handling quality assessment from now on, scepticism is the reasonable response. Particularly since the person at the machine is often right: They know the exceptions the model is missing.

Which is why it works better when the people who run the process help build the prototype. Not out of politeness, but because their knowledge of the exceptions is the actual gold. Anyone who has watched their own hard-won experience turn into a rule inside the system has a completely different attitude towards it than someone who gets shown the finished result.

There's a second problem this solves on the side: those same people can afterwards judge whether the next vendor's pitch is realistic.

A first use case in two days

For manufacturing companies, the fastest route to an answer you can rely on is a format where exactly those people sit down together: someone from the shop floor, someone from IT, someone who owns the budget.

The AI Hackathon is that kind of format. A 2-day AI Innovation Workshop where your team takes a real process out of the plant, pulls it apart, and ends up with a running prototype. On your data, on your case, built by your people.

What you're left with at the end isn't the question of whether AI works in manufacturing in general. It's the answer to whether it works in yours.

Check out the AI Hackathon

Back to overview