LNS Research

Ask most manufacturers what they want from industrial transformation, and the answer sounds reasonable enough: find what works, standardize it, and then scale it across the network. Sounds clean, logical, and efficient, right?

The plants, however, have other ideas.

Of course, every facility carries its own history. The equipment was chosen at different times (some, quite long ago) by different leaders with different priorities, and it shows.

Some companies have tried to impose top-down standardization all the way to the equipment level; we have even seen some try to go remarkably far down that rabbit hole. But most find it becomes an endless project that never quite delivers and has to start over every time there's an acquisition or a technology refresh.

What we're seeing work better is an entirely different kind of standardization. Instead of trying to make the equipment itself uniform, you standardize how you interact with it. This builds the layer that lets different machines, systems, and data sources speak a common language to the business above them, without requiring everything underneath it to be identical.

Sure, it's less glamorous than a uniform factory floor, but it's also much more realistic.

The goal has never been about making every plant look the same. It's about being able to see, understand, and act across all of them with trust and transparency.

Making every plant identical and being able to operate across all of them aren't the same problem, and a lot of transformation programs run into trouble when they're treated as if they are.

There's a mismatch at the heart of most industrial AI deployments that doesn't get examined closely enough.

Manufacturing is built around precision and repeatability. However, generative AI, by its nature, is probabilistic. That's not a flaw, it's just, it is what it is.

But those two things are not in the same neighborhood, and glossing over that gap is where a lot of projects get into trouble. It gets even more complicated when you factor in that manufacturing isn't just a single decision; it's a sequence of them.

Think about something as routine as a batch release process: an AI flags an anomaly in raw material testing, a second system assesses whether it affects downstream quality, and a third determines whether the batch meets spec for release. Each step depends on the one before it.

If each system is highly accurate but not perfect, those imperfections don't cancel each other out. They stack. By the end of a five or six-step sequence, your cumulative reliability might look very different from what it looked like at step one.

None of this means AI doesn't belong in manufacturing. It absolutely does, in some form or another. But deploying it without understanding where that risk lives is how you end up with outcomes that are hard to explain, and even harder to defend.

The real issue isn't can AI do this. It's where does probabilistic output create acceptable risk, and where does it create unacceptable risk. Those are two different things that require very different approaches.

There's an assumption baked into a lot of industrial AI thinking that autonomous is somehow the more advanced option; that it's where you end up when you've really figured things out.

But it's worth pushing back on that thought. Because it actually gets the relationship backward.

Think about how most automation got built in the first place. Someone watched a skilled operator, documented what they did, and eventually codified it into a system. At some point, the rules became clear enough to hand it all off to a machine.

AI plays a different role.

What we're seeing at LNS Research is that AI tends to do some of its best work in situations where the rules aren't fully defined yet, and where there's still ambiguity to navigate. It helps organizations learn what's actually happening in a process, and that learning is genuinely valuable, often more valuable than the automation that follows it.

In that sense, AI isn't the destination. It's how you get to automation faster and with more confidence than you could before.

What this all means in practice is that reaching for AI because it feels like the modern answer can sometimes be the wrong call. What some of those situations truly need first is a better understood process and a more reliable system to run it.

So, bottom line, before you leap, it's best to know which one you're actually trying to solve.

We surveyed 300+ manufacturers about quality, and the results were a little uncomfortable.

The data told us that most companies, despite clear advances in technology, people, and processes, still treat quality as something that takes place near the end. This generally looks like teams finding problems before product ships out.

But that's not really quality, is it? That's just inspection with a fancier org chart.

What we see at LNS Research is this: the manufacturers that are pulling ahead are the ones that have moved quality upstream. Their operational leaders have built it into product design and supplier conversations; into decisions that happen long before anyone touches a production line.

One Chief Quality Officer we spoke with said something that really resonated: "We stopped measuring defects. We started measuring decisions." Talk about reframing the conversation with a simple ah-ha moment.

If your quality team is still measured on things like reject rates and audit scores, you're measuring how well you're catching problems. But you aren't measuring whether the problem should have existed in the first place.

In his latest research, LNS Research Senior Analyst James Wells looks at five things the companies getting this right are doing differently. LNS Research members, you'll find the link in the first comment below.