AI won't fix your factory. Data will.

Anyone who knows me knows I believe AI can genuinely transform manufacturing.

But only if you get the basics right first.

I've sat through countless conferences where everyone is obsessed with AI. Then I ask about their MES/MOM strategy, or how they're collecting data from the shopfloor to feed those AI ambitions, and I get... nothing. Silence. A slide about a pilot.

Think about ChatGPT, Claude or Gemini for a moment. The models are remarkable. But they learned from the internet: decades of text, data and knowledge that people put online. Take the internet away and the most advanced model in the world has nothing to learn from. It's redundant.

Your factory works the same way. Shopfloor data is your factory's internet. AI is the technology that sits on top of it. Without the data, there's nothing for AI to work with.

So let me put my position plainly. AI won't fix your factory. Your data foundation makes it possible. And then AI will make it much, much better.

The slide everyone wants

There's a slide I use when I teach this subject. It says one thing: "Use AI to optimise your factory."

Simple, right?

Then I zoom out. Above that one line is a very long list of things you need first:

• Defined, documented processes
• An MES/MOM platform running those processes on the shopfloor
• Machines connected, with data collected automatically
• That data contextualised: which order, which material, which shift, which reason
• A data platform that holds it over time
• Dashboards people actually use, and baselines to measure against
• Training, buy-in, and a team that trusts the numbers
• A use case, a hypothesis and a model

AI is the tiny last line at the bottom.

Most people want to invest in that last line before the data architecture, the applications and the processes that let it succeed. They're still struggling to get basic downtime reasons recorded accurately, but their expectations are already at AI level.

AI isn't a black box

AI doesn't magic up answers. It finds patterns in data.

Machine learning, specifically, is a way to train an algorithm to predict an outcome based on what it's seen before. You need a clear question and data to learn from. Then you train the model, test it and refine it. It's statistics at scale. Useful, powerful statistics. But statistics all the same.

Which means the quality of what comes out depends entirely on what goes in.

Feed a model manual end-of-shift entries, estimated scrap figures and backflushed consumption from ERP, and you'll get confident-looking predictions built on sand.

What a good foundation looks like

A good data foundation has a few properties. None of them are glamorous.

It's granular. You need individual stops, not shift totals. Individual measurements, not daily averages.

It's automatic where it can be. Machine counts and states come straight from the equipment. People record the things only people know, like the reason for a stop.

It's contextualised. A signal that says "machine stopped" is worthless on its own. One that says which line, which order, which product, which shift and why, for how long and how often, is something a model can learn from. I've written about this in more detail in Raw machine data is worthless.

It's connected across departments. This is where MOM matters. Predictive quality needs process parameters, inspection results and material lots in the same place. Predictive maintenance needs machine condition alongside what was being produced at the time. If production, quality, inventory and maintenance each live in their own system, much of that context is lost at the integration boundaries before any model sees it.

Where AI is already earning its keep

Here's the part people forget. AI and machine learning have been in MES/MOM for longer than most realise.

Statistical process control has spotted drifting processes for decades. Condition-based and predictive maintenance use algorithms to detect patterns that come before a failure. Neither is new. Both are proven.

The newer applications are real too. The technology behind them is ready. What's lagging is the number of applications built to take advantage of it:

• Process parameter optimisation. Using history to find the speeds, temperatures and pressures that minimise scrap for a given product and material.
• Predictive quality. Predicting whether a part will pass inspection before it's inspected.
• Anomaly detection. Spotting unusual patterns in production data that no human would catch.
• Generative AI for people. Asking questions of your operational data in plain English, generating shift summaries, or explaining what a screen is showing. I cover practical examples in Practical GenAI for the shopfloor.

Each of these depends on the same thing. Good data, collected over time, with context.

You can't skip the steps

Everyone wants to be at the top of the maturity model. Predictive, autonomous, AI-enabled. Very few are.

Most manufacturers I speak to are somewhere between paper-based and partially digital. That's not a criticism. It's a starting point. But the path to AI goes through the middle, not around it.

Get the processes defined. Put MES/MOM on the shopfloor. Connect the equipment. Build up clean, contextualised data. Then point AI at the problems that matter most.

Do it in that order and AI amplifies everything you've built. Skip the steps and you'll be back at the start in two years, with a very expensive pilot to explain.

So where should you start?

Not with a model. With a question: what decision would you make differently if you had perfect data about it?

Then work backwards. What data would that take? Where does it come from today? How much of it is accurate, automatic and in context?

That gap is your roadmap. And for most factories, the first steps on it look a lot more like MES/MOM than AI.

What's the one decision in your factory you'd most like AI to help with, and do you have the data for it yet?