Before You Invest in AI, How Well Do You Know Your Factory?

AI is being discussed in almost every manufacturing company today. Many SMEs are looking at how they can use it to improve productivity, reduce costs, and make better decisions. But I think there is a more fundamental question we need to ask before deciding where AI fits into a factory.

How well do we actually understand what is happening on the shop floor?

In some factories, the owner knows almost everything. In others, information is spread across operators, supervisors, spreadsheets, and software systems. As the business grows, the challenge changes. What was once understood through direct conversations now has to pass through people, processes, and systems.

The information may exist, but does it reach the person who needs it, with enough context to make the right decision?

1. How much does your business know beyond what its people know?

Consider a machine that stops for 30 minutes. The operator knows what happened, the maintenance technician may know the reason, and the production supervisor knows how the stoppage affected the shift. But does the business retain all this information in a way that can be used later?

In many SMEs, operational knowledge depends heavily on experienced people. They know which machines are troublesome, which processes need attention, and where delays are likely to occur. But how much of this knowledge is recorded, and how much remains with the people who do the work?

As the business grows, the owner cannot be involved in every decision. New employees join, responsibilities change, and the same problem may occur on another shift without anyone recognising the connection.

What happens when the person who knows the answer is not available? Can the business still understand what happened, or does it have to start investigating from the beginning?

2. Do your numbers explain what is happening on the shop floor?

Most manufacturing companies have some form of production reporting. Some depend on spreadsheets and registers, while others have ERP, MES, and machine monitoring systems. But having numbers available does not necessarily mean we understand the reasons behind them.

Production may be below target, downtime may have increased, or material consumption may be higher than expected. The reports show the variation, but the underlying information may not explain why it occurred.

There is also a question of consistency. If two operators record the same problem differently, can we compare the records meaningfully? If the production report does not match what people experienced on the shop floor, which information should we trust?

In larger systems, the problem can be different. The data may be recorded correctly, but the context may be spread across production, maintenance, quality, and planning systems. Each system tells us something, but bringing the information together to understand the complete event can be difficult.

A number can tell us what happened without explaining why it happened. How much of the information we currently use for decisions actually helps us understand the factory?

3. Are we choosing AI based on the problem or the technology?

When we discuss AI in manufacturing, it is easy to start with predictive maintenance, production optimisation, quality inspection, or forecasting. Each has its place, but the value depends on the problem we are trying to solve.

If machine stoppages are not recorded properly, would predictive maintenance be the right starting point? If production data is already reliable but recurring losses remain unexplained, could AI help identify patterns that are difficult to find through conventional analysis?

And if information is available across several systems, is the main challenge the lack of AI, or the inability to connect and interpret what we already know?

These are different situations. One factory may need better data capture, another may need consistent recording practices, and another may be ready for advanced analysis. There is no reason to assume that all three need the same solution simply because they are all interested in AI.

I think this is where we need to be careful. Are we selecting AI because it addresses a real operational problem, or are we looking for a problem to justify the technology we want to introduce?

4. How will we know that AI has improved the business?

Suppose we implement AI to reduce machine downtime. What would tell us that the investment has worked?

If downtime was already being measured, we should be able to compare the results before and after the intervention. If it was not measured reliably, how would we establish the improvement? And if an AI system identifies a problem but nobody acts on the insight, what value has it delivered?

The same applies to quality, material consumption, production planning, and delivery performance. Generating predictions or identifying patterns is only part of the process. The real question is whether the information helps someone make a better decision and whether that decision improves the outcome.

For an SME, this matters because technology investments compete with other operational priorities. Would it not be more useful to demonstrate value by solving one important problem before extending the same approach across the factory?

The answer should be visible in the business, not just in the software.

Where does your factory stand today?

When I look at AI in manufacturing, I don't think every company needs to begin at the same point. Some businesses struggle to capture information. Others have records but cannot rely on them consistently. Some already have reliable data but find it difficult to connect that data with the reasons behind operational problems.

These are different challenges, and the right approach depends on which one exists in the factory.

Before asking what AI can do for your business, perhaps we should look at what we already know, what we are still unable to explain, and where that gap affects our decisions.

How well do you really understand what is happening in your factory today, beyond what your reports tell you?

That is the question I would start with.