AI Is Getting Easier. The Physical World Isn't.

What industrial systems taught me about users, context, and better products.

I've worked across AI, software products, SaaS, integrations, and industrial systems. These experiences looked very different on the surface, but over time they led me to the same conclusion:

Good products start with understanding the user. Data is how that understanding compounds.

1. The physical world made this obvious

In software, a lot of context already exists. Documents contain text, APIs have schemas, databases have fields, and applications generate logs. The information may be messy, but much of it has already been translated into a form software can work with.

The physical world was different. A machine doesn't explain what is happening. It vibrates, heats up, makes different sounds, and behaves differently depending on operating conditions. Sometimes the most valuable information isn't in the data at all. It is in the experience of the person standing next to the machine.

Someone might know that a certain vibration is normal under one condition but unusual under another. They may notice something from the sound, the installation, or simply from years of working with the equipment. That context was often not written down anywhere.

So we had to go and find it.

We visited factories, looked at how equipment was actually operating, talked to users, checked how sensors were installed, and tried to understand why two machines that looked similar on paper behaved differently in reality.

That experience made one thing clear:

Collecting data and understanding what the data means are two different problems.

2. The workflow matters as much as the technology

At first, I thought the challenge was fairly straightforward: collect good data, build a reliable model, and detect problems accurately.

In reality, the model was only one layer. We had to deal with sensor installation, communication, security policies, software, vibration analysis, mechanical systems, and different operating environments. But even those technical problems were not the hardest part.

The harder questions often came after detection.

Who receives the alert? How do they decide whether it matters? What do they inspect? What action do they take? And what happens after that?

An alert could be technically correct and still create no value if nobody knew what to do with it.

The hardest part wasn't detecting the problem. It was understanding what happened before and after the detection.

That meant the user's workflow had to become part of the product itself. We needed to understand not only the machine, but also how people worked around it.

Over time, I started thinking about the product as a loop:

Observe → Understand → Decide → Act → Learn

The system observes what is happening, tries to understand it, helps someone make a decision, and leads to an action. Then the outcome needs to come back into the system.

Without that last step, the product keeps making predictions.

With it, the product starts learning.

3. Data is how understanding compounds

We often talk about data as something used to train AI. I think that definition is too narrow.

For a product, data is also a record of how users behave, what decisions they make, what went wrong, and what happened afterward. It is how the product gradually builds a better understanding of both the user and the environment.

This becomes especially important as a system scales. With a small number of users, experienced people can fill in missing context manually. They can talk to the customer, inspect unusual cases, and remember what happened previously.

At scale, that stops working.

The product itself has to capture more of that understanding. It needs to preserve context, record outcomes, and create feedback loops that make the next decision better than the previous one.

This is why I increasingly think the quality of a product depends not only on how good the technology is, but on how well it can continuously learn about its users.

AI is getting easier. Understanding users isn't.

AI is becoming dramatically more capable. It can read documents, analyze images, write code, call APIs, use software, and execute increasingly complex workflows.

Software intelligence is becoming cheaper and easier to access.

But better AI doesn't automatically create better products.

The same questions remain:

1. Who is the user, and how do they actually work?
2. What context does the system need to make a good decision?
3. What happens after the decision, and does the system learn from the outcome?

These questions mattered in industrial systems.

They matter just as much in AI products today.

AI makes intelligence cheaper. It doesn't make understanding users automatic.

Good products still begin with understanding the user.

And data is how that understanding grows over time.