How Industrial AI Is Really Used in Plants

There are two things to know about industrial AI first.

1. Using industrial AI is not the same as implementing ChatGPT in your plant.
2. Industrial AI has been around well before ChatGPT existed.

Now that you know these two points, let’s get into what industrial AI is, how we are currently using it and things you need to know before implementing industrial AI in your plants.

This is a good basic guide for you if you are an operations manager who is considering implementing AI in their plant.

What is industrial AI?

So what is industrial AI really? As I mentioned, industrial AI has been around well before all the buzz created by ChatGPT. We have used and still use AI in a plant for things like predictive maintenance. For example, a simple AI can let you know before a pump is failing so you can fix it.

But for you to better understand industrial AI, you need to get a clear understanding of the main types of AI.

The three types of AI

There are three main types of AI. Each one has a different job. Here is a simple way to see them:

1. Predictive AI

Predictive AI is the oldest type. As mentioned, this type of AI was around well before ChatGPT existed and it is fully safe to use for industrial plants. Predictive AI works in a simple way. It uses the data that already exists in a plant and learns the pattern. This learning from data is called machine learning. Based on what it learns, it can tell you when a machine is about to fail. It can also detect a bad product using a camera. Checking parts with a camera like this is called computer vision.

We also use this type of AI in our everyday life. For example, the spam filter in Gmail uses the previous data to predict which email should go to your inbox and which one should go into the spam folder. Or your Netflix account uses the data of what you have already watched to give you recommendations of what you may like to watch next.

2. Generative AI

Generative AI is the ChatGPT style AI we use these days. This type of AI is good for things like answering questions, generating text or creating images. For example, you can ask ChatGPT to write an email for you. In an industrial plant, you can use generative AI to do things like generating draft PLC code, or helping you understand a PLC program that is written by someone else.

3. Autonomous AI

Autonomous AI is a type of AI that can do things on our behalf. An example of Autonomous AI is the self-driving car built by Tesla named Robotaxi. This car takes you from A to B without needing you to drive. An example of using Autonomous AI in a plant is to investigate a production problem, decide what action to take, and handle the supporting steps on its own. But it does not fully control the process. A person or the PLC still has the final say. Autonomous AI is the best type of AI for an industrial plant because it can think more like a human. For example, autonomous AI can understand if you do this, then that happens. This type of thinking is something generative AI does not have.

Generative AI limits for industrial plants

As I mentioned, when we are talking about generative AI, we are talking about ChatGPT style AI that are designed to generate text, image, code, etc. These are very useful AI but not the best type of AI for a high-risk environment like an industrial plant.

Here are three reasons why generative AI is not the best type of AI for industrial plants.

1. Hallucination

Like it or not, ChatGPT-style AI sometimes make things up and when they do that they usually do it with a very high level of confidence which makes it worse. And this is not a flaw in the system, this is how the system is designed. This means that this system is designed to sometimes make things up. This is because generative AI is designed to never say “I don’t know”. The model is designed to always give an answer, even if it’s not correct.

On the other hand, the model is also designed to receive feedback from the users. Those little thumbs up and thumbs down buttons you can click after each answer is how generative AI knows what type of answers you like. And it turns out people hate vague and unsure answers.

So overtime, generative AI has learned that if it gives answers with a higher level of confidence, humans like it more.

But this makes the issue even worse, because now the AI can give you a wrong answer with a very high level of confidence which makes it harder for you to find out if this is a made up answer.

So generative AI is designed to sometimes give you a made up answer which is bad but the feedback it receives from the users makes it give that wrong answer with an even higher level of confidence which is very bad.

If you’ve been using generative AI for a while, chances are that you have dealt with this hallucination thing a few times already. But there are also a few public examples of this.

For example, in 2024, Air Canada’s website chatbot told a customer about a refund rule that wasn't real. The customer applied to get a refund but Air Canada refused it because the refund policy did not exist. But later, the court said Air Canada had to pay because its AI gave a made up answer.

Another public AI hallucination example are the two lawyers in New York City who used ChatGPT to help write the court papers. ChatGPT made up some fake court cases that did not exist. The lawyers did not check it. The judge found out and fined them $5,000 each.

These may sound like funny or worst-case embarrassing examples for those poor AI users but imagine if this happened in a high risk environment like an industrial plant. Here the AI hallucination can result in someone dying or a whole batch worth of hundreds of thousands being destroyed. This is not a risk we can take.

2. No thinking

Generative AI does not actually think like a person. It can’t understand “if I turn this up, then this will happen.” Instead, it just guesses the next word that sounds right, without really understanding what causes that. On a plant floor, that kind of cause-and-effect thinking is exactly what we need to protect people and expensive machines, and generative AI simply does not have that.

3. Different answer every time

If you ask ChatGPT the same question five times, it gives you five different answers. This is fine for writing an email but, again, for a high risk environment like a plant floor, this is not something we can use. Here we must have the same correct answer every single time. 100 times out of 100. Generative AI does not have that either.

Where generative AI is okay to use

Ok, with all that said, you may ask, should we stop using generative AI for industrial plants? The answer is… “not, really”. We can still use it but we need to intentionally use it in a low risk way.

For example, when a company that built the machine puts its own generative AI chatbot inside that machine, that is fairly safe. Why? Because, usually, nobody knows that machine better than the people who made it. So this is low risk generative AI. However, you take a general chatbot and start using it on all kinds of new or old machines in a plant that is very high-risk.

Another use case of generative AI in an industrial plant is to use it to generate draft PLC codes. And I emphasize on the word “draft” here because this should be treated as a draft not as a final code. The PLC programmer needs to review the code, revise it and finalize it before downloading it to the PLC. So using generative AI for generating draft PLC is good but taking that as a final code and downloading it to the PLC is bad.

Closed-loop AI for industrial plants

When we bring AI into industrial automation, one big question always comes up. Can the AI run the process on its own?

Closed-loop AI or “human out of the loop” is when the AI controls the whole process in a plant without any person being involved. This may sound cool in theory but in practice, no serious plant does this. This is because in a fully closed loop or human out of the loop AI system, if the model makes a mistake nothing can stop it and this can result in hurting personnel or causing damage to the machine or the batch which, as I mentioned before, can cost hundreds of thousands of dollars. So we never let the AI fully control an industrial process and we always make sure that there is always an operator or a PLC that says the last word and controls the process.

So the role of AI in industrial automation is to predict and advise, not to control.

What is AI drift?

In a plant, we need to calibrate sensors and actuators regularly to keep them sharp. This is because after some time these devices are not as accurate as day-one. The same applies to the AI models we use in a plant. Regardless of whether you use a Predictive AI for prediction or an autonomous AI for more advanced industrial applications, after some time, the AI starts to drift. Meaning it won't be as accurate as when it was installed. This is because AI learns from how a plant works today. But the plant slowly changes. Machines wear down. Materials change a little. And after a year passes, the process is not exactly the same as last year. So over time the AI starts to receive slightly different inputs for learning and as a result it gives a slightly different output which slowly makes it less accurate. This slow drop in accuracy is what we call AI drift.

So what does this mean? This means that you need to be aware that implementing AI in your plant is not a “set it and forget it” thing.  Because AI models drift, they need to be recalibrated once or twice a year and this costs time and money. So as an operations manager, you need to take that into consideration when budgeting.

Where do we use AI in Industry?

Here are some examples of real use cases of AI for different industrial facilities.

Load shifting

Industrial plants use a lot of electricity. For many of these plants, electricity is one of the biggest costs. And electricity is not always the same price. At some hours it is cheaper and at some hours it is more expensive.

Now if a plant does more of its heavy work when the electricity is cheaper, it can save a lot of money. And this is what we call load shifting.

This load shifting can work in many industries. For example, a steel plant can do more of its melting work during the cheap hours. Or in a food factory, the cold storage unit can cool things down at night when power is cheaper, so it can work less during the day. Or as another example, a car factory can build up a stock of parts during the cheap hours and slow down when the power is expensive.

So as long as the output stays the same we can shift the load or the work to the cheaper hours so we can save on electricity cost.

This is something we can set up with AI. For example, a simple example of such a system can be a combination of two AI agents. One AI agent watches the electricity price all day and another AI agent talks to the PLC and plans to run the heavy machines when the power is cheaper and makes sure that the output stays as planned. So with this setup, the plant makes the same amount of product but pays less for electricity.

Predictive maintenance

Picture a machine in a factory. This machine has some small sensors that constantly check (sense) how much the machine is shaking, how hot it is and what sounds it makes. This is kind of like a smartwatch that checks your heartbeat all day.

An AI agent watches all this data nonstop.

One day, the machine starts shaking in a weird way. This usually means that a part inside, (like a bearing) is about to break.

Now, if the AI is a Predictive AI, it can send a warning light on the HMI and wait for a human to see this and take action on it.

However, if it is an autonomous AI, it can take some further follow up actions on its own. For example, for this case, it can take the followup actions below:

1. Check the factory’s computer system (ERP) to see if we have a spare part available.
2. Book a repair technician to come and fix it.
3. Start filling out the paper work (the work order) so the task is ready to go.

So instead of just saying "hey, something is wrong here," the AI takes the next steps on its own. It finds the part, books the technician, and does the paperwork before the machine even breaks down. A person still approves and does the actual repair, but the AI has already done all the supporting work. That "taking the next steps on its own" part is what makes it an autonomous AI.

How to start without wasting money

Here are a few simple points that can save you a good amount of time and money if you are starting to implement AI in your industrial plant for the first time.

Start with a real problem, not with “we need AI.” Just because you have added a budget line for AI does not mean that you should start looking for AI vendors and spend your money. The right way is to first look at your plant operations and ask this simple question: Where do we lose the most time or money? If so, then you can see if AI can help there. Do not buy AI first and look for a use later.

Is this a real bottleneck? This is a simple but important question that you need to ask. Just because AI can save you some cost does not mean that you should attack this problem with AI at the moment. If the project is not a real bottleneck of plant operations, it should not be considered.

How long does it take to get the investment back?  Do a simple math in your head. If you are investing to build a system that saves you $30K per year but costs you $300K, you should ask if this is really worth it because it takes 10 years to get your investment back, not counting any potential maintenance for the system. In this case, it might be a better idea to keep paying the current $30K per year cost and use that $300K for a real bottleneck of the plant.

Get your data ready. AI needs data to work. If you’re not collecting data or your data is not clean, maybe a good first small project is to start collecting and cleaning data.

AI is not a shortcut. Many people think that AI is so powerful and so magical that it can fix messy plant operations on its own. That is not true. Just like any other area of life and work, there is no shortcut. AI can work well on top of a good industrial operation. It makes what you already have better, not magically fixing things. So please adjust your expectations. Otherwise, you will get so excited about AI and what you pay will be only the cost of “learning how to implement AI” and not actually the cost of “implementing AI”.

Okay, that’s all. You now know the basics of industrial AI more than most other operations managers in the industrial space. I recommend reading this article again right before you are about to have a meeting with an industrial AI vendor to keep your mind sharp.

And if you happen to visit the SPS (Smart Production Solutions) trade show this year in Nuremberg, Germany or Atlanta, USA, it’s good to know that our main focus this year is Industrial AI. So it is a great opportunity for you to learn more about industrial AI and connect with people and companies in this space. You can get your ticket on this page. Good luck with your industrial AI journey.

FAQ