PLC Copilot for industry
What if AI could address the most persistent challenges in the industry? In this episode of AI at Scale podcast Malini Nambiar, Digital Customer Innovation Director for AI Applications at Schneider Electric, tells the story of how customer insights helped develop a GenAI assistant that solves some of the most pressing pain points of industrial companies: talent gap, long learning curves, legacy programming languages and time-consuming code customization.
The solution: Industrial Copilot for PLC. Built on Schneider Electric’s trusted PLC libraries, it helps developers generate robust, working code- fast.
In this episode, you will learn:
- How Industrial Copilot accelerates PLC code generation using Generative AI.
- Why multilayer guardrails ensure safety and compliance in automation workflows.
- How Copilot helps bridge the talent gap in the industry and boosts engineering productivity.
- What the future holds with agentic AI and autonomous system orchestration.
If you want to look under the hood of industrial AI solutions, this conversation is a must-listen.
Listen to the AI at Scale podcast
Listen to Malini Nambiar: How AI is changing industrial engineering episode. Subscribe to the show on the preferred streaming platform (Apple Podcasts, Spotify) or play the episode on YouTube.
Or play the episode on Youtube.
Transcript
Gosia Gorska: Welcome, everyone. This is the AI at Scale podcast by Schneider Electric. My name is Gosia Górska, and I’m the host of this program. My guest today is Malini Nambiar. She is the Digital Customer Innovation Director for AI Applications at Schneider Electric. Malini has spent over two decades driving innovation in industrial automation, and she’s currently leading the charge on how AI, and especially Gen AI, is reshaping the way engineers work. Today we are discussing Industrial Co-pilot and the future of AI in engineering. You will discover how AI is transforming the way engineers create automation code for industrial machines, the role of human oversight in this process, and key challenges in adopting AI in engineering workloads. So, I’m very pleased to welcome Malini.
Malini Nambiar: Thank you, Gosia. It’s a pleasure and honor to be here today.
Programmable logic controller (PLC)
Gosia: Yes, thank you for joining. Let’s start with the three letters that we will probably be mentioning a lot in this conversation: PLC, Programmable Logic Controller. I remember when we first met and I didn’t know this abbreviation. You explained this to me. Could you also explain it to our audience?
Malini: Sure. A Programmable Logic Controller, or PLC, is a digital computer, designed for industrial control and automation. It’s very rugged so that it works well in an industrial manufacturing setup. It’s typically the brain of the machine or the process which it’s monitoring. So, it’s monitoring the input devices, it’s executing a program written by the user, and it’s controlling.
Copilot how does it work with PLC code generation ?
Gosia: Yes, exactly. We are discussing today the Industrial Co-pilot created by Schneider Electric. My first question is what inspired the development of such a co-pilot in the first place, and how does it specifically address the challenges of PLC code generation?
Malini: Certainly, I would love to share about this journey with you. When we started with the idea of a Co-pilot, initially, when we heard about ChatGPT, you know, how magically it was creating code for any language under the sun, we thought, “Well, PLC also needs to generate control code.” Structured text and function block diagrams are some of the languages for PLC code generation. Much of it is legacy code, legacy languages, but it still continues to be very widely used in the industry.
So, we told ourselves, “Let’s put this to the test. Let’s see how ChatGPT handles PLC code generation.” We had to validate this. What we did next was to set up a task force comprising some of the bright minds from the Industrial Automation CTO office and the AI hub teams, and they did a quick technical investigation for us. After they finished this investigation, they came back with very positive results, saying, “This is real. PLC code is certainly generated by ChatGPT.” Once that was established, our next target was to validate this from the market. Typically, we do that by making a market evaluation to find out how it’s seen in the market, if there are any references. But Co-pilot and ChatGPT being very new, we had absolutely no references available. So we decided to rely on our internal expertise.
We had a lot of expertise already, which I think was a very wise step. So, we got our Subject Matter Experts (SMEs) and Application Design Engineers (ADEs) together, and we wanted to understand from them what kind of problems they face when they create applications or support our customers when they design or engineer a certain application. From our brainstorming, we got a good list of potential customer pains. Once we had this, we wanted to validate it. We wanted to hear it from our customers as well to see if they corroborate what we found. So, we went forward and we went to the OEM market customers. We interviewed them and what we got is a variety of concerns that they deal with every day. To name a few: there’s an aging workforce problem; years and years of expert knowledge which is at risk. There is new generation employee churn because new generation employees want to work on more recent languages like Python and not legacy coding languages, and the learning time taken to learn these languages. The amount of time they spend dealing with code at the OEM side, the extent of customizations—all of these problems came out when we conducted that voice of customer.
What we had was the pains, the frequency of such pains, the gravity of which was very convincing for us to say, “Okay, we just need to build our own Co-pilot to address all of these customer issues.” From there, there was no looking back, and that’s how we started on this exciting journey of creating our own Co-pilot.
GenAI and benefits for engineers
Gosia: I see. So the need for the assistance was absolutely clear. But how did you come up with the idea that Gen AI can also be used for automation of PLC code generation? And what are the key benefits for engineers and also for system integrators?
Malini: Great question. When we were embarking on this journey of the Co-pilot, one of the fundamental questions we asked ourselves was: now that ChatGPT can generate PLC code in whatever languages, what would the customer need additionally from us? Why would they need another Co-pilot? What is the additional value that we can provide to our customers? When we dug deeper, we realized that the code coming from ChatGPT does not carry any Schneider PLC context. It’s very generic, boilerplate kind of code. None of our PLC libraries, which are our wealth, none of that information is available with ChatGPT to design our applications. We had our answer: we have to customize the OpenAI based LLM with our PLC libraries so that we could generate non-hallucinated, genuine code that could run applications using Schneider PLCs. That exactly is the key differentiator and benefit that our own Co-pilot provides over and above the vanilla GPT.
Yet another benefit that was very clear from our Voice of Customer is the program workflow feature. When we talk to our customers, one of our biggest findings in these interviews was about the time our OEMs spend dealing with a lot of legacy code. In many cases, almost 80% of the application is fixed; it doesn’t need to change. But what needs to change or customize is the 20%. Essentially, they spend so much more time in understanding the application itself, figuring out where the customization needs to be done, than the actual customization. This kind of a tool comes in super handy in order to chop down the entire application control flow. We do this in very simple visual steps. Any engineer, without having much of the legacy application knowledge or the PLC programming skill, could just go ahead and understand the program workflow and then perform this customization. What we do in addition is also to provide a possibility to test and document the changes that were done in totality with the entire application, and this enables machine delivery in quick time. So, what we bring to the table with this Co-pilot—specifically the generation tool and the program workflow—is enabling the customer to reduce the engineering time required to build a new machine.
Safety of the GenAI solution
Gosia: I see. So basically, you have discovered a new need among the customers: not only to automate the code but also to work on the legacy codes. This is extremely helpful. But as you said, we are not offering a vanilla GenAI. It’s a GenAI that is based on Schneider documentation. But still, Gen AI is probabilistic. So, how do you ensure that this Gen AI solution is safe and reliable?
Malini: Very important, very critical question, and something that we get asked by our customers often. When we generate code from our Co-pilot, we are also very careful to have this code go through a supervision of multiple layers of guardrails. It’s built into our knowledge repository. The first and the most critical one, I would say, is to have our specific Schneider PLC libraries. That acts as a guardrail to ensure that the LLM that is generating code is both valid and compliant with the exact PLC hardware that is being used in that application. It’s adding relevance there. The next layer consists of the coding best practices and guidelines to ensure that what we generate out of the Co-pilot is high-quality output according to the industry standards. Finally, we also have the provision to top this up with specific customer application knowledge, which is critical to generate code that is very close to the specification of the customer. So there are these multi-layer guardrails that are in place, and this is what gives us the confidence that the output that is finally generated is high quality, consistent, safe, reliable, and very close to what the customer is needing.
Human oversight in PLC code generation
Gosia: So you have covered this potential risk with a specific architecture of the solution, but I was also wondering if there is still human oversight in PLC code generation and is there a feedback loop between the engineers and Co-pilot?
Malini: Absolutely. It is something very critical. I think we discuss this topic very often. I think humans play a very important role in building all of this, without which we just cannot get something close to what we have today. So human oversight plays a very important role, specifically in the high-quality outcome that is delivered by a Co-pilot. Our ambition here when we use the Co-pilot is essentially to enhance efficiency. The way to do this is basically to augment the efforts of the human who is involved in the process so that we can enable them to deliver faster, standard, reliable output. I already mentioned how we build our knowledge repository based on expert feedback. We have adopted a three-step strategy: generate, validate, and train. Based on a certain user prompt, the Co-pilot first generates the code, making references to the inputs that we have given to the LLM. The Co-pilot also lists these sources that it references. So, you know that we are not generating code out of thin air. It’s very much from the references that were already fed, and the LLM was taught using these inputs. The human in the loop can now validate the output by referring to these list of references or validating the references, and then the committing of the code can happen. While this validation is happening, there is systematic feedback that can be provided back into the knowledge repository system. By means of that, we enable the training of the repository. Over a period of several iterations, the knowledge repository becomes richer with the sources plus expert knowledge that is going in, and then what we get is high-quality output that can be generated.
Misconceptions about AI
Gosia: Yes, so it’s like you are also improving with the knowledge that is coming from the usage of the tools. That’s a really great idea. You know my favorite question about AI: AI is science, but it also sits under innovation and discovery. My favorite question is always about the myths and misconceptions. What are the misconceptions that you would like to fight about AI and industrial automation, especially when it comes to PLC code generation?
Malini: Absolutely. I think this is one of the most commonly asked questions as well. It’s true that AI is more a reality than hype today. We have several solutions that are proven in the market, even if I think there is still speculation about the ROI. But we must be very aware of the changes it brings—not to be afraid, but to be very prudent so that we can embrace it and adapt AI to our daily work. About the misconceptions, I would say the most common one we hear all the time is about AI replacing PLC programmers and control engineers. This is the most common and also the most emotionally charged misconception, I would say. AI is primarily an augmentation tool. It’s not a full replacement, and for the foreseeable future, AI will continue to excel at tasks like optimization, prediction, and automation of repetitive tasks, such as generating boilerplate code or optimizing tuning parameters. But the core task of creating the core logic, handling safety, managing non-standard system integration, will still remain very human-driven. So, I would say AI will change the programmer’s job, making them AI-assisted engineers, but not eliminate the role. The risk is not the elimination of the job, but certainly there is a change in the nature of the job which is internalizing AI quite heavily, and this is something that we must be aware of. That’s one of the misconceptions about replacement of jobs.
The other one I can talk about is AI being ready for critical real-time control applications or control functions, which is very much the case for our industrial automation setup. The reality is that traditional PLCs are built for deterministic, hard real-time control and safety-critical operations. AI algorithms, particularly machine learning, require a massive amount of clean data for training, which is also a very big concern. Gathering data is not an easy job; digitization itself is quite an important action that needs to be done in order to gather this clean data. This makes AI very less reliable for the instantaneous, guaranteed response kind of control. So, in my opinion, AI’s current strength is still in the higher level of optimization—maximizing or minimizing energy consumption—these kinds of activities which run alongside the stable control of the PLC, not really replacing it. Of course, as the tech evolves, this will change over a period of time, but I think for the foreseeable future, we could see that AI is still not very ready for the critical real-time control functions.
And there’s one more misconception that we already talked about, which is about not requiring human oversight. Like I said, hallucinations and errors that can be made by AI systems are quite common. It depends on the training that we do. It is crucial to ensure that human oversight and human-in-the-loop are continuously being managed in order to generate code out of AI so that we are making safe and correct decisions.
How does Copilot cooperate with other tools ?
Gosia: Yes, exactly. Thank you for debunking the myths for us. I also wanted to take a helicopter view and take a look at other offers by Schneider that are also in the space of industrial automation. In what ways does the Co-pilot integrate with existing engineering tools and workflows, such as EcoStruxure?
Malini: Sure, I’d love to share about that. Our ambition is a full integration of the Co-pilot inside our EcoStruxure Automation Expert Platform. We call that EAEP. Application engineers can rely on the Co-pilot to create a full application given a set of input artifacts, like requirement specification or control in case of a process application. They can also start from creation of a project solution, addition of required assets, creation of control logic, creation of HMI, and testing the application. All these are different steps. In the case of a machine application, which mostly needs customization, like we talked about, the Co-pilot enables understanding of the existing code base, identifying pieces that need optimization, and even finally suggesting where you can add this code. That’s our ambition where we want to get to. With our EAEP offers, all the offers are coming together under one platform. We also plan that one Co-pilot will enable switching context based on intent detection to ensure that it caters relevantly to the request that it receives, no matter where we are in the system. This will ensure that the customer has a very unified user experience while navigating through the different spaces in the platform, simplifying the complexities of application building.
Future of automation systems
Gosia: Basically, a bright future is ahead of us, and this would be my last question about the future. How do you see this technology evolving in the next three to five years? Could we reach a point where entire automation systems are designed and deployed?
Malini: Yes, the future is always fascinating, and we must talk about that. AI, of course, is evolving very fast, and it touches on the most cutting-edge developments in industrial automation as well. I think the next three to five years will be a period of intense evolution primarily driven by Generative AI, and more specifically, Agent AI is going to play a big role. The most significant leap will come from Agent AI and its usage for system-level orchestration, which means systems can be designed to pursue a certain goal by planning, by reasoning, and executing tasks, often by coordinating with other agents or other systems.
As an example, on the engineering side, we talked about the work that has started around creating highly capable, intelligent layers and tools that make human engineering more productive: code generation, HMI generation, and all the others that we talked about, like documentation and explanation. All of these are available today as functions that can be triggered with specific prompts. You want to generate code, you put a prompt, and it generates code. You want to generate HMI, you put a prompt, and it generates HMI. But imagine if instead we only have to give one prompt, one goal. For example, say, “Design a conveyor system to move 1,000 boxes per hour.” And all of this gets done. You don’t have to tell the system or prompt the system for every step; everything is completed. All these different agents of code generation or HMI generation work in tandem to deliver this conveyor system for you. That would be incredible. Another similar example on the operations space I could envision is about an autonomous optimization agent. If we give it a goal such as “Maximize throughput while maintaining product quality above 95%,” it would adjust set points in the underlying PLCs without any human intervention and get the task done. That’s the fascinating future that we’re looking toward.
Wrap up
Gosia: Yes, and it really is fascinating and optimizing for me because in past episodes we were discussing different applications for AI, such as finance and energy management. Every time we discuss how the usage of AI is actually making the job of specific experts much easier, much simpler. I hope that this is also the impression that our audience will get from this episode—that it is a fascinating assistant that can help us, especially with tasks where we are having a lot of documentation ahead of us, a lot of legacy codes, as you said in this case. This is a great assistant. Do you feel excited about this for your work as well? I think you have an amazing career in the AI space.
Malini: Thank you. Yes, absolutely. 100% super excited about what and how the future unfolds for us in industrial automation.
Gosia: Yeah, that’s great. So thank you, Malini, so much for all the insights from our conversation today.
Malini: Thank you very much, Gosia. Thank you.
Gosia: And also thanks to all of you who are listening and watching us. You can follow us on podcast platforms and on YouTube. So stay tuned for next episodes. Thank you.
Like what you hear?
- Visit our AI at Scale website to discover how do we transform energy management and industrial automation with artificial intelligence technologies to create more sustainable solutions.
- Listen to the other episodes of the AI at Scale podcast.
AI at Scale Schneider Electric podcast series continues!
The first Schneider Electric podcast dedicated only to artificial intelligence is available on all streaming platforms. The AI at Scale podcast invites AI practitioners and AI experts to share their experiences, insights, and AI success stories. Through casual conversations, the show provides answers to questions such as: How do I implement AI successfully and sustainably? How do I make a real impact with AI? The AI at Scale podcast features real AI solutions and innovations and offers a sneak peek into the future.
Add a comment