AIで「見る・考える・行動する・学ぶ」：Honeywell Technologies の Suresh Venkatarayalu 氏

Going beyond automation, what are its four layers, and why does having a more nuanced understanding of this application of technology matter? Find out on today's episode. I'm Suresh Venkatruailu at Honeywell Technologies, and you're listening to Me, Myself, and AI. Welcome to Me, Myself, and AI, a podcast from MIT Sloan Management Review, exploring the future of artificial intelligence. I'm Sam Ransbotham, Professor of Analytics at Boston College. I've been researching data, analytics, and AI at MIT SMR since 2014, with research articles, annual industry reports, case studies, and now 13 seasons of podcast episodes. In each episode, corporate leaders, cutting-edge researchers, and AI policymakers join us to break down what separates AI hype from AI success. My guest today is Suresh Venkatarailu, Chief Technology Officer at Honeywell Technologies and President of its software business.

Honeywell makes the control systems and software that quietly run a lot of the world's refineries and chemical plants and utilities and buildings. Suresh has been at Honeywell for about 30 years, and now he's leading the shift from automation to autonomy. And that's important because the mistake that's a bad answer on the screen is kind of something we can laugh at and ignore. But a real-world problem is something very different. Suresh, thanks for being here. Great to be here, Sam. Delighted to join this podcast. So tell us at first, what do you do at Honeywell? What does the Chief Technology Officer do? Well, my job is all about defining the vision for the company, the roadmaps, the future, predicting the trend, and really driving the transformation, innovation engine for the company. The best part of my job is to shape the future. Well, I've got to think that AI is going to be a big part of that future, or it certainly seems to be. Tell me about this idea of automation to autonomy. What's your thinking there?

For us, automation to autonomy means adding more intelligence on top of your automation system that drives efficiency, eliminates waste, and then probably makes the worker and human's life a lot easier to get things done. You know, Sam, if you really go back and look at Honeywell's history, for 140 years we've been seen as a control system company. What do we do the best? We build systems within industrial plant or a commercial building or industrial environment. We built a system that does to the precision point, your inputs have to deliver a predictable outcomes. Some of them would ask me a question, do you use AI? I would actually say for all our 140 years, we built AI, but we used to call them built around deterministic models, heuristic models, right? And these were built around more mission critical and safety critical standards.

What has happened in the last few years, we have started seeing what we call AGI, artificial general intelligence, which is probabilistic model with limited knowledge and insight, ability to train and learn. You allow the system to learn on its own. So now in Honeywell, we are trying to really bring in the best of the world on both sides. We are clubbing our probabilistic model, which we call AGI, on top of our deterministic model. And we are actually trying to really build the learning system on top of your safety critical standard that had automated all our 140 years. Give me an example of something that takes a previously deterministic model and layers learning on top of it. There are four layers to autonomy, which is see, think, act, and learn. That's how you kind of really define your progression cycle for an autonomy.

What we have done throughout in the last 140 years, we have done two things so well. We have done C, meaning we have incorporated ACT. What it means is C is you have multi-million sensors and actuators providing a real-time input and feedback with a high level of precision. and then you had workers who took decisions and actions based on predefined rules. Now, with the advent of AI, we are incorporating more logic and intelligence into the system. Assume that you're actually taking 20 to 60 percent of false positive coming from sensor inputs or from field devices. the systems are able to eliminate false positive, get the much more accurate inputs and data, guiding the users to take decisions. In certain areas, it's not only guiding, it's providing more additional reasoning. In certain areas, if you start trusting the system well,

you can actually hand over the reins back to the system. For an example, you ask for a very specific example. In a commercial building like this, you have a building management system that is designed to really manage efficiently your operations, energy, and multi-systems connected to a building management system. In the past or even in the current form, your building operations center, the operators are managing your building efficiently. They're throttling between heating system, cooling systems, and multi-system. They're taking decisions all together. Now, when you have more AI with more agents incorporated into a system, they are observing the patterns of the building, looking at efficiency, inefficiency. They start guiding and suggesting to the operator what else can they do to really prune the system and to improve the efficiency.

Now, you can start to teach and learn. and as the workers are taking actions, the system is also learning. Learning is part of the reinforcement process. So it will actually pop up to say, do you think this is false positive? I think the operator is saying yes. And if he continues to say yes, he's actually training two things. One is he's reinforcing the algorithm that I have a high confidence your prediction is right or he or she is going to go back and change the way you operate your building your SOP your process framework So I think it a very beautiful orchestration of how we built a system for a very predefined control loss. And now you have a system that's allowing us to think and optimize and evolve. And I think that's what I think about as the future of a continuous learning systems embedded on top of your automation that makes it interesting between semi-autonomy or autonomy. Yeah, I think that makes sense to me.

Maybe a little simplistic here. I was thinking about my thermostat on the wall here, which I'm not really convinced is connected to anything, to be honest. But, you know, we don't directly turn the AC on and off. We tend to set a temperature. And then we rely on the proportional integral derivative type controls to reach that steady state that feels like the temperature we want. So that's a clean example that's been working for a long time. It sounds like you're expanding this to really everything you can get your hands on in terms of sensors and that learning. Maybe that building operator before knew what was a false positive and what was a false negative. You're trying to get that information into a system so that it can learn. What's that going to do about what people need to know to run a building, for example? But you might not need to know exactly as much about the details of how the boiler works. Or do you need to know more details? I'm not sure how that goes. Maybe three years ago, when the frontier models were launched, like GPT 3.5 and Gemini initial version,

we started by taking all our legacy knowledge from service tickets. We took our technical publication, manuals. We took most of the artifacts. We trained the system. We did that for two reasons. One, we'll continue to see a skill shortage or labor shortage. Either way, in the U.S., we have a labor shortage, and then the developed market has a skill shortage. We felt like we need to really offset that disparity, and we built our system. I think the first innovation was called maintenance assist. We felt like every service technician in the field had to have certain knowledge hands-on, just like we would actually use ChatGPT for a general purpose. They should be asking a question around the systems or subsystems or the asset that they are repairing or fixing things altogether. Can they actually get the highest level of accuracy? We spent probably the first two years taking all our systems from warehouse, commercial building, industrial plant,

and we took most of the repository, trained it. We started with an accuracy of probably 45, 50 percent to start with. Over a period in time, once we started training the system, we went closer to 85%. I believe 85% above is good enough to start with where you're starting to get the thumbs up from your end user. We saw a massive productivity uplift with our field service engineers, with our field operators who had been supporting our customers. And the fact that we got a high throughput, we said now the system is ready, not just in the field engineering side. We took it down to our product engineering and project engineering side. Now, as we went through the whole product lifecycle management arena, and we started really building this frontier models on top of your lifecycle from requirements, designing, implementation, and then actually going through the field, you know, user acceptance testing process.

As we got it, we actually called it as project assistant, design assistant, proposal assistant. We saw a huge uptick in our productivity. Then we actually saw, look, this is actually getting to the point. Now we can bring those models back into our control system products so that it can actually start to support our systems because the control systems were built to run an operation, But if it has more inbuilt intelligence around the building, building operation, if you can contextualize the building assets and their operating conditions, the system can be more intelligent. So we are probably into the second generation of adopting those intelligence as AI getting better and better over a period of time. Now you actually see agents are starting to operate seamlessly because they are actually using this basic intelligence that's been trained over a period in time. Remember, AGI is absorbing all the information from Internet. What we are trying to do is we're taking industrial data knowledge,

which we have more implicit or explicit knowledge. We've been training the system for the last couple of years. I think it's a generational shift which leads into Sam where we started defining it as physical AI because physical AI is all about systems and subsystems, and can you really bring in intelligence closer to the physical world. Yeah, so it's kind of interesting. As you were talking about that, some of the things I was categorizing were, hey, I'm sure these engineering people put together these huge tomes of operating manuals that they feel good about and they're constantly frustrated that nobody reads because they're giant and huge. And so that's about how your systems work. But then the other part you brought in was this context of, like, well, what is different about this building? And the idea of meshing those two together, that seems like it has a lot of potential. That's right. The industry uses this word called ontology. And to really define it in an easy term, we started using this tag called Teach the AI the World of Your Domain,

meaning the world of your domain is a building or an industrial plant or an environment. How do you teach? How do you teach how the building operates? And within a building, how these subsystems between HVAC, BMS, FHIR, Security, Access Central, and ITOT system, how do they operate? And then if you double-click it, there are categories of a building. You may have a building which is a data center type, a hospitality hotel, could be a commercial building like where we are right now. So we are going through the journey right now. So what's fascinating is the approach we have taken for the robotic systems in the past. Now we are actually trying to really bring in that context into a world where we believe these buildings are becoming an industrial robot in its own terms. And if AI can take the lead and if you can teach the context of a building or an environment altogether, and if they become intelligent over a period in time the prospects we going to see are much different than what we have seen so far Yeah at a minimum I hope just the elevators show up when I ready for the doors to open and they ready Sam I should add something back to you We implemented in our North Carolina campus

with IDMIA your ability to scan your face and then it determines which floor you belong to. It understands who you are, so we don't need to really flash an access card altogether. Yeah, actually, I feel like it's a Devil Wears Prado moment where the operator could have my coffee and take my coat right when I walk in the door. That's my fantasy is to run my life like that. I don't envision myself as the assistant. I envision myself as the devil in that scenario. I was reading about one of these petrochemical industries that you were working in, in Buruj-Abu Dhabi, that was trying to predict problems sort of five or ten minutes ahead. Now, what does that do? Five minutes could be a huge time. What difference could that make? Think about this way. This is one of our largest ethylene plant where we have close to 20,000 to 30,000 sensor data that is processed every second. And when you process every second, you will end up seeing closer to 15,000 to 20,000 alerts or alarms that keep popping up.

your ability to really dissect between what's a false alarm or a clean alarm, that process of eliminating and getting a highest level of accuracy to guide your operator, that you actually gain more time. And that's what we have actually done. We actually said we will incorporate this. I think we use a word called Experian cognition, which sits on top of this ethylene plant where we process thousands of sensor data points with feeds and inputs, the intelligent system ability to dissect between false positive and false negative. And not only that, it provides an operator insights, recommendation, and also a factual data pattern, what the prediction could be. And that's where we find it fascinating, where we bring in the AI intelligence with a deterministic model, which is an algorithm that guides them what the potential root cause is,

what the potential scenario could be, and then guiding them what they need to do. So we have done a couple of pilots so far. The feedback has been exceptionally good. In fact, they have gone ahead and published externally that this is the good proof point for the future autonomous industrial plant, and we are engaged with them in a series of next few cycles. The idea that you could get sort of a five-minute head start when something is happening feels like it'd be pretty important. Sam, there's a reason we use the word called mission-critical or safety-critical standards, because sometimes even 20 seconds would make a huge impact. I think Honeywell is known for helping, supporting from a destruction or a disruption that could actually be catastrophic in nature. Yeah. Did you ever see the Galaxy Quest movie? If you saw it, it was funny. But the premise there is that they have some magic device that gives them the ability to roll back 60 seconds in time.

And, you know, the whole movie is, ah, what can you do with 60 seconds? And it turns out, of course, it's a giant plot point that you can do a ton different in 60 seconds. Maybe let's dig into this a little bit more in terms of I think it all sounds good when the alerts are good, when you're getting them filtered. What about when they're, I think you mentioned 85% right is good. What is the 15%? How are you thinking about how to handle the cases where the systems aren't right or don't give good advice? You know, that's the reason that we believe that these systems are not going to replace operators or humans. We absolutely believe this is supposed to augment and enable and help. If I can remove 85% of their inefficiencies out and help them to take a lot more better actions and decisions, Sam, I think the throughput is going to be exceptionally high. We took an example in a building. If I need to go back and do a dynamic set point optimization, flip your temperature based on external temperature,

that's not massively catastrophic in nature. once you're starting to train the system, saying that, okay, I'll get from semi-autonomy to autonomous more. Now you understand the weather pattern. You understand the temperature pattern. Set point optimization. I can trim down 10 to 15 percent of energy efficiency. Okay, trigger off autonomy. But you would want the operator to enable that autonomous condition, not the other way around. Yeah. You know what I was thinking, you know, is maybe to push back a little bit. You said, oh, yeah, if we have 85% right and the 15% wrong, people can pay more attention to. I think it's kind of like marketing, though. And what's the famous marketing quote? Half my marketing dollars are wasted. I just don't know which half. Well, the 85%, 15% makes sense if you knew that this is the 15% that's a problem. But the uncertainty, I think, is the problem there. How are you conveying the uncertainty in information that these systems are giving to people? I actually said our guardrails are around the deterministic model,

meaning we have the probabilistic model that could be 85% accurate, but we give the reins back to the control system that we have built, which always assured 99.9999%. So the way we operate, Sam, is it's not like an autonomous system that is built without a guardrail. It is built with a guardrail where we have a control system for industrial environment, and then we call EVI, Enterprise Building Integrator, in our building management automation system. And those two were built in the last 20, 30 years, assuring the highest precision point. But what we've really done is we've brought additional intelligence on the top. That could be an 85% assurance on top of whatever we have built, but then you have a guardrail to protect you. Yeah, I think that's a good framing because I think this layering was missing from my take on it. the idea that this is in some sense no worse than it was when you had the 99% accurate deterministic models.

If you're adding this on top, then it becomes additive versus let's just, you know, throw an LLM at it. That's exactly right. Yeah, I do believe in the deterministic models, and I do believe in the engineering behind those kinds of things. What's hard about all this? It good that you asking me the question I tell you the learnings that we have gone in the last three four years in my job Access to your data infrastructure wasn an easy one because it not like you putting a router in a building and collecting the data. You're talking about putting a gateway in an industrial environment where the data is disparate, systems are disparate, and information are disparate. this pattern. So your ability to collect the data, connect the systems, and then curate and ship the data back wasn't an easy one. Two years ago when we started, you know, a building which is a million square feet, 10 floors down, it would have taken anywhere between 120 to 150 hours.

Now, if it takes 120 to 150 hours to connect a million square feet with probably a thousand assets, you're automatically losing the trash. And then we work not directly with customers, we work with system integrators, our channel partners. Now, you're talking about the channel partners who have specialized technical skill sets to a very small system integrator may not have enough technical skill set to go back and deploy this gateway who can provide an assurance to our customers about cyber protection, data connectivity, and GDPR, you name it. I think we have come a long way in today we can connect anywhere between 10 to 20 hours within the building context. We deployed a massive program called Forge Academy. We went ahead and trained, certified most of our system integrators. We offered them a massive program, giving them a gateway, giving them a partner portal so that they can onboard this building, onboard the system.

They can issue licenses. So there are a lot of upstream processes. So number one, I would call it as connectivity wasn't an easy one. I think we believe that we got that right in terms of frictionless connectivity. Just to give you a fact, Forge is our platform, or we call an intelligence layer on top of our physical systems. Two years ago, we had probably 700,000 assets connected. As of July 2026, we have 5.3 million assets connected to our Forge platform. So the rate in which that we have connected has gone significantly up. We believe that our install base is close to 11 to 12 million assets. So we want to get those assets connected faster. So we brought the whole idea of ontology and deploying at scale. We have made a lot more progress in the commercial building side. We're making a first step of process in our industrial environment because an industrial environment is 10x complex than a commercial building, Meaning I can connect to an asset, but I also need to connect to a process units like they are more sophisticated to understand how the process works,

because you're talking about some of the catalyst performances that needs to be digitized to allow an AI system to understand what it means. So we are in the early days at this point in time. So second one was contextualizing it. I think the third one is once you have the data discovered, once you contextualize it, then can you actually build an agentic platform that agents or AI can start to use contextual data to take decisions and action? So that's where in the last four to five months we've been running a number of pilots. It seems like a great opportunity for digital twins because you can simulate what all these sensors would do under different temperature conditions and those kinds of things. I've always wondered about how the fidelity of these twins, how do you get them to actually model? You've got a bunch of data from real buildings, and you're trying to build digital twins of those buildings. What's the process of making sure that that simulation actually reflects what's going to happen in a real building? That is interesting.

Six months ago, if you've asked us to go and simulate and emulate, the process itself would have taken a building like this three months to just render the three-dimensional artifact. Today, we've evolved to a much higher level of precision. We were able to feed the CAD drawing of a building and the PNID documents that are available today. Within a day or two days, we are able to render the whole building and the context of the building. And then once you have an ontology which provides the contextual nature of the building, it's almost like a brain behind the building comes in light. So it's not to say that it's 100% there. It's probably 80, 85%. We are fine-tuning as preceding it. For me, you do not have a digital swan. I don't think we are ready for a first-generational autonomy conversation. I have built a robotic system, supply chain automation. I've built, in the past, truck loaders, truck unloaders, packet picking.

We have built a number of different applications. But the first thing that I was always taught was the first principle is the ability to simulate and emulate. Once you can actually do so, then you can go in real-time action. I think right now that's probably what's getting ready in this world that we are living with. Actually, I like that because the idea of being able to get to that good simulation quicker then will sort of unleash perhaps a bunch of other things that you can then try it out. And I think we're seeing that in lots of things. We're seeing it in healthcare and we're seeing it in other places. As I think about Honeywell and I think about others in this space, One of the things that I squint at it, what I look at is a lot of pieces that have been developing over a period of 20 or 30 years, and it seems like we're at a moment where these things are all sort of coming together. The compute capacity, the ability to say, hey, I have 20 or 30,000 sensors out there that can be connected. I have these large language models that are coming out all the time. All these things are coming together, and it seems like all that's happening about the right time.

I think it's an exciting time to be here. Thanks for taking the time to tell us about what Honeywell is doing. Thank you so much, Sam. Thank you so much for your time. Hi, everyone. Thanks for listening for the last six years. I have had a fantastic time hosting Me, Myself, and AI alongside our producer, Allison Ryder, and our engineer, Dave Leshansky, and originally with my co-host and fellow chemical engineer, Shervin Cotabande. Along the way, we've interviewed some inspirational thinkers and fantastically capable technology leaders. Please don't ask me to pick favorites, because I honestly can't. We hope you've enjoyed listening to this archive, and will continue to do so for many years to come. Thank you.