Next in Tech | Ep. 268: Autonomy

View Full Transcript

Eric Hanselman   00:00:00

Welcome to Next in Tech, an S&P Global Podcast with the world of Emerging Tech lives. I'm your host, Eric Hanselman, Chief Analyst for Industry Research at S&P Global. And today, we're taking a deeper dive into autonomy in AI and industrial control systems. And joining me to discuss it is Cody Falcon, Global Digital Products and Technology leader for ABB's Energy Industries division. Cody, welcome to the podcast.

Cody Falcon   00:00:26

Thank you for the invite, Eric. Looking forward to the discussion.

Eric Hanselman   00:00:30

Well, it's great to have you here. This is actually a conversation that we had started back at CERAWeek and the CERAWeek Agora on an autonomy panel there. And it is such a -- I was going to say, it's fascinating, it's relevant. There are so many different angles when we start talking about what does autonomy mean. Now with the rise of Agentic capabilities, we've got everybody's various open claw systems running around doing things on their own. I guess with so much buzz on Agentic capabilities and talking about autonomous operations, when we start to translate that into industrial processes, autonomy really starts to take on some very different -- well, not really so much meetings, but very different aspects that we've got to be aware of. I guess how should we be thinking about autonomy when we start thinking about process control?

Cody Falcon   00:01:26

Yes. It's a great question, Eric. And I think you're spot on it. And we started this conversation at Sara, right? And when we had some other panelists, I think we had a panelist from Microsoft, for example, right? And you could see the difference or the dichotomy between applied AI, specifically Agentic and autonomous workflows in the business side being pretty far apart from what we do on the industrial side, right? And I think there was an analogy given using Copilot inside your office suite to help you sharpen an e-mail is a lot different than an agent making a set point change in a live production process, right? The risks are much greater. The guardrails have to be tight. And at the end of the day, right, what we do in agents and AI, it has to be right, right, if we're going to make these decisions. So we tend to lag the business side of innovation in this area, intentionally so, right, because we have to make sure that, again, our recommendations and agentic actions are right every time. We cannot afford to miss or be wrong, which is quite different from what you see from Open clock, for example.

Eric Hanselman   00:02:31

Well, it's something where -- I mean we're a long way from sort of clippy sort of popping up and saying, hey, it looks like you're running a fractioning column.

Cody Falcon   00:02:40

That's right.

Eric Hanselman   00:02:42

Where do you want to crack these molecules?

Cody Falcon   00:02:44

And you know what's interesting, when we talked about this a bit, I think, at CERAWeek was one of the biggest, I think, misconceptions or maybe learning we have with our clients, right, is understanding that this industrial autonomous journey, right, is a spectrum. It is not a binary on-off or switch, right? It is a journey, right? And of course, ABB, we published our 6 level 0 to 5 reference autonomous map in 2019, but SAE and other standards are out there, right, from assisted to supervised to conditional to full autonomous control. And what we found is most customers today, right, most sites, most processes, most plants are somewhere between level 1 to 3 of 5, right? And so I think the average, the last I looked was close to a 2, 1.5 to 2 is really where most are at. And so it's not about leaping from Level 1.5 on autonomy up to Level 5 fully lights out plant. That is way off in the future. I think it's a bit more assessing where you are, being honest, where you are today and making those incremental steps, right, getting closer and closer to the next level up. So that's the big one. The big difference for me is thinking about autonomy as a spectrum or a journey versus a switch or a binary.

Eric Hanselman   00:04:08

Well, to your point, this is something where -- I mean, we've been using machine learning and what you could identify or some forms of AI in control systems for a long time. Well, we've been doing that in ways that help to augment control capabilities typically to be able to actually do observational work to be able to actually run data analysis stuff I guess what should be the kinds of capabilities that we have to take in place to be able to take those next steps to be able to go from those level 1s and 2s up. How do you -- what do you have to build in order to support that journey?

Cody Falcon   00:04:50

Yes. Look, I mean, I can speak to it from our side, what we're having to do, right? Of course, you have to have what we call a data foundation, right? These agents have to have contextualized data, whether it's from a historian, unified name spaces. There's a bunch of kind of enabling technology or plumbing underneath that allows these agents to work, right? And oftentimes, we overlook the technological plumbing underneath and some folks may believe, hey, I can just put an agent in play tomorrow, and I'm not really sure where it runs or how it's going to work. There's some real heavy lifting required on the technology piece, specifically on the data, the run times if the agent is working between applications, for example, right, you may have some API changes. That's something that we're working through, for example, when we talk about our various applications that run on top of the control system data, right? These applications now work in a coordinated way using agents. Well, those agents use APIs to work between the applications. So you may have some API changes needed at the application layer. a few technology enablers there. And you made me think, Eric, when you opened that question, you mentioned around using AI for some time in this space. I think the conversation for us is shifting from clients asking, can AI predict this to can AI decide and act on this, right? That's really what's shifting, and we're moving it away from dashboards or insights that a human has to interpret to where our agents are now kind of interpreting that data, logging into the next system, gaining a bit more context logging into the next system, building up their decision tree before they take an action. And so again, I think the shift is, can AI predict something or forecast something or make a recommendation? Can -- now it's can my AI decide and act on my behalf.

Eric Hanselman   00:06:47

Well, you talked about one of those things that I think we stumble around a lot, which is data and the access to data pieces and ensuring that we've got data where we need. A big part of that also is trusting the data that we've got and building an infrastructure to be able to go establish that foundation of trustworthy data. We've got more than enough examples of challenges about not being able to get to the data we need or having things go sideways in terms of how we're actually accomplishing the kind of telemetry we need to be able to get systems to operate well.

Cody Falcon   00:07:23

Yes, that's right. I mean the data has to be there, right? But one interesting observation we've seen is -- it's less about -- particularly with agentic workflows. It's less about massive data lakes, right, and having your data all aggregated and something has to run there because now your agents can go log in and get what's needed from the various data sources, databases, data lakes, applications, right? I think just a few years ago, I think there was a big push on enterprise lake strategy where you bring all the data to one place and everything can point at that. Well, now it's shifting a bit to where just leave everything where it's at, right? The agent can now go log into the various systems and get the context that it needs.

Eric Hanselman   00:08:06

Well, but then you've got to raise that level of concern about data security, of course, in terms of making sure that you've got -- the right agents are getting access to that data. That's right. This is, again, one of those things about the trust infrastructure is not only ensuring the data itself is robust and durable, but that the access infrastructure you build around it has got to be a challenge. And those are areas in industrial controls, we haven't had that same level of history in terms of understanding access and some of those capabilities. It seems like in many cases, that can be complicated to overlay.

Cody Falcon   00:08:44

Yes, it is. And it opens up -- we think about it as an agent is just like a human, right? So a human, what sort of access control, permission, authorization, escalation, privileges, et cetera, that you would have for a human being logging into your system, your network, your application, you need that same level of robustness or security control around an agent, right? You cannot have just a general admin privilege and agents running around, right? You have to treat these agent workflows as if they are human users with the right level of permissions, right, the right gated controls, the right authorization levels, et cetera, just like a human user on an IT network.

Eric Hanselman   00:09:24

Well, it's one of the conversations that we've been having around AI data architectures is that challenge of understanding that you -- yes, you've got a whole bunch of data, yes, you've got controls. And to think when you get to that next step for agents, now you've got what is this entity that actually can start to bring some of this data together. So if you're looking at data partitioning to be able to manage security, those are areas where agents can start to leap over some of those preconceptions about that you may have based some of your security controls on and ensure that you've actually got those aggregations of data as well protected as well.

Cody Falcon   00:10:05

That's right. Yes, you're spot on, and it's something we think a lot about, right, is that what we call the agent memory, right? And I explained this to my kids. We were talking about agents. We've got college age kids that are thinking about what the future looks like, and we talk all the time about agents and where AI is headed. And one of the ways that I explain it is an agent is just like -- it's similar to how a human being gains context when they log in across applications, right? So if you're going to think about a problem that you need to solve, you may need to log into 3 or 4 different systems. And every system that you log into for a few minutes, you get a few data points and you catalog that. You store it in short-term memory and then you log into the next one. And you're just building that kind of memory up and then finally say, you know what, I have enough information to make this sound decision and you go make the decision and take the action. And then it goes away. And so agents sort of work in the same way where they have a context memory window, right, where it's logging into various systems, it's pulling back context. And it doesn't necessarily have to store everything it's ever seen in a massive database, but there is some memory involved as it logs in and pulls context and gains understandings and reasons with itself before making a decision. You're spot on, Eric, that memory is a big topic, right? How much memory is stored? Where is it stored? How long is it kept, et cetera? Can it bleed over and be referenced in other queries? Yes, the agent memory is a definite topic.

Eric Hanselman   00:11:34

Well, you raised an interesting point there also about one of the challenges with aggregating information, which is when do we have enough? And what's the sufficient level of context that you have to build in order to make a decision that's one of the things that we've had some interesting conversations with clients about what that shift is about handing over some of that responsibility and really understanding what are the nuances of the decision-making process because that can be one of the harder parts of getting to a agentic operation, which is really understanding what is it that the human side of this equation understands either intuitively, the sort of level things. And you can sort of see -- do you have enough insight? Do you have to actually think about second order issues of how fast is a temperature going up, all these kinds of things that human operator may have just sort of put together somewhere internally. And that get feel for it.

Cody Falcon   00:12:38

That's a classic case, Eric. In fact, we were just talking about this in my team just last week, right? If you -- I think about our process operators, right, and our clients. If you go into a plant or a manufacturing facility and you go into the control room and you look at their operator notes, and let's say that Eric was the operator yesterday. And you could see Eric making decisions and changing things. I can promise you, inside those logs, you will never have any insight on what was going on in Eric's head while he made those changes, right? And so why is the expectation that we can explain, right, of how agents think and reason and logic? But that's the expectation we have, right, for transparency and trust in the system. But you're right, it's very, very difficult, right? And if you think about that operator example I gave you, right, if you go look and you shadow over the shoulder and you watch an operator work, and he or she's bouncing between trend charts and lines and PNIDs in a process control context and adjusting things, there is no -- there are no logs, right? There is no system of record inside that operator's mind and tell you how she's thinking and why she's doing that. So trying to programmatically define that, right, into an agent is very difficult.

 

Eric Hanselman

It's a challenge. And yes, you can learn from some of the control inputs, but we always get into that question of for all the situations we've seen in the training process, we -- hey, those who got covered, it's all of the ones that we haven't seen that we can react to. And I guess to that end, how do we start to balance some of the existing control capabilities we've got as we start to integrate more AI-driven functionality around this? And you talked about really sort of building on that capability, sort of marching up that level of control. How do we start to balance it? I mean, because again, one of the things we were talking about at CERAWeek was the idea that, well, geez, we've got all these controls already built in, in the control systems themselves, they'll take care of everything and not quite where we want to be heading with, I guess, some of those conversations.

 

Cody Falcon  

Right. So yes, I think at CERA week, we're talking about the kind of built-in guardrails, right? In process control context, I think about there's 4 or 5 layers, right? You have or like your safety system, your regulatory controls, your advanced process control, your model predictive control, your AI optimization and then these autonomous agent things sit on top of that, right? And I think at CERAWeek, we said, hey, look, an agent should never be the last line of defense or the last decision-maker, right? You have a lot of built-in safety and guardrails all the way up, right, that keep that agent in check. And I think on the panel, it was a great contrast because I think we had a colleague from Schlumberger, who gave an example of how hard it was in his space, right? Because in our space of process control, you know the physical design and safety limits and these protection mechanisms are kind of hard coded in there. If you're interpreting 3D seismic in his case, right, and you're trying to bet on where to place a land lease maybe for the next well, that is very unconstrained and difficult, much more difficult to control than how we think about it in a process context. So I think the layers, right, safety, regulatory, APC on up, that helps a bit. And on the human oversight piece, I think that humans tend to scale inversely with the risk, right? The higher likelihood or the higher risk the operational decision is, the more humans will be involved, right? Lower risk things that are easily reversed, agents will take those up first naturally, right? So that's how we see kind of the adoption. And the last quick thought on here is we talk a lot about human on the loop because we see scaling challenges with human in the loop, right? And so if you have a human in approving every single decision times thousands across the fleet, for example, multiple sites, you reach the limit, right? So we have kind of human and supervisory mode, human on the loop as being more realistic than human in the loop at every decision.

Eric Hanselman   00:17:02

Well, because scale is one of those -- both big advantages and one of the challenges to manage -- but the extent to which you can have automated systems that are doing things like alert triage and consolidation. You got a whole bunch of alarms going off. How do you have a system that makes sense of how do you prioritize it, which one do you look first and managing things like the human operator saturation problem and those kinds of things that helping prioritize can be relatively low risk. Absolutely and really high benefit.

Cody Falcon   00:17:42

Yes, absolutely. That is a classic AI case is alarm rationalization, right? It's the needle in the haystack, we're trying to remove some of the hay, right, to let the operators focus on what really matters, right? And I can't tell you the number of control rooms that I've been in where the operators -- the panel is just alarming, alarming, alarming. And the operator is looking at me having a conversation and just hitting the space bar to acknowledge them all, right? -- because the tone is different, right? You can tell by the audible tone how severe it is. And there's a real problem there, right? Where can you bring AI into that rationalization in real time to kind of weed out all of the noise and ultimately tune the system over time to where you get less noise and only the most relevant things that need your attention are flagged up.

Eric Hanselman   00:18:28

Well, maybe we do get to a clippy version of autonomous things, except that maybe it's a clippy sort of popping up and saying, "Hey, it looks like overpressure in that vessel there. You're right. It's due to a pump failure. Yes, you got a whole bunch of arms going on over there, but I trace them all back to this one pump that's really not working well.

Cody Falcon   00:18:48

Yes, you're spot on. You are spot on. I think we like to say we like to treat AI, whether -- I mean I'm wearing my vendor hat or wear my operator hat. We treat AI or agents like new operators. right? I mean they have to earn trust through performance, not claims, right? So when you onboard a new operator into a control seat, right, that new operator takes months and sometimes years of observation and they ramp up and get greater and greater control. That's kind of where we're at with Applied AI and agents in the space, right? We treat them like a new operator and make -- allow them to earn that trust through performance.

Eric Hanselman   00:19:28

Now how do you typically track that kind of progress? Is this something where you're looking at agent versioning, retraining for some of the models that are driving them? How do you watch that progress and help to actually ensure that you're actually tracking the trust that gets built?

Cody Falcon   00:19:46

Yes. So there's a couple of ways to do it, right? So I think at the very top level on the modeling front, there's an entire practice around machine learning ops or MLOps, right? How we -- every deployed model is constantly monitored for performance and degradation and alerted, pulled out, retrained, redeployed, et cetera. So there's an MLOps practice around modeling, trust and KPIs in the system. I think one of the most common ones that we see in the autonomous work is the number of human interventions, right, or maybe the number of manual interventions, right? And you can measure that, right? If you look at the end of every month, how many manual operator interventions did you have in the system across all shifts, right? And you start applying the technology, and you can see those numbers fall over time. So that's one way to do it. A second way to think about is how can you interpret reinforcement. And so for example, if I give you a recommendation, it highly unlikely that you're ever going to give me any sort of feedback. We see that in the commercial tools out there. They don't get much feedback. But one of the ways that you can do this is if I give you a recommendation and then I see you take my recommendation within 30 seconds, it's a pretty safe bet that you trust it or you acted on my recommendation. And then you can monitor the outcome. Did it stabilize the process? Was the recommendation correct? And you can kind of double check that and then bring that back as reinforcement into the model, right? So there's some ways that you can do it without having direct feedback. But I think at the top level KPI, what's most common is number of operator interventions, manual interventions, maybe the human-to-alarm ratio is another one you can look at. But then behind the scenes, I think there are some things you can do to look at the actions taken on the agent's recommendations.

Eric Hanselman   00:21:38

All a matter of actually keeping track to ensure that you can understand where you are from a trust perspective.

Cody Falcon   00:21:44

That's right. That's right.

Eric Hanselman   00:21:46

So as organizations are looking to be able to get more autonomous, what should be their priorities for next steps in terms of being able to build that? We talked a little about data. Are those primary focus areas? What do you think are really the better next steps on this?

Cody Falcon   00:22:02

Yes. Look, I think the first thing -- and of course, we have our own assessment practice and framework that we do with our clients. But I think the first one is an honest readiness assessment. If you think about that Level 0 to Level 5 levels of autonomy, right, an honest assessment on where you are, let's just say you're at Level 1, and then there's a readiness component, right? To get to Level 2, I need a data foundation as an example. Do I have that in place? Do I have the data, the controls, the cyber, the governance the people in place to do that, right? And so that's my first step, right? There's an honest assessment of where you are. And then second half of that is a readiness assessment of how quickly you can get to the next level and what is required to get you there.

Eric Hanselman   00:22:49

And it's that process that we go through in any technology transition, right? If you're going to figure out where you're going, you got to figure out where you're starting from accurately.

Cody Falcon   00:22:58

Figure out where we're at first.

Eric Hanselman   00:23:00

No, you go.

Cody Falcon   00:23:02

And then I think from there, Eric, it's probably -- it's similar to the digital transformation track we were talking about years ago, right? It's from there, once you have the assessment, pick 2 or 3 bounded use cases or maybe applications that you can get deployed to start building that trust and start gradually moving up that curve, right? It's not 0 to 5. It's not an enterprise moonshot. It's I might level 1, I can get to Level 1.5 in 6 months by deploying, let's say, APC, right, on some of the critical processes, right? That's how we look at it. 2 or 3 bounded guaranteed or safe wins that build trust to keep climbing the curve.

Eric Hanselman   00:23:48

So much of AI adoption in general is about ensuring that you've defined the use case well and you scoped what you're trying to get done in something that's reasonable and actually making sure that you're -- in many ways, you're not biting off more than you could choose.

Cody Falcon   00:24:05

Yes. And one of my final recommendation, maybe step 3, right? And of course, it's a bit biased coming from me, but I would recommend you partner with people that understand or maybe own both the OT and the AI stack, right? What we see is if you bolt on AI and things around the outside, it becomes very brittle over time. And someone has to maintain those integrations and dependencies for the life of the asset, right? And so I think there are some leading players out there that understand both sides of that and kind of integrate the OT with the AI stack altogether, that's a natural starting point, especially if you already have those vendors in play, right? Those vendors may already be running your process or your facility or your plant, that's a good starting point if they have both sides of this equation.

Eric Hanselman   00:24:56

Well, I mean, the operational technology aspect is -- we are in an age in which it's so tempting to just be able to throw technology in various forms at a number of problems. And yet operational technology around the OT world is one that operates in different spheres, different context, different time scales. These are areas that those operational considerations have to be such a key part of that process of moving forward. However tempting the technology pieces are.

Cody Falcon   00:25:29

That's -- you're spot on. And now it's interesting, you see a lot of -- including us, right? You see a lot of the major OT players collaborating and partner with leading technology players to bring that latest and greatest in, whether it's the hyperscalers or NVIDIA or others, right, the leaders in this space, you see these partnerships and collaborations coming together with the OT players, bringing that technology to bear into the OT landscape and offering both sides of that equation, right? The OT as well as the AI stack.

Eric Hanselman   00:26:00

It's a combination that can be powerful when it works together well. That's right. You just have to make sure that you're considerate in terms of how you put it together.

Cody Falcon   00:26:08

That's right.

Eric Hanselman   00:26:09

Well, this has been great, Cody. I appreciate all your insights. It's been great to be able to continue this conversation.

Unknown Speaker   00:26:16

Absolutely.

Eric Hanselman   00:26:16

There are so much more that we can head towards, but we are at time for this episode. But thank you very much.

Cody Falcon   00:26:22

Very cool. Thanks for inviting me. Great conversation, Eric.

Eric Hanselman   00:26:25

Well, it's great to have you on. I look forward to talking further.

Cody Falcon   00:26:28

All right. See you soon.

Eric Hanselman   00:26:29

But that is it for this episode of Next in Tech. Thanks to our audience for staying with us, and thanks to our production team, including Sophie Carr, Feranmi Adeoshun, and Kyra Smith from our marketing and events teams. If you liked this episode, please like and subscribe. Please keep in mind that statements made by persons who are not S&P Global employees represent their own views and not necessarily those of S&P Global.