Podcast: Best practices for implementing AI agents in manufacturing workflows

Below is an edited excerpt from the podcast:

PS: You’ve got sweeping responsibilities at IBM, especially with Maximo, and I’m so glad you're here. We could talk for ten podcast episodes, I’m sure, but today we’re going to focus specifically on data-driven asset management—especially AI.

So let me ask you the first question: when it comes to data-driven approaches, what do you consider best practice for the kind of people we’re speaking to today—the plant managers, operators, especially those in the reliability function?

CN: You know, I would say it all comes down to data. As you just mentioned, data is a huge, huge topic—and a huge challenge. As we start to talk more about AI in this conversation, you’ll see that data is at the core of everything that’s happening today.

But it is a huge challenge. I think in the world of manufacturing—and just in general—we’re collecting more and more data and information, whether it's around assets, vibrations, defects, temperatures, or the work people are doing with work orders. There's so much information out there. Some of it is structured, some of it is unstructured—like notes or other formats that are difficult to reach. 

I kind of see it as three challenges, and they all start with the letter A. First is accessibility—sometimes it’s just not easy to get to the data. Second is actionable—you might have all this data, but what do you do with it? What does it mean for you? How does it provide any value to make it useful? And third is accuracy—we see a lot of issues there, whether it's someone entering incorrect data because they misunderstood something, or just not bothering to enter it at all. We hear about this all the time when talking with customers—you have these work orders where someone just selects a generic dropdown, like "N/A," and it’s just not super helpful when you're trying to find patterns, or some trending thing with an asset, and you can’t see that information because somebody forgot to put that data in there.

So, yes, there are a lot of data challenges with data as a whole. Having the right product—obviously, the right software product—having the right process, the right tools absolutely helps. Having the right people in place to make sure it’s really happening helps.

We believe that you get a lot of value as well tracking from beginning to end, so within my world, we talk a lot about asset and asset management, but we’ve really transitioned the conversation to focus on the entire lifecycle of the asset. From the moment you’re planning what you need and the financials around that, to receiving the assets, tracking the assets, seeing what’s happening around them, the work around them, everything all the way through to disposal—each one of those steps gives you additional insight.

We’ve started to transform our conversations to be more holistic, as opposed to just managing assets. And we feel that leads to a good foundation for AI, because AI is going to be heavily dependent on having good data to work with.

PS: There are certain sectors, like power generation, that require five-nines reliability. I’m curious—from your perspective, how are you seeing this approach to asset management permeate other market sectors, like pharma, food, or utilities? Are they all taking cues from power gen, or is it  moving through sectors at different rates?

CN: I feel like the impact of AI is hitting all industries. The challenges with data—and the regulations around data—is hitting everywhere. We’re all realizing there’s so much untapped value. There are opportunities for optimization, for efficiencies, for safety. And there’s a lot of opportunity for things like optimizing power consumption. There’s so much untapped opportunity that we feel it's transforming every aspect of life that we’re looking at, every kind of business.

A couple of examples of that would be to combine power generation and healthcare, where you look at the facility management. There’s a lot of questions around, can we do maintenance better? Can we do compliance better? Are there ways to optimize energy usage? How do we look at new ways to improve occupancy and repairs? 

It’s all rooted in data, in looking at patterns, at making insights actionable. A lot of that is going on regardless of what industry you’re looking at, and there’s so much opportunity based on what’s happening from a technical standpoint—with AI helping as a tool to get there.

PS: So on this podcast, we’re here to talk about one specific flavor of AI for the next couple of questions: AI agents. Before we jump into how they’re applied, can you explain what AI agents are exactly? What kind of flavor AI is it, and how are they moving into manufacturing?

CN: Yeah. If we go back to the idea of a traditional AI idea, what it does it it’s looking at the world as a whole, it’s looking at leveraging basic queries—basic questions in a natural language—and interpreting them for people. In manufacturing, it could be having the ability to use computer vision to identify defects and flag problems. AI is built for things like that—it can look at the world around you, and say hey, there’s this thing here that may be an anomaly, or hey there’s this thing here that may be a little bit different from what we expect.

The next generation that we hear quite a bit about is generative AI. That’s about creating new artifacts from what’s out there today. It’s leveraging algorithms and it’s creating new content based off of patterns from existing data. When you think of things like ChatGPT, it’s doing that same thing, it’s looking at lots of text, it’s looking at patterns of “when people say thing kind of thing, these kinds of responses are what’s expected.” It’s building off of recognizing those patterns, and figuring out what the next pattern-step is. 

AI agents then piles on top of that and it goes “OK you know you have all this stuff that's generating new content. You have the ability to query what's out there. Everything is very forward-looking, right? It's one kind of interaction where you ask it certain things and then it just kind of gives your response. But there's no ability to go back and actually look at how you revise or improve, or add on to. There's not really a whole lot of opportunity for that. With AI agents, what it's doing is it's actually letting you stack on top and revise and build. 

An example would be if you're creating a marketing plan or an essay. If you were to use GenAI today and you say, “I want to write an essay about my job, and it spits out some content for you. But if you wanted it to be a lot more accurate and a lot more specific, you'd want it to say, “well, what kind of jobs do you have? What are some statistics or interesting things about your job that would enhance the contents being created?” And you have like a version 2 and a version 3. That's where something like an AI agent is helpful because what it does is looking at what needs to happen. It's giving you multiple opportunities to add on to the output and iterate on it, and create something new and better from it. 

If we tie that into what we do today in the world and in manufacturing, you can think of agents as, “hey, I am doing a certain kind of work,” and the agent goes “I recognize what you're trying to do. You need to request a work order.” And it can go out and request the work order, create the work order for you. It maybe can look at the type of work that you're doing and say, “hey, I recognize that we have other kinds of work orders or similar to yours. Let me make some recommendations for you on what you need to populate.” Or, “I see what you're doing and it looks like it could be tied to this error code or failure code.” It starts to have a multi step process on providing you additional value and additional recommendations that you wouldn't have been able to do with a simple AI query.

PS: Interesting. It sounds like part of what you're saying, too, is that it's helping extract certain aspects of tribal knowledge which may otherwise get trapped out there with frontliners, as the AI agent itself seeks to improve the quality of the work plan or the work order itself by asking the right questions.

CN: Exactly. I'm going to go with the art of the possible for a moment, but eventually, what's really fun about the agents is that they're not bound to one application or one solution. They're really operating within a model of a business case or a use case. When you're thinking about it from the standpoint of, you have knowledge of the inputs, you have knowledge of what agents have access to (and this is from a more back end perspective), and you have knowledge of the type of output you want from it, you have multiple agents that support that. 

An example would be, let's say that you run a renewables energy plant, and you're dealing with solar panels. And let's say that in your system or solution you're able to capture data and patterns of data around power generation. So you know typically what happens during the daytime, you know what typically happens at night, you know what happens during certain times of the year. And there's impacts depending on weather and other things. Now all of a sudden an AI system, an agent is monitoring and it notices weirdness happening in the pattern of what's expected for power generation. Now what it does is say “I see something that's not right. I know what the pattern should look like, and it's not quite there.” And then what it does is it calls another agent and says, “hey, you need to check out the solar farms in this location in California, something's going on there.”

So then that agent calls to a drone, and a drone is sent out, does a scan of the location, brings back information, calls to another agent that says “I'm now going to do assessment on the video that was captured. I see that there is an issue with whatever panels, they have dust on them” or whatever it is that’s impacting its ability to really generate as much power as expected because a dust storm went through. We're also able to look at weather information. Then, it calls to another agent and it says “create a work order for me. Find an available technician. Here's where there's a problem.” That person is going to be available within a given time period that we need, and then that person receives a text or a phone call, and they're told, “go out there and fix the problem.”