Aired Sep 29, 2026 · 25 min

Listen on: YouTube · Spotify · Apple Podcasts

What could AI agents actually do for a utility today? Power Perspectives host Matt Chester talks to AI strategist Jeff Wilser about the work beyond the buzzword.

They explore how AI agents could search years of filings, and turn interviews with retiring employees into searchable institutional knowledge. Jeff also explains why human judgment remains essential when mistakes carry real consequences for customers and the grid, and highlights the limitations of the technology.

For utility leaders ready to experiment, he offers practical starting points that you can start implementing today even if you’ve never used AI agents before.

Matt Chester: Agentic AI, you've heard the term, but what is it really? And how is it already changing the way that energy companies are working? Our guest today on Power Perspectives is AI strategist, Jeff Wilser, and he's here to answer those questions for us. Jeff, welcome to the show today.

Jeff, let's start with the basics. When you say agentic AI workflow, what exactly does that mean? And how is that different than say, someone at a utility company using a chat bot to ask a question?

Jeff Wilser: Sure. I know you've had a lot of recent great episodes on AI. So I'm sure many of your listeners have some basics here. But for those who are new to the agent world, I think when most people think of something like ChatGPT or Chatbot, it's mostly about just asking a question and getting a response, right? So simple analogy, hey, I want to go to Tokyo for a vacation. What do you… would actually do the things for you. So instead of just getting an answer, like here is a good itinerary in Tokyo, you can say, hey, I want to buy plane tickets. Please buy me plane tickets, book me hotels. So you can think of it as actually the arms and legs of doing things, not just the brain itself.

Matt Chester: Gotcha. No, that's definitely helpful. It's a little bit more action-based. And for our audience that's in the utility and energy world, what's an example of something that maybe one of their employees spends hours doing today that an AI agent could realistically handle?

Jeff Wilser: Absolutely. And Matt, I think a lot of folks hear about AI agents and they immediately race to the most provocative, scary, crazy thing that can be done. Like some agents who has full autonomy of a control room. But there is a lot of middle grounds and a lot of far less glamorous, more basic things that can be done before you get to that point.

I think the number one thing, and this is true, so many industries and top of utilities and energy is paperwork. Virtually every industry, including you guys, we all have more paperwork than we want to do. And that ranging from compliance documents to trying to find what your rates were in some niche area from 12 years ago, there are buckets and buckets of paperwork that no one really has time to do. And AI agents can be built to actually sift through the paperwork and summarize and actually draft documents. So kind of the two-year-ago and one-year-ago AI thing was, was to manually, let's say, feed a PDF into something like ChatGPT or Claude or Copilot and say, hey, summarize this for me. Now that could be done with far more efficiency where agents can actually romp through stacks and stacks of paperwork and glean insights and help you create summary documents.

Matt Chester: And is the benefit just that it's able to do it that much quicker and more efficiently? Or do you find that the AI agent can actually do it better too?

Jeff Wilser: So in your world, think about a compliance issue that no one analyst could possibly have time to read the hundreds of precedents and documents behind this, but an agent can go through. And not just do it faster, but have access to this sprawling amount of paperwork to turn up insights you wouldn't otherwise have known to even look for.

Matt Chester: Yeah, that's definitely, I think, something that the utility executives would be eager to learn more about. And you mentioned that people jumped to the idea of like, oh, the AI is running the control room. And then there's a little bit of an aversion to that. But even when you're just looking at the paperwork and the regulatory aspect of it, it's still important. And I'm curious how much you find that there's oversight or fact checking still needed from the human side of the equation when the AI agents are doing these tasks.

Jeff Wilser: For sure, there's a cliche that I'm sure you've heard, but it's one of those cliches that happens to be true, and that's keeping a human in the loop. And that can happen in a couple of different levels, right? You could have the human spot check and look for errors. Obviously, you want someone to be sanity checking it before it goes anywhere sensitive.

But one thing I think that one thing I share when I work with companies and workshops is, yes, human loop. But one additional thing you can do to help with control and help guardrails is actually using AI agents and AI models kind of to cross-check each other in an adversarial, constructive way, right? So imagine that you, to keep it very simple, imagine you're working with something like Claude to create some agentic workflow and your agent says, aha, I am now going to create this compliance document. You can then feed it to a different system entirely. Let's say in ChatGPT, give it the rules of the road, give it the context and say, hey, I want you with fresh eyes and to get the nice splash cold water here, cold water, I want you to evaluate this and find blind spots. So by using AI system to check the other, that can help actually determine reduce errors and give yourself more of a more, um, again, I wouldn't use the word certainty. A human should always be in a loop, but it's helpful to have AIs check each other and that can reduce errors as well.

Matt Chester: Is that one of the go to solutions that we hear about hallucinations or just, you know, faulty conclusions sometimes come from the AI? And, you know, the natural inclination is to have the humans be the one checking that, like you said, spot checking. But do you find that it's a better practice or almost time saver to have that, you know, AI checking its own work?

Jeff Wilser: For sure. It's, again, not to take the human out of this, but it's like one additional level of protection before the human, right? And it's also, it's kind of a lot of fun too when you have these two different models and you almost surprise them and say, hey, by the way, I asked Claude the same question here's what it found, and you give it its answer, that will almost, is a way to spur the first model to make sure it's buttoned up, and to actually second and triple and quadruple check itself.

Matt Chester: I know in the utility sector, we're in an environment where these companies are seeing massive liabilities with things like wildfire management, outages that impact vulnerable customers. The utility leaders don't really feel like they can afford to let a mistake through. And this is great advice for how to capture and minimize those mistakes. But on the chance that there is an AI agent involved and a mistake does get through, who ultimately is held accountable when something goes wrong?

Jeff Wilser: There's one group that is definitely without question will benefit from AI, and that is lawyers. There are going to be so many gray area legal questions about who ultimately is accountable. That is a legal question beyond my pay grade as far as the actual liability. That said, I think the North Star, the way I view all of this for your industry, is something that I call the red line theory and where you are on the red line. And what I mean by that is what I found is the most successful way companies can use AI, whether in utilities or in publishing or in healthcare, is to think clearly about where is the red line of what you hold sacred that AI should not touch, that humans, to your point, Matt, ultimately must be accountable for.

So for one example, in media and editing and journalism, Time magazine, they said, okay, we're not going to use AI to write AI slop to write our articles and do our research. That is still the work of humans. But the chief operating officer of Time, who I interviewed recently, what he told me was, yes, the AI is not going to replace the human writers. However, that's a red line. That's what Time is sacred for. That is their sacred editorial judgment and credibility.

On the other side of the red line, they want to use AI ambitiously, and they created AI agents to do far more things that cannot be done with time. Now they have AI agents and AI bots that let you interact with Time Magazine and explore its 103 year history in ways you couldn't do before. They added new functionality. The goal is not to take humans out of the loop of using judgment with respect to the wildfires and all the tricky questions you're facing. The goal is to give humans more time to think deeply about edge cases, about the sensitive issues, by automating the more drudgery stuff, the mundane stuff, to give humans more time to do what they are really good at.

Matt Chester: Yeah, that's interesting. It makes me think of, we ran a survey a little while back asking utility executives, what do they think AI will not replace in the utility sector? And the main tranche of answers came into, one, there's the physical work of repairing a line and that much was obvious. Oh, the robots are coming. I'm sure the robots are coming imminently and they'll do all of that. Eventually, yeah. But the other aspects that were cited multiple times were empathy for the customers and, you know, having kind of that human connection. And then the sort of judgment that leaders have to make based on all the information that AI can help collect for them. But ultimately, a human has to make, you know, weigh the pros and cons and the different tradeoffs and make that ultimate decision.

Jeff Wilser: Yeah, well said. And along those lines, think about flying and aviation. That's similar to your world, right? The stakes just as utilities. When you're flying a plane, human lives are quite literally at stake. When autopilots, when that technology developed years and decades ago, pilots were understandably skittish about it. They did not want to trust flying to a computer. Like, fair. That is an understandable reaction. They also worried they could be replaced by robots and by self-flying planes. And what happened was not that pilots were replaced or that pilots stopped using their judgment. What happened instead was pilots, they still had to know how to fly, but they still made the judgment calls, but they're actually able to make their role more about the judgments as opposed to doing more of the manual things.

I think utilities can be looked at the same way, right? Ultimately, it's the human's job to apply the judgments. And if we can automate some systems to make that job easier, great. But I would also encourage utilities to think about it from the lens, not just of automating and cutting, but even growing. And by growing, I don't necessarily mean doubling revenue. But rather, what else could you do that you couldn't do beforehand? That's a big light bulb I see when I work with CEOs about this across industries is most companies and most organizations are thinking very narrowly about, okay, maybe I can use AI to cut this one thing or to automate this. And sure, automation is great, but it's also just scratching the surface of what you can do.

I'd encourage everyone listening to your podcast to do is to take a moment and imagine, let's say theoretically, your headcount has doubled without spending an extra nickel. What then would you be able to do? What additional safeguards could you create? What additional services? Where are you able to add more value if your head got doubled? And then think, okay, how can AI allow us to actually get some of that new resource and capability? Okay. So I think that's one way that people in utilities can use AI, not just for like saving off a bit of cost at the margin, but to more ambitiously get more done in your organization.

Matt Chester: Yeah, that's a terrific thought experiment and how to approach it. Yeah. Kind of going back to the drudgery of the paperwork we were talking about earlier, you know, I know a longstanding challenge in the power sector is siloing, you know, information processes. They're always separated between different departments and the right hand doesn't always know what the left hand is doing. So given that part of that problem is that utilities have so much siloed data buried in PDFs, filings, engineering documents, old records. How close do you think we are to a world where, you know, through an agentic AI, an employee can essentially have a conversation with all of that institutional history?

Jeff Wilser: I think it's as close as you want to be. The technology exists right now, and it's not something that requires billions of dollars of investment. It's mostly about that. The irony, Matt, is I think this is one thing that's kind of lost in the conversation. There's so much AI hype out there and all these buzzwords. Even the word agentic AI can sound kind of magic wandy. Ultimately, an AI agent is mostly about context and context files. And by context files, kind of fancy way of saying really just like, it could be as simple as text documents. You mentioned PDFs, right?

And so what's making or breaking AI adoption, both in the power industry and most industries at large, is not about are you using the most cutting edge, powerful frontier model. It's less about the model's power and more about how well are you feeding it context. What I mean by that is, as you point out, Matt, there could be PDFs living in, there's somewhere in IT, there's all kinds of PDFs. Somewhere in procurement, in marketing. Every different division has their own weird warehouse figuratively or literally of all these different files. Getting those files, having a smart infrastructure that is feeding the relevance files and takeaways from the files into a system that AI agents read is, that's the whole ballgame.

Matt Chester: The other area where, you know, utility leaders traditionally see knowledge spread out is where it's not written down at all, but it's stored in the minds of longstanding employees. And this is a common topic in the utility sector as much of the workforce is aging closer and closer to retirement. Are we risking losing that knowledge? I'm curious what you think AI agentic tools can do to capture that institutional knowledge and maybe re-reporting plug in that gap where, yeah, hey, there might have been that 40, that employee who's been there for 40 years, who knows, oh, this is how this substation really behaves. How can we set up the agentic AI to be able to understand that as well?

Jeff Wilser: Yeah, I'm so glad you brought that up. That's a great one. I think you're right. It's super relevant to the industry here. And some might call it tribal knowledge that is a danger of being lost. And I think it's one often overlooked use case of AI agents. And one of my favorites is technology interviews and actually having... Now, this could be done a couple of ways. You can set up an interview for almost like a series of prompts that this salty engineer about to retire can then answer, crucially, answer verbally, right? If it's... If we're asking for type response, that's not going to be nearly as effective. No one likes to type a 5,000-word essay, right? But if you ask them questions about things that, frankly, they'll probably be kind of flattered. They're asked about, like, oh... Yeah, I recall well that ice storm of 1987, and no one's thought to ask me that in years, right? And have them go through and give verbally detailed responses that can create this library, this corpus of knowledge.

And here's where agents really come into play. In a world before AI, if you were to, let's say, spend, and I'm making this roadmap up, but bear with me here. Let's say you have maybe two hours a week you devote to interviewing this, the salty veteran engineering knowledge, right? So over the course of a year, you have a hundred hours of them talking about all kinds of esoteric things they've seen. And do that at scale. So you've now interviewed, let's say, let's say there's thousands of hours of conversation of these folks. In a non-AI world, good luck getting anyone to go through that. It reminds me of the last scene of Raiders of the Lost Ark with the crates and crates and crates in the warehouse and all these things. No one can go through those crates in the warehouse and make sense of this. The agents can go through this, at this point, this massive library of interviews you've conducted with all these folks who are about to retire and make heads and tails of it.

It could synthesize things. You can actually have a few different cuts of this library. You can say, hey, I want to create a masterclass about ice storms or about substations and how they reacted in this kind of emergency. And you can create an entire documents and dossier around these thousands of interviews. And then once you have that effectively, this massive database of tribal knowledge in one place, in the future, you could ask AI agents to search into it, to probe, to ask questions. A more advanced thing would that mean to create an actual, you can create effectively your own chatbot based off of this database to let you real-time ask questions to it as well. So I think we're going to see, that's still relatively fast, fringe and bleeding edge use case right now. In the future, my prediction, that sort of use case will go mainstream across many industries.

Matt Chester: Yeah, that's fascinating. And I definitely think that the utility sector is primed for that. Like I said, the aging workforce, this long institutional knowledge, it's too valuable to just let it go. So maybe the AI agents are coming at the perfect time to fill in that gap that's so needed. But I guess as we wrap up our conversation, Jeff, I want to ask the advice-based question. If somebody's listening in, they're running utility, they want to experiment with agentic AI this week, not two years from now, not those fringe cases. What is the first workflow you would tell them to try and how would they know whether it's actually working?

Jeff Wilser: One thing that, and if your listeners already tried this, great, move on to the next, but I think something that everyone busy in the utility industry can deal with. But before you think about how can we inject AI in most sensitive workflows of our power grid, whether we like it or not, we all have busy inboxes. We all struggle to make the meetings on time. We all have a hard time keeping track of our to-dos. I would create your own chief of staff, your own AI agent who is able to triage email for you, who is able to hoover up all the projects you have. Most systems now let you easily connect in your email, company Slack, to-do list systems, and so on.

And crucially, my advice to even be more specific is to give it lots of context. Spend an hour walking through what's important to you, what your goals are, what are risks you're worried about, and have it then each morning, give you a briefing for the day and say, hey, what are the three most things I must do today? What's most important? This, I think that's a, one reason I like this starting point so much is that one, it is very de-risking, right? You are not putting the grid at risk. You are starting small with your own personal workflows. And two, I think most people will get a ton of value out of it. When I work with CEOs and show this in workshops, teams are like, oh my God, it is amazing how a little bit of agentic systems in your own workflow can handle so much busy work.

And one more bit of very practical advice that I think listeners will benefit from is when you're working with AI, in this manner. The way to speed things up and the way to give it even more context is actually through voice, not through typing. Reason why is that almost everyone, certainly including me, despite typing for a living, is much faster at talking than typing. It's also more enjoyable to talk than to type. And so when we're going back to when we interviewed the salty 35-year veteran about to retire, they're going to be more likely to share lots of rich details if they're just casually, conversationally chatting. If they're typing, they cut corners, but the corners are where the good stuff lives. The corners are the edge cases. That's the nuance. The more nuance you give, the better. When you're talking, you're more likely to just share a random parenthetical the AI will actually store that parenthetical and that's useful context in the future, right?

And so, and last, not to be overly mechanical here, but when you hear voice mode, people might be tempted to actually turn on the kind of like voice mode in ChatGPT or Claude, or mainly ChatGPT. That's not what I mean, because that, when you go back and forth with voice, almost like Siri, and we all know how bad Siri is, you kind of only get the lightest, most surface level version of the AI. It's not the more powerful thinking versions of the model. Instead, what I mean is you can use a dictation service. And for those who are still not sure exactly where to start, literally after you finish listening, Matt, to your great podcast, they can go in and say, hey, chatbot of choice. I want to create my own chief of staff AI agents. Show me how to do it. They will show you how to do it. And the last bonus thing they can do too is say, because this is moving so fast, AI is evolving so fast they should say in this prompt, also go and do some latest research to make sure you're up to date on what's possible. Because the irony is, Matt, AI is evolving so fast that a lot of these chatbots don't even know what they're capable of because the training date is cut off and things have happened in the last two months since training data. So that's where I would start and you will get your agentic system before you know it.

Matt Chester: Jeff, again, this is a topic that's been on the top of everyone's mind. You mentioned we've been covering AI on the podcast a lot, but you definitely gave us some new insights to chew on. And we thank you for joining us today.

Jeff Wilser: Thanks so much. Had a great time.

Matt Chester: And as always, I also want to thank our audience for being with us today on Power Perspectives. And we'll catch you next time.