Episode 5 – AI in Utilities: Moving from Pilot to Production

Marc (00:44)
Today’s episode is titled AI in Utilities Moving from Pilot to Production. Utilities are under increasing pressure.

Pressure from multiple directions. Load is growing again, driven by data centers, industrialization, and electrification. The system itself is becoming more complex with distributed resources and new technologies. And at the same time, utilities are managing workforce constraints and an explosion of data from AMI systems and other sensors deployed in their infrastructure. Against that backdrop, there’s a growing conversation about how artificial intelligence, or AI, can help utilities operate more efficiently and make better decisions. But while many utilities have experimented with AI, far fewer have successfully scaled it. For many, AI still feels more like experimentation than execution. So the question becomes how do you move from pilot projects to real sustained value from AI? To explore that, I’m joined today by Jon Kerner,

A partner and AI expert at ScottMadden, who works closely with organizations on successfully applying AI and advanced analytics. Jon, let’s start with the big picture as we like to do. AI has been talked about for years, but it feels like the conversation has begun to shift recently. Why is AI becoming more relevant for utilities now than it was even a couple of years ago?

Jon (02:20)
A few things came together at once, and that’s really the story. For years AI was a slide in a strategy deck. Interesting but not urgent. What changed is that new industry pressures and data raw materials showed up at the same time.

On the pressure side, as you mentioned, load is growing again after two decades of being basically flat.

Data centers, manufacturing coming back onshore, electric vehicles, and building electrification, all of it landing on the same grid. And a lot of that grid was built for a different era. At the same time, the system is getting harder to run. Solar, storage, and EVs have turned what used to be a one-way system into something that flows both directions and changes by the minute.

On the raw materials side, the AMI rollouts of the last decade quietly created an enormous amount of data.

A utility that installed smart meters now has interval data on every customer, every 15 minutes. Most of that data has been sitting in a data warehouse doing very little. AI is what turns it into something you can actually act on. And then there’s the workforce, a big share of the people who understand how the system really behaves are retiring. And you’re not replacing that experience one-for-one. So you’ve got more complexity, more data, and fewer of the people who understand how this works.

That combination is why this has moved from someday to now.

Marc (03:52)
So there’s a lot going on, for sure. If you had to pick one factor, what is the biggest driver that’s making AI relevant right now?

Jon (04:03)
If I had to pick one, it’s the workforce, but not the way people usually mean it. It’s not just that crews and engineers are retiring. It’s that the knowledge is walking out the door faster than the data and the tools are filling in behind it. You’ve got more information than ever, and fewer people who know what it means. AI is the most practical way to close that gap, to take the judgment that used to live in a few experienced heads and make it available to everyone running the system.

Load growth and complexity make that gap matter more, but the gap itself is the driver.

Marc (04:40)
When people hear AI, it can mean a lot of different things. It’s become a term thrown around for a a broad category. We also tend to use the term more broadly at ScottMadden. When we discuss AI with utilities, what are we usually talking about in practical terms?

Jon (05:01)
Honestly, most of what gets called AI today is not new to utilities. Strip away the label and there are really three things. The first is analytics and machine learning, which is just finding patterns in data and making predictions from them. If you can predict which transformers are most likely to fail this summer, that’s machine learning. Utilities have been doing versions of this for years.

The second is automation, software following rules to do work a person used to do by hand. That’s not really intelligence, but it usually gets lumped in. The third, and this is generally the new one, is generative AI. The technology behind tools like ChatGPT. What’s different is that it works with language and documents, not just numbers. It can read a thousand-page rate filing, draft a first response to a regulator, or pull the one relevant clause out of a maintenance manual in seconds. My advice to leaders is don’t get pulled into the vocabulary. The question that matters isn’t “is this AI?” It’s “what problem are we solving and does this help?” We do see a lot of teams start with the technology because it’s exciting and then go looking for a problem to attach it to. That’s backwards and it’s one of the most reliable ways to waste money.

Marc (06:26)
Where are utilities already using these tools today, Jon? Even if they don’t always call it AI?

Jon (06:33)
They’re using it all over the place. Load forecasting is the obvious one. Every utility forecasts demand. And the better forecasts have used machine learning for a while now. Outage prediction and restoration is another. Using weather and historical data to stage crews before a storm hits. Vegetation management has quietly become one of the biggest users. Utilities are flying imagery over their lines and using software to flag which spans need trimming instead of sending someone to drive every mile of it.

On the customer side, your call center is probably routing and triaging calls with these tools. And anyone with smart meters is running analytics on that data, even if nobody in the room calls it AI. The point I’d make to leaders is that you’re not starting from zero. You already have people doing this work and getting value from it. The opportunity isn’t to invent an AI program from scratch. It is to take what’s already working in pockets and do it deliberately across the organization.

Marc (07:37)
That’s a great point, Jon. And the next question is where AI is actually creating tangible value to the business. So let’s make this concrete. In your work, where are you seeing the most practical and immediate value from deploying AI in utility environments today?

Jon (07:58)
Sure, and and this is where it gets real. I’d group the value into four areas, and they’re not equal. Grid operations is the first: better load forecasting, predicting where outages are likely so you can pre-stage crews, and optimizing voltage across feeders to cut losses. These are close to the core of what a utility does, and the payback is direct.

Asset management is the second, and for a lot of utilities, it’s the fastest win. You’ve got transformers, breakers, and other equipment with enormous amounts of money. And historically, you’ve maintained them on an inflexible schedule. Replace it every so many years, whether it needs it or not. Predictive maintenance flips that. You can use the equipment’s own data to tell you which assets are actually headed towards failure. So you spend your maintenance dollars where the risk is. That’s real money and real reliability. And the data to do it usually already exists.

The third is the customer and commercial side. Forecasting demand, segmenting customers so you can target programs, and reaching out proactively before someone’s bill spikes or before they fall behind.

The fourth is the back office, and this is where generative AI is changing things fastest. Think about everything in a utility that involves reading and writing documents: rate cases, regulatory findings, work orders, compliance reports, the call center knowledge base. These tools can take a huge amount of manual reading and drafting off people’s plates. It’s less glamorous than the grid work, but the value shows up quickly and the risk is low because a person is still reviewing the output.

If I had to point a leader at where to start, I’d say predictive maintenance on your high-value assets and the document-heavy work in the back office. Both have fast, provable returns, and neither one requires you to touch the control room on day one.

Marc (10:04)
That’s great guidance, Jon. How should utilities think about the maturity of these different applications of AI or as as we call them sometimes, use cases? What’s still in pilot mode across the industry and and what’s actually been scaled up and is now running in production?

Jon (10:25)
Market varies a lot by use case, and that’s worth being honest about. The things genuinely running in production across the industry are the most established analytics: load forecasting, meter data analytics, vegetation management from imagery, and predictive maintenance at the utilities that have invested in it. These have been scaled and they’re delivering.

The things still mostly in pilot are anything that touches real-time operations, anything fully autonomous, and most of the customer-facing generative AI. A tool that drafts a response for an agent to review is in production in places. A chatbot talking directly to your customers about their bill with no human in the loop, that’s mostly still being tested, and for good reason. The way we think about it is almost everything in our industry is still in one of the early stages. Either you’re experimenting to see if something works, or you’ve proven it works and you’re scaling it. Very little has reached the point where it’s fully built into operations and just part of how the lights stay on. It’s not a criticism. It’s just where the industry honestly is, and it should shape how you set expectations with your board.

Marc (11:41)
Where are utilities overestimating what’s ready for production now?

Jon (11:47)
The big one is real-time operational decisions. A model that’s right 90% of the time is fantastic for deciding which transformers to inspect first. That same 90% is nowhere good enough to let it make switching decisions on the grid itself. People see a strong demo and assume it’s ready to run unsupervised in the control room, and it almost never is.

The second is fully automated customer interaction. The technology can hold a conversation, but in a regulated business, being confidently wrong about someone’s bill or their service is a real problem, not a minor glitch. And the third, honestly, is the gap between a demo and a production system. A prototype that works in a conference room on a clear sample of data is maybe 20% of the work. The other 80% is making it run on messy real data at scale, wired into your other systems every day, when nobody’s watching. A lot of utilities see the demo and think they’re almost done. They’re not. And underestimating that gap is exactly what kills pilots, which I suspect is where you’re headed next.

Marc (12:58)
Yes, definitely, Jon. The integration with enterprise systems and real world application is certainly interesting. And, you know, this discussion leads directly to one of those biggest challenges utilities face with AI. Many have experimented, but far fewer have successfully scaled it across their operations. In in your experience, what typically breaks down between a successful pilot and a production deployment? Where do most pilots actually fail in getting to production?

Jon (13:33)
Marc, you read my mind. And the answer surprises people because it’s almost never the technology. The model usually works. What breaks is everything around it. The most common pattern we see is what I’d call pilot purgatory. A team builds something, it works, everybody’s impressed, and then it just sits there. It never scales. When you dig into why, it’s the same handful of things over and over.

First, there’s no clear owner. Pilot was run by an innovation group or by IT off to the side, and nobody in the actual business has their name on the outcome. So when it’s time to fund the scale up, there’s no one whose job depends on it. Second, there’s no real business case. “It works” is not a business case. If you can’t tie it to a number someone in operations or finance cares about, it dies the moment budgets get tight.

The third, the data problem you didn’t see coming. It ran beautifully on a clean sample. Then you point it at the full production data and it falls apart because real data is messy and incomplete in ways the pilot never had to deal with. The fourth, integration. The pilot lives on in its own island. Making it actually useful means connecting it to your OMS, your CIS, your GIS, the systems that people already work in. And that’s where months disappear.

The fifth, the people. If the crews or the analysts or the agents who are supposed to use this thing aren’t part of building it, they won’t trust it and they won’t use it. You can have a technically perfect tool that nobody touches. The thread running through all of this is the same. The pilot was treated as a science project instead of the first step toward something operational. If you build it disconnected from the business, the business won’t be there to catch it when it’s ready to scale.

Marc (15:30)
Those are lot of great observations, Jon. What what have you seen that separates utilities that have successfully scaled AI from those that are still stuck in pilot mode? You pointed out some of the issues, but what specifically are those successful utilities doing differently?

Jon (15:49)
Sure. They do a few things deliberately that the stuck ones don’t. They pick a problem that matters, not the most technically interesting problem. The one tied to a metric leadership already loses sleep over with a business owner attached to it from day one. And that owner isn’t IT. It’s whoever owns the operational result. They build from production from the start. This is subtle, but it’s huge.

The teams that scale don’t build a throwaway demo and then say, now let’s rebuild it for real. They build the prototype on a path that can actually grow into a production system, even if it starts small. You don’t over-engineer it and you don’t paint yourself into a corner either. They put it where the work happens. The tools that succeed show up inside the system the line workers and the analysts already use as part of their normal day. The ones that fail are a separate dashboard somebody has to remember to go look at.

And leadership actually makes choices. The utilities that scale don’t run 40 experiments and hope. They pick a small number of bets, fund them properly, protect them, and give them time. Then they manage the whole thing as a portfolio. Some will work and some won’t, and that’s fine, as long as the winners are allowed to grow. None of this is exotic.

It’s mostly discipline around ownership and focus and building the right way from the beginning.

Marc (17:20)
Sounds like scaling AI is certainly not just a technology challenge. It has real implications for how utilities are organized and and how people go about their work. What does adopting AI at scale mean for this, for the way utilities are organized and how work gets done?

Jon (17:41)
It’s really less about a big reorganization than people expect, and more about how a few things work day-to-day. The biggest shift is that the business and IT cannot operate in separate worlds with handoffs between them. The teams that get value from this are small and mixed. Someone who owns the business problem and someone who can build it, sitting together, talking to the actual users.

The old model where the business writes a requirement document, throws it over the wall to IT and waits a year, this is where these efforts go to die.

You do need some skills, but my strong advice is to build more than you buy. Bring in an outside help to get started on is smart. Outsourcing the whole capability is not, because then you never actually own it.

The goal should be that a couple of years from now, your own people can do this without us.

And you have to let two speeds coexist. The part of the organization keeping the lights on has to be careful and methodical. That’s correct. Don’t break it. But the part that’s experimenting needs room to move fast and to be wrong sometimes. Forcing the experiments to live under the same rules as grid operations is how you smother them before they ever prove anything.

Marc (19:03)
On that point, Jon, where does accountability for AI actually sit in a successful organization? Who owns it?

Jon (19:11)
It sits with the business, full stop. The person accountable for an AI tool should be the same person accountable for the outcome it’s supposed to improve. If it’s a predictive maintenance tool, accountability sits with whoever owns asset health and reliability, not with a data science team off in the corner. Your IT and data people are essential, but their job is to enable it, not to own whether it delivers value. The failure mode I see constantly is that AI belongs to an innovation lab or to IT, with no one in the operations on the hook for it. It produces interesting pilots and no lasting value because nobody whose real job depends on the result is driving it. The moment you put a business owner’s name on it, someone who’ll be asked in their performance review whether it moved the number, the whole thing gets serious in a hurry.

Marc (20:05)
Makes sense, Jon. How should utilities think about governing a portfolio of AI initiatives without stifling experimentation? You’ve you’ve hinted at a few areas, but as they get into this, how should they think about that?

Jon (20:22)
Marc, we think this is one of the most important points in this discussion. The key idea is that you don’t govern everything the same way. The instinct in a utility, for very good reasons, is to apply the same rigor to everything, because that’s how you run a grid safely. But if you put control room level governance on a back office experiment, you’ll kill it before it ever proves anything. And if you put experiment level looseness on something that touches the grid, well, that’s generally dangerous. So you match the oversight to the stage and the risk. Early on, when something’s just an experiment on safe data, keep it light. A one-page description of what you’re trying to learn, a quick weekly check-in, and the freedom to kill it fast if it’s not working. As it proves out and starts heading towards real systems and real data, the rigor then goes up. Architecture review, security, the works.

By the time it’s running in production, it lives under your full enterprise governance like anything else. And you manage it as a portfolio, not as individual projects. Going in, you should expect that most experiments won’t make it. That’s not failure, that’s the model working. You’re running a lot of small, cheap bets so that the few that pay off more than cover the ones that don’t. The discipline is being willing to kill the ones that aren’t working quickly, so you’re not funding zombies, and protecting the ones that are so they get a chance to grow.

Marc (21:53)
As utilities think about where to go from here, the key question is what they need to do next. If you were advising a utility executive today, what are the first one or two steps they should take to move from experimentation to real value with AI?

Jon (22:11)
Sure, two things. And they’re both smaller than people expect. First, pick one problem you could show real progress on in a few weeks, not 18 months, using data you already have. The most common mistake is waiting, waiting for a perfect data strategy, a governance framework, a big platform investment. You will wait forever, and your data is never going to be as clean as you’d like. Start with one valuable bounded problem and prove something real. The early win does more to build support and unlock budget than any strategy document ever will.

Second, put a business owner on it from day one, someone who owns the outcome, not just a technical lead. That single decision is the difference between a pilot that scales and one that sits on the shelf. What I’d steer them away from is opening with a big strategy and governance exercise. There is a place for that, but it comes after you’ve proven value, not before.

Prove it works on something that matters. Then build the structure around what you’ve learned. Start where you are with what you have and let the early wins fund the rest.

Marc (23:20)
That’s great advice, Jon. And thank you so much for joining me today. I appreciate you sharing your perspective on how utilities can move from pilot projects to real value with AI.