The top three questions utility leaders are asking about AI

Raiford Smith
Global Market Lead, Power & Energy, Google Cloud
AI doesn’t have to disrupt utilities — but it does take proactive leadership, careful data control, and a willingness to experiment to be empowered by AI.
AI is evolving at lightning speed. With the constant deluge of new models, enterprise solutions, and consumer products, the pace of change can feel overwhelming. However, the excitement is real and so are the opportunities to transform operations.
No utility leader wants to miss out. They want to understand where and how best to apply AI. Whenever I talk with them, I often hear the same three questions.
Question 1: How will AI change our jobs?
When I worked at a municipal utility company, planning reported to me. My team was concerned that the analytics we were developing would affect their jobs. I asked them, “If we can run more analyses, run them faster, and get more insight out of each one, does that shrink our role or expand it?” They realized it would create new opportunities to add value. And that’s exactly what happened.
Planners look into the future to make decisions about which resources to apply where and when. Using models that take hours to days to run, they only have time to analyze a handful of scenarios before they have to submit their plan.
AI dramatically accelerates analysis; our DeepMind team built a security-constrained optimal power flow model, CANOS, that runs 28,000 times faster than the traditional one. This creates room to test far more scenarios. With a probabilistic view, planners can concentrate on the ones that carry real risk instead of spreading equal attention across the handful they had time to examine.
A similar shift is happening in the field.
Managing weather, late-arriving materials, and an ambitious schedule has always landed on the construction site foreman. AI can look ahead at the remaining work, which trades can do which tasks, which equipment has to be in place first, and then run simulations to pinpoint the least cost, least time path to finish. The foreman makes the final call. NextEra is doing something similar with maintenance, using a predictive model to anticipate equipment issues and optimize crew deployment against supply chain and weather constraints.
Many will still be doing the same job they do today, just differently.
Question 2: How can I trust what AI is telling me?
One of the topics I most often hear about in executive meetings is AI hallucinations, when a model produces something that sounds right but isn't grounded in anything real. That is, you can’t trace the conclusion back to actual data. Some call that a defect, but I'd frame it differently: The same behavior that produces a wrong answer also produces a creative one.
Trust often comes down to control, and you control more than you might think.
For example, it’s up to you how much latitude to give a model. Safety procedures and regulatory compliance require strict if-this-then-that heuristics, so for those use cases you turn down the creativity dial and tell the model to follow the rules to produce a more deterministic output.
But when you're looking for a design nobody has tried, or to understand nuances in human speech during a customer call, you give the model more latitude to reach past the obvious answer. Sometimes it comes back with the wrong, but sometimes it gives you a response your team would never have considered.
When you're looking for a design nobody has tried, or to understand nuances in human speech during a customer call, you give the model more latitude to reach past the obvious answer.
You also control which model you use. In fact, you want the flexibility to run several, which Google built the Gemini platform to support. You can experiment with different models to determine which ones give you the best answers, with the fewest hallucinations, for your specific tasks.
No matter which model you choose, its accuracy and creativity depend heavily on well-governed data; the better your data, the less room your model has to invent things from whatever it can find. And even if you fine tune the model and carefully govern your data, you must still validate what comes out before incorporating the results into your operating and planning decisions.
Question 3: Can I use AI and stay compliant?
NERC CIP standards don’t bend because a new technology shows up, nor should they. But people may assume that if your systems can’t connect to the cloud, you can’t use AI. That’s not what the standards say.
CIP applies to the assets supporting the bulk electric system, not to everything a utility runs. Your planning environment, load forecasting, finance group, and work management systems generally sit outside that perimeter.
For the systems that do fall inside, or that have to operate fully disconnected for other reasons, the AI can come to you. Google, for example, offers air-gapped solutions that run on your premises, under your control, with our AI capabilities built in. We’ve also published guidance on using our services in a way that supports CIP compliance.
This comes back to the point I made about control. You control more than you might think, and that includes where the model physically runs. An air-gapped deployment keeps your models inside your own perimeter, on hardware you manage, working only with the data you’ve put in front of it. Nothing crosses the boundary.
The case for optimism
AI isn’t something happening to our industry, like potentially inclement weather we have to prepare for. What AI does for — and does to — our industry depends on the decisions we make: How we redesign our work to capitalize on AI's strengths, how well we match each model to the job we're giving it, and where we put it to work across planning, building, and operations to drive efficiency and reduce waste.
Whether you're a large investor-owned utility or a municipal one serving part of one state, the tools are now interactive enough that you don't need a specialized team to implement AI. My advice is the same for any company: Start trying AI, start using it. There’s no shortage of use cases, and you'll have plenty of employees saying, "I've got an idea."


