Utilities turn to AI to manage rising power demand — from AI

As electricity demand from artificial intelligence surges, the power sector is looking to the same technology for new ways to plan and maintain the grid more efficiently.

The utility industry is a long way from allowing large language models to take action affecting physical infrastructure, executives say. For now, most AI deployments in the sector are niche tools designed to support human decision-making on topics like load forecasting and vegetation management. 

Barriers to broader utilization include uncertainty over what to do with AI outputs that are uneven or that can’t be audited in a highly regulated sector with strict safety and reliability standards.

“Many demonstrations perform well in controlled conditions, but evidence of consistent performance at scale remains limited,” Rick Lesley, executive vice president and energy and utilities sector lead at QualityAI, told Utility Dive in an email. “Most deployments assist people with forecasting, document retrieval, anomaly detection or workflow automation.” 

Even with these limited AI uses, some have raised concerns that utilities are going too fast, too soon. Earlier this month, the Federal Energy Regulatory Commission released a cybersecurity audit of critical infrastructure entities that found “increasing use of super intelligence” tools, warning that their “evolving capabilities may introduce access control risks.”

One review of federal regulatory filings by Guardrail Technologies, which provides AI security and governance software, found that roughly 70% of energy and utility companies identified AI as a cybersecurity risk — but just 23% of utilities and 9% of energy companies documented a process for managing that risk. 

“Most utilities could not reconstruct what an AI system read, what it did and why,” said Guardrail CEO T.J. Marlin. “Reconstruction requires a record, and a record only exists where a governed process captures one. The utility and energy sector's disclosures say that process mostly isn't there yet.”

It’s unclear ‘how much AI is already running’

The lack of documented oversight also makes it difficult to assess how widely utilities are using AI and for what purposes. 

Some states are moving to tighten scrutiny. New York regulators recently directed utilities to file inventories of their AI use. On Oct. 13, Connecticut Attorney General William Tong and Consumer Counsel Claire Coleman petitioned their state regulators to open a similar review.

“What's still unclear is how much AI is already running and what data it's running on,” said Stephen Rose, CEO of Render Networks, which provides software for managing electric infrastructure.

Rose said the quality of the input data is especially important. In his experience, utility records do not always reflect what is physically in the field, he said.

“If you automate against bad data, you just get the wrong answer faster, and nobody can tell you how often it's wrong,” Rose said.

Among utilities, AI adoption has been strongest in tools designed for specific functions, said Pradeep Tagare, interim president of National Grid Partners, National Grid’s venture capital arm, which invests in AI energy startups.

“Where most utilities are finding success today is in what I will call the ‘vertical AI’ applications within utilities,” Tagare told Utility Dive in an interview. “So, these are not the broad horizontal AI platforms that get deployed companywide, but rather specific, function-related applications.” 

Many utility AI projects remain in development or pilot stages, Tagare said, as utilities are still working out what cybersecurity risks they pose as information technology and operational technology networks become more interconnected.

“[For] anything that touches the physical assets directly, we haven't seen widespread deployment of agentic architecture yet,” he said.

Nearly 80% of utility leaders NGP surveyed this year in the United States, the United Kingdom and Europe said their organizations had fully deployed at least one AI application for large-load customer planning. 

Notably, 87% of respondents said outdated regulatory and rate case frameworks are hindering returns on AI and innovation investment, and 83% said the cost of building infrastructure to serve AI data center demand is being passed on to residential customers through higher electricity bills.

Tagare pointed to several companies NGP has invested in as examples of AI deployment, including GridCARE, which uses hourly demand modeling to identify available grid capacity. Portland General Electric said its work with GridCARE identified more than 80 MW of capacity for data center connections years earlier than initially expected. 

Another NGP portfolio company, Emerald AI, says its platform can reduce data center power demand by up to 40% in response to grid curtailment requests while keeping critical workloads running.

Bridging domain expertise and AI expertise

Virtual power plants are one area where machine learning is already being used to coordinate distributed energy resources, Tagare said. Although VPPs using machine-learning systems have been in use for about five years, they do not represent the same level of automation as systems that independently control core grid operations.

The concept is still in its early stages, but Tagare said a large language model trained on utility operations and equipped with industry-specific guardrails could eventually be combined with a representation of the physical world for more active grid management. 

“I expect to see this gain more momentum in the next couple of years,” he said. “That’s where the innovation is going to come from.”

But getting there will require substantial investment in talent. According to NGP’s survey, “the hardest roles for companies to fill right now are domain experts with AI fluency, data engineers and architects, and AI/machine learning engineers.”

Tagare said he expects the shortage to become even more apparent as AI expands into mission-critical operations, which may depend on empirical knowledge that has never been written down. 

“You still need a lot of domain expertise, you need a lot of AI expertise, and you need to figure out how to make the two work together,” he said. “That, I don't believe, is going to get automated anytime soon.”