Powering Strong Communities
Grid Modernization

The Risks and Possibilities of AI for Utilities

This story is based in part on a presentation at our 2023 Customer Connections Conference that featured Jared Combs, business intelligence analyst from CDE Lightband in Clarksville, Tennessee; Juan Corredor, chief technology officer at Open International; and Brian Lindamood, vice president of content & and marketing strategy for Questline. 

Since OpenAI released its ChatGPT platform over a year ago, many articles have been written about the potential for artificial intelligence to transform how each of us approaches our work and various aspects of our daily lives. As people have gotten more used to discerning whether an image or text has been produced using generative AI, the data pouring into these systems has also helped train AI platforms to produce material that seems less uncanny valley and more “real.” (Note: AI didn’t help write this story).

From enhanced operational awareness and forecasting capabilities to assisting with complex modeling, predictive maintenance, or system planning, AI offers the potential to help the people who work in utilities to be more efficient and more effective. But before AI transforms utility jobs, people working in utilities need to gain some understanding of what trying out such applications could mean for the quality of their work and the safety of the system.

Knowing the Terms

Although generative AI systems have been the focus of attention over the past year, there are several distinct types of AI that can serve different purposes.

Heuristics are about solving problems or wading through complex decisions. They are a foundational function of AI and have been in use for a while, underlying the algorithms behind website searches. For utilities, heuristics could be established to identify potential problems with equipment or parts of the system.

Machine learning takes a step beyond mining through data and attempts to learn from the data and to recognize patterns. Machine learning systems are often about trying to help predict future actions, such as for load forecasting.

Generative AI is about creating and adapting work, such as writing and images. Some of these platforms use large language models. For utilities, generative AI could be used to come up with potential responses on a customer service chat or to develop images and messaging for the utility.

Potential Concerns with AI

The people working with AI in any field have several areas of concern, often relating to safety, ethics, and accuracy.

The Department of Energy is working on a report about the potential benefits and risks of using AI in critical energy infrastructure. Key findings from the draft DOE report include that there is a need for the workforce who interacts with critical energy infrastructure to better understand AI applications and their risks and to do further research into the implications of using AI within the energy sector.

As an emerging field, the use of AI is largely unregulated. As regulations are imposed on its use, utilities that have implemented AI will need to be aware of potential additional compliance measures.

The rapid proliferation of AI-based applications — and the data that feeds them — poses another challenge to utilities: increasing demand. In its 2024 Electricity report, the International Energy Agency projects that in the U.S., demand from data centers will increase to 260 terawatt-hours by 2026, up 30% from the energy use from data centers in 2022. The IEA’s report predicts that data centers will account for about a third of the overall 1.5% estimated annual growth in electricity demand for the next two years. This growth is concentrated in areas seen as most favorable to data centers, where rates are low and supply reserves healthy, but the rapid increase in demand threatens both factors.

Part of the accuracy factor is for utilities to have “clean” data sets to work from. Higher confidence in the data translates to higher confidence in the analysis. Still, there are continual opportunities to maintain data and check for inconsistencies, either as part of training and maintaining the AI or in ensuring it isn’t developing misinformation.

Misinformation and bias can crop up when using generative AI and should be checked. For example, ask for sources for any claims made in a piece of writing and then review them to see if the sources are credible or if the claim was taken out of context. Another known problem is that AI can “hallucinate,” which is the term for when it makes up information. Using open-source AI code poses data security risks, from the potential to download malicious code that could harm a utility’s system to the potential for sensitive information being accessible if data gets shared into an open system.


The biggest plus that utilities bring to the table in being ready to use AI is access to data. The AI then would be a tool in running various analyses on the data. Benefits including scaling up efficiency, understanding customer needs, and saving time. The idea is that AI allows utilities to analyze customer preferences not just on a few data points, but on all data points in a timely manner.

Utilities could automate the sending of certain communications, such program promotions or predicted peak events, based on real-time customer needs. It could advance customer service by giving customer service representatives additional information and make recommendations for the most appropriate programs to suggest, based on eligibility or other factors. The recommendation is ultimately a human one, but analysis makes it possible to do more quickly.

Jared Combs, an analyst for CDE Lightband, the department of electricity in Clarksville, Tennessee, shared how the public power utility has used AI in several applications in recent years, including for load forecasting, customer segmentation, and analyzing organizational documents.

For the latter, the utility used an open-source large language model but set it up to only run within the utility, to search through utility documents and the full text of the Inflation Reduction Act to find the best matches for opportunities for the utility within the programs created by the law.

CDE Lightband was also able to create more meaningful segments of its customers by doing detailed cluster analyses based on advanced metering infrastructure data. This type of clustering allowed the utility to model different scenarios of how its system would be affected if customers with different load profiles were to adopt certain behaviors or technologies, such as electric vehicles.

AI could also allow utilities to create more tailored messaging and refine customer segmentation based on specific factors. For example, Brian Lindamood from Questline mentioned working with the Long Island Power Authority to first analyze how customers could be affected by proposed rate changes, and then create targeted communications based on those changes. Relying on detailed usage profiles, the authority was able to share with customers ways they could save energy and how much savings could be possible under different behavior and usage changes. In Iowa, Lindamood mentioned that Questline also worked with Duquesne Light to have AI analyze which segments of customers were more likely to adopt practices such as paperless billing, and then the team identified potential hurdles that might be keeping people from doing so. AI-created newsletters can also pull in content that can be tailored to readers’ interests, resulting in overall higher engagement. Lindamood mentioned working with AEP Ohio on such a newsletter for commercial customers that led to significantly increased engagement metrics.

Others pointed out the possibility of AI being able to connect with an outage management system to help pinpoint fault areas or other system issues more quickly, or to aid in predictive maintenance. It could also help predict and automate sending expected restoration times to customers in the event of an outage, based on history and other factors.

Your Passenger, Not the Driver

AI needs hardware to run on, coding to tell it what to do, training to make sure it is performing correctly or improving its performance, and maintenance to ensure it isn’t creating or running into problems. While machine learning and generative systems require ongoing training and maintenance, heuristics require legwork to ensure the coding is correct.

New technologies and platforms are making AI cheaper and more accessible, but the cost to develop or train your own system can involve upfront expenses to learn how to code such software, paying a third-party to do so, and the operational costs to run the machine(s).

It is important to take the time to ensure utility leaders and other stakeholders are aligned on what business goal the AI would help to achieve.

Even if engaging a third party to develop AI for your utility to use, it is important for utility professionals to be become literate in it — understanding the basics of what it does and doesn’t entail and the limits of what it can and can’t do. With the ongoing difficulty in being able to recruit or retain certain technical staff, utilities might feel that the level of expertise needed to be able to use AI safely and effectively is out of reach, but AI literacy doesn’t mean getting a degree in computer science. Having a basic level of AI literacy will make it easier for utilities to ask the right questions of any developer and to appropriately define the scope of any program.

Lindamood noted how people might already have some familiarity with AI through examples we’ve become accustomed to, pointing to features such as recommendations from Netflix and voice assistants on mobile phones.

“Because it is new, people are learning together,” added Combs. He mentioned that there are a variety of high-level courses, including from prominent universities, available online through sites like Coursera that don’t cost too much and that help people understand some of the basics of AI and machine learning. He said such courses can help utility professionals learn which options are decent tools and what problems or different situations the tools are suited for.

While critical infrastructure owners and operators could establish applications for AI, the key piece of advice in moving forward is to ensure that such tools are set up to help humans make better decisions, not for the AI to be empowered to make those decisions directly.

While AI might be able to run analyses or offer suggestions, it isn’t sophisticated enough to develop an overarching strategy, especially for areas like communication.

“The AI can make a human more effective at their job, but without a human is much less useful,” noted Lindamood. He said that being able to use generative AI will be a skill that can help, but not replace, creators.

NEW Topics