A tool named in honor of EPRI founder Chauncey Starr is just the start.
When EPRI was created in 1972, Chauncey Starr was asked to become the organization’s first president. Starr already had an impressive career, having worked on the Manhattan Project and helped pioneer the academic study of risk analysis.
Starr had one non-negotiable condition to accept the leadership role at EPRI: Research had to be independent and of service to society at large. Starr believed that science and technology existed to serve the public and improve the quality of life for all people. Starr was also convinced that communication and accessibility were vital for research to have a beneficial impact on both the electric power industry and the public. In fact, in the inaugural edition of the EPRI Journal, published 50 years ago, Starr noted that the publication’s primary objective was “to report on the results of EPRI research.”
EPRI Research is More Needed than Ever
Today, the need for easy access to EPRI’s research is more urgent than ever. Utilities, regulators, policymakers, and the public are all navigating a power system being transformed by rapidly increasing electricity demand, an influx of distributed energy resources (DERs) and renewable generation, digitalization, and climate change. A challenge facing member utilities and other stakeholders interested in leveraging EPRI research, however, is the sheer volume, breadth, and depth of five-plus decades of work.
“It is very easy to get lost searching for research reports. EPRI has this very spiky corpus of research, and we deal very deep in questions consultants and the market don’t answer,” said Justin Thibault, EPRI.ai product manager. For example, Thibault recalls an EPRI research project launched after the manufacturer of a grease used in many nuclear safety applications ceased production. EPRI responded by studying various lubricants to determine whether they could serve as adequate replacements. “How do you keyword search that?” Thibault said.
Keyword searches on the EPRI website are no longer the only option for people seeking to benefit from the organization’s research legacy, much of which is available for free. Generative artificial intelligence (AI) has made it possible to transform how both EPRI members and the public engage with and benefit from EPRI research. A new EPRI product, appropriately named Chauncey, allows users to ask questions in their own words (including misspellings and inelegant phrasing) and receive plain language summaries grounded in EPRI’s research that answer the question.
Chauncey, powered by the EPRI.ai platform, is part of a broader EPRI effort to leverage artificial intelligence in collaboration with the electric power industry. EPRI.ai sits alongside related initiatives including the Open Power AI Consortium, which brings together utilities, technology companies and researchers to develop AI tools, models and datasets for the power sector, and SAFERai.power, an EPRI-led effort focused on assessing and managing the risks of deploying AI in utility settings.
A False Start Leads to an Important Pivot
EPRI’s journey to launching Chauncey began with a failed experiment. In late 2024, EPRI experimented with feeding publicly available research reports into a chatbot. The goal was to determine whether the approach would yield a user-friendly interface that would allow EPRI employees to ask a question and receive an answer, similar to how ChatGPT and Claude work.
It didn’t work largely because of the depth and specificity of EPRI’s research. For example, someone seeking insights into asset management for the distribution system could instead receive answers relevant to nuclear power assets because of how they asked the question. Put simply, the phrasing of the question determined whether the answer was of any use. “It was really easy to get lost because EPRI goes very deep in areas like asset management,” Thibault said. “We learned very early that we were not going to put that out into the market and say, OK, this is our solution.”
Instead, the AI team pursued a hybrid approach that combined aspects of how people have traditionally found EPRI research in the form of a Generative AI-enabled assistant. To put it more precisely, what eventually became Chauncey blends the keyword search users have long used to find reports on EPRI.com with semantic queries that seek to grasp the meaning of a question or request. In other words, a keyword search informed by the result of a semantic search. The semantic search delivers results based on its understanding of the user’s meaning, even when there are typos and non-technical language, and also generates the summary within the context of EPRI’s research.
A Different Development Approach
The philosophy behind the development of Chauncey differed from other projects Thibault has worked on at EPRI. In the past, Thibault says technology development projects started with the technology’s features and capabilities, only later connecting them to a valuable use case. In the case of Chauncey, however, the approach was to start with the problem they wanted the technology to solve: the difficulty members and others would encounter when searching for research. “We really wanted to fall in love with the problem first,” Thibault said. The result was a product that leverages generative AI’s semantic search capabilities to answer user questions in plain language, while a traditional keyword search runs simultaneously.
Here’s what that looks like for users of Chauncey: A query will produce a summarized answer based on a distillation of EPRI’s body of research and the reports that would come up from a traditional keyword search. In addition, Chauncey provides users with additional information, including relevant EPRI contacts and events relevant to the search topic. The results are also filtered based on the access their membership allows, meaning that two members could receive different answers to the same question.
It wasn’t just the product development process behind the creation of Chauncey that was different; what was also novel was the speed at which Chauncey was built. Thibault and his team first showed mockups, known as wireframes, of Chauncey’s user interface to member focus groups in the fall of 2025—with a broader launch following in the spring of 2026. One important pre-launch step was to ensure that Chauncey would provide fundamentally different and more valuable answers than those from other widely used AI tools.
To do that, Chauncey underwent an independent benchmarking process that evaluated its answers against those of other tools. “We didn’t want to put something out there where you are going to get a better answer from the free version of ChatGPT,” said Thibault, who notes that Chauncey doesn’t replace other AI tools because it doesn’t have general knowledge, only power-sector, EPRI research-based knowledge. Since its launch, EPRI technical staff have provided feedback on Chauncey’s answers, helping ensure their accuracy and value.
A Preference for Chauncey, and What’s Next
Based on one measure, EPRI members prefer the use of Chauncey over traditional keyword searches. To reach that conclusion, Thibault examined a seven-day rolling average of Chauncey usage. Anyone who logs in to Chauncey will be given the choice to use either the new tool or the “Quick Search” function, which is keyword search. What Thibault discovered is that members engaged with Chauncey rather than relying on Quick Search only at a rate of five-to-one. Because Chauncey is the default solution and Quick Search requires users to navigate to it, some preference would be expected. But Thibault says the design feature would only explain a two-to-one ratio, suggesting that members are actively choosing to use Chauncey.
Chauncey is by no means the only tool being developed on the EPRI.ai platform. Among the other tools in development are Research Insight Assistants (RIAs), AI assistants that specialize in a specific technical area, such as transmission assets. Research Insight Assistants will be able to leverage EPRI’s published research and internal resources, such as meeting transcripts and presentations. Also on the development roadmap are research agents. Unlike assistants like Research Insight Assistants that answer specific questions, agents can be given a more general task and pursue it independently.
In the future, members could ask Chauncey a question, transition to working with a Research Insight Assistant as their needs become more technically focused, and then collaborate with a research agent. While there are nuances among the specific tools, the vision driving them is the same. “The product vision is really to make EPRI’s collective knowledge actionable in the flow of work,” Thibault said.
EPRI Technical Expert:
Justin Thibault
For more information, contact techexpert@eprijournal.com.
