It sounds like the start of a bad joke: How many nuclear engineers does it take to change a lightbulb?

But Josh Bell, a member of Dominion Energy’s innovation team, says the answer is actually quite a few.

Any event that happens at a power plant, even something as minor as a burned-out lightbulb, gets logged in what’s known as a “condition report.” Dominion’s two Virginia nuclear plants in Louisa and Surry counties, its plant in Connecticut and one in South Carolina, collectively log roughly 50,000 condition reports per year. Surry alone sees around 14,000 reports annually.

Sorting through them all has become a time-consuming task, one the company hopes to streamline with artificial intelligence it will pilot at Surry later this month.

Every day, a team of five to 10 employees at the Surry plant is tasked with reading the reports that came in over the prior 24 hours. Reading them, ranking them by urgency and sending them to the appropriate department can take the team 20 to 30 hours per week, Bell said.

That’s why Dominion is in the process of rolling out software by Phoenix, Arizona-based AI developer Nuclearn that’s designed to help automate the process. 

Currently, the condition report team “is equally reviewing lightbulbs out as they are major equipment challenges, so the idea of this pilot is to provide a tool that can let them focus on the most essential responses and kind of expedite the pathway through the more mundane type situations,” Bell said.

Nuclearn, according to its website, was co-founded in 2021 by nuclear engineer Bradley Fox and computer scientist Jerrold Vincent, who’d developed an earlier version of the software during their tenure as employees of Arizona’s Palo Verde nuclear plant.

“At Nuclearn, our inspiration for developing this program came from firsthand experience working in the nuclear industry,” Fox and Vincent said in a joint statement to the Times. “We saw how much time and effort engineers spent on documentation and administrative tasks — time that could be better spent solving complex engineering challenges. The initial version of this solution was developed at the Palo Verde Generating Station in Arizona, where we recognized the opportunity to apply AI to streamline workflows, enhance accuracy, and ensure compliance with regulatory requirements.

Nuclearn says its software, dubbed CAP AI, is now deployed at 48 nuclear plants across the United States, Canada and the United Kingdom. Bell said the Surry plant will be the first in Dominion’s fleet to begin using it.

Dominion recently tested the AI by feeding it 150,000 condition reports the company’s four East Coast nuclear plants logged over the past five years. Bell said the AI correctly classified and distributed 75% of those reports.

If the AI doesn’t have a high degree of confidence in how to categorize a condition report or where to send it, the AI will flag that report for human review. Those reports account for the remaining 25%. The minimum level of confidence is something Dominion can control, Bell said.

“We can raise that as high as we want until we have determined that it’s going to do what we expect it to do and make the decisions we would expect it to make,” Bell said.

But there’s also a tradeoff, he said. The higher the confidence Dominion requires, the less it can automate.

Starting in late March, Dominion’s human review team will begin a six-month first phase of working in parallel with the software, where both they and the AI will review the same reports and then compare them to see if they reach the same conclusions. This will help determine the minimum confidence level the Surry plant requires of the AI while at the same time act as a failsafe to ensure the AI doesn’t misclassify anything critical, Bell said.

“We are rolling out our deployment plan over the course of 2025 and the expectation is by the end of 2025 we will have the data flow automated and will be in that … level-setting phase on the confidence level and automation rates that we are going to employ,” Bell said. “Our teams have gathered and are meeting and determining the right pace to pull this system into our normal work processes.”

Bell said the AI won’t put anyone out of a job. Even when the AI’s minimum confidence threshold is set and the redundant human reviews are no longer needed, the Surry plant will have the same number of employees it did before.

“Everyone who’s on those condition report routines has additional roles and responsibilities,” Bell said.

What this will do is allow those team members to “focus on those top 25% of truly significant condition reports that require problem solving” without having to “sift through the mundane” to get to them, Bell said.

The reason the AI is being piloted at Surry first is because the North Anna plant in Louisa County has a scheduled refueling outage this spring. Once Surry completes its six-month first phase, North Anna and Dominion’s Millstone nuclear plant in Waterford, Connecticut, and its V.C. Summer nuclear plant in Fairfield County, South Carolina, will begin their own rollouts.

Dominion spokesman Tim Eberly said Dominion has purchased six other software licenses from Nuclearn to roll out at its four plants over the next several years once CAP AI proves successful.

Editor’s note: This story was updated at 11:35 a.m. on March 7 to correct that Dominion has already purchased the six additional Nuclearn software licenses.