This is the fourth of a five-part series at AIxEnergy and mirrored at Michael Leifman's Substack
Part 1 introduced three distinctions (vendor capability versus utility adoption, reliability versus safety, task versus role) and a twelve-row sample of the US electricity workforce.
Part 2 developed the four-stage mechanism by which AI capability translates into workforce change in regulated industries and traced it through aviation, freight rail, and maritime.
Part 3 maps how AI is reshaping the U.S. electric-power workforce role by role, showing that task content is already shifting across the industry while headcount changes will unfold unevenly depending on whether work is routine, reliability-critical, safety-critical, or institutionally protected.
This article explains that AI and robotics are not only replacing existing power-sector tasks, but also creating new capabilities by letting the grid coordinate more complexity than human operators can manage and sending machines into hazardous places where human bodies should not go.
Years ago, in a training on power plant fundamentals at GE, I came to understand how much hazard is packed into the ordinary machinery of making electricity. A generating station is full of places hostile to the human body, from the energized switchgear where an arc flash can reach temperatures above 35,000 degrees Fahrenheit, nearly four times as hot as the surface of the sun in a fraction of a second, hot enough to vaporize the copper and aluminum it travels through, to the confined and superheated interiors of boilers and the high structures that were never meant for a person to stand on. The lesson that stayed with me was that across much of the power system the binding constraint on the work is not the skill of the worker but the vulnerability of the human body, which is why so much of it is hedged with safe working distances, protective equipment, and procedures written because the failure mode is measured in milliseconds.
The boiler is one of those places. In the fall of 2012, a Grove City College junior named Jake Loosararian toured the Scrubgrass Generating Plant in Kennerdell, Pennsylvania, and the plant manager told him that a worker had died the previous year while inspecting a boiler for defects. Loosararian and his classmates built a forty-pound robot with an ultrasonic scanner that could climb the boiler walls and hunt for the cracks and thinning that precede a failure. That senior project became Gecko Robotics, valued at $1.25 billion by June 2025, whose magnetic-wheeled robots now scale the walls of boilers, tanks, and other structures across the power sector, carrying sensors and, the company says, capturing far more of the surface than a rope crew ever could.
The first three parts of this series traced power-sector work that AI and robotics can now do in place of people. This part takes up the other half of the story, the half that does not register as a subtracted job, because it is work people could not do well, safely, or at all. Some of it was out of reach because no feasible number of people could do it fast enough. The rest was out of reach because a human body could not be where the work had to happen, or could not survive being there.
Two kinds of capability are arriving at once, and it helps to keep them separate. The first is cognitive: AI can now hold more variables in view, and act on them faster, than any control room of people could manage. The second is physical: robots built in shapes no human resembles can now enter spaces that injure or kill an unprotected person. When I wrote about AI and labor in 2018, I expected both cognitive and physical work to be exposed, brain and brawn together. What I did not expect was the sequence. Cognitive automation arrived first and fast, through language models, while physical automation lagged, slowed by the stubborn difficulty of getting machines to manipulate the real world. The power sector is one revealing exception to the general rule. Here, physical capability is arriving early in a specific set of niches, not because the robotics got easier, but because the human was never a safe option to begin with.
The first engine: cognition at a scale people cannot reach
Hannah Kaplan, writing on this platform, describes the shift underway in grid operations as a move from reactive reliability to dynamic intelligence. For most of its history the grid was operated reactively. Something failed, an alarm sounded, and a crew responded. The system Kaplan describes is different in kind. It coordinates millions of devices at the edge of the grid, rooftop solar, home batteries, electric vehicles, and smart thermostats, in something close to real time, shaping supply and demand continuously rather than waiting for the next emergency.
That coordination is not a faster version of what a human dispatcher did. It is a task no dispatcher could do at all. Consider what happened in California on the evening of July 29, 2025. More than 100,000 home batteries discharged together for two hours, delivering an average of about 535 megawatts to the grid, the output of a midsize power plant, assembled out of appliances sitting in tens of thousands of garages. The event was a test of California’s Demand Side Grid Support program, coordinated by the state’s Energy Commission and its grid operator, with the batteries pooled by aggregators including Sunrun and Tesla. The act of conducting that many devices, second by second, so that they behave like a single plant, is software doing something no staff of people could.
The same pattern appears in the queue of projects waiting to connect to the grid, which I covered in Part 3. Reviewing an interconnection request means running power-flow studies to see what a new generator or large load would do to the surrounding network, work that has historically taken human engineers months per case and produced a backlog measured in years. AI systems can run those studies at a speed that compresses the arithmetic from months toward minutes, which matters because the queue, not the supply of equipment, is now one of the binding limits on how fast new power can come online. Kaplan points to the same kind of capability in wildfire country, where utilities use predictive models to score ignition risk across an entire service territory and decide where to cut power before a fault becomes a fire. Scoring a whole territory continuously, weighing weather, vegetation, and equipment condition together, is once again a problem of more variables than a person can hold at once.
The second engine: places the body cannot go
The robots showing up in the power sector look nothing like people, and that is the point. I have made a version of this argument before: we long ago stopped designing our fields for the ox and our roads for the horse, and there is little reason to design the next era of work around the human form either. A human shape is the right design only when the work was built around human bodies, and the work in question here was not. It happens inside narrow gaps, atop tall structures, underwater, and in radiation fields, so the machines that do it take whatever form the task demands.
The substation is another such place. In 2024 the utility Avangrid began a pilot at two of its United Illuminating substations in Connecticut, sending a four-legged robot named Spot, made by Boston Dynamics and nicknamed Sparky, on inspection rounds to read gauges and capture thermal images of transformers and breakers, with software that flags anomalies. The value, as the utility described it, is partly that the robot can inspect far more often than a crew making periodic visits, building a record of how equipment behaves across seasons and loads. It is also that every inspection the robot performs is one a person does not have to perform inside an energized yard.
The same constraint operates inside the machines themselves. A large generator is normally inspected by pulling the multi-ton rotor out of the stator, a process so costly and slow that the examination was historically left out of outage schedules when time was short. Today a robot small enough to pass through the roughly one-inch air gap between rotor and stator crawls the length of the machine carrying cameras and test instruments, performing the inspection with the rotor left in place, and GE Vernova offers a related tool that scans a generator’s retaining rings for stress-corrosion cracks without removing them. This is inspection, not repair, and the distinction matters. Getting a robot to see inside a place a person cannot reach turned out to be far easier than getting one to fix what it finds there, which is the same lesson the wider robotics industry has been learning, that locomotion was the solved problem and manipulation the hard one. In-place robotic repair inside turbines and generators exists today mostly at the prototype stage.
The same shift is visible on wind turbines, where the leading edges of blades erode under years of rain and grit a hundred meters in the air. That repair has meant rope-access technicians, who climb or hang from the blade and can work only in calm weather. The company Aerones now runs a ground-controlled robotic system, suspended by cables, that cleans, sands, fills, and recoats eroding blades, and can work in higher winds and a wider temperature range than a person on a rope can. By 2025 the company projected servicing more than 10,000 turbines, with deployments at US wind plants including two in Oklahoma. Notably, Aerones is explicit that it is not removing the technician: the robot does the leading-edge work while technicians operate it from the ground and handle the repairs the machine cannot, a division of labor worth holding onto for what follows.
The hardest limit of all is radiation, because it is a biological ceiling that no protective equipment lifts. Inside a nuclear plant, the surfaces of the reactor refueling cavity must be cleaned and decontaminated during outages, work historically done by people after the water was drained, slowly, and at the cost of radiation dose to those workers. The company Diakont sends a hybrid swimming-and-crawling robot into the flooded cavity instead, operated by a small team from the edge of the refueling floor. At the Perry plant in Ohio, the robotic decontamination was thorough enough that no manual cleaning was needed afterward, and it kept people out of the cavity entirely. Robotic inspection of reactor internals and of the thousands of narrow tubes inside a steam generator follows the same logic, sending a machine where a dose limit forbids a person.
Described separately, the two engines read like different stories, but they are beginning to act as one. Imagine an inspection robot inside a grid-scale battery enrolled in a program like California’s, picking up the early thermal signature of a cell starting to fail. The same coordination layer that conducts a hundred thousand batteries could take that one unit offline and rebalance the rest of the fleet within seconds, faster than any operator watching a screen could react. The physical engine reaches what no person could reach, and the cognitive engine acts on what it finds at a speed no person could match. That combination, more than either engine alone, is where the capability becomes something new.
The edge of what does not exist yet
Everything to this point either is in service or is close to it. The more interesting boundary is the one just beyond, where the most consequential capabilities are still being invented. Putting power lines underground would end much of the storm and wildfire damage that overhead wires suffer, but at more than six million dollars a mile it is so expensive that utilities rarely do it. In January 2024 the Department of Energy’s advanced research arm launched a thirty-four-million-dollar program, GOPHURRS, to cut that cost by shifting undergrounding from digging to drilling. One of its dozen projects, GE Vernova’s SPEEDWORM, is a robot that mimics the way an earthworm moves to dig and lay cable in a single pass, with the goal of installing a thousand feet of line in two hours from the back of a standard pickup truck. It does not exist as a working machine yet. If it comes to work, it would give the power system a capability it has never had, the ability to put lines underground at a cost that makes undergrounding ordinary rather than exceptional.
The harder version of the same problem appears during a storm, when drones can map the damage while it is still too dangerous to send people, but the repair itself cannot begin. After Hurricane Helene, Duke Energy used drones to help restring a transmission line, and robots can now inspect energized lines on their own, as EPRI’s “Ti” does gliding along the wire above an energized transmission line in Ohio, with a few systems able to make limited repairs on live conductors while a crew stands by. What does not yet exist is the capability that would matter most in a storm, a machine that could find and mend a fault on its own while the weather still made the site too dangerous for people. Until then the repair waits for the storm to pass and for crews to reach the poles, which is why a bad storm is still measured in days without power, and why the storm, not the technology, sets the clock.
What gets created, and what does not
Both engines create capability that did not exist, and it is tempting to read that as unambiguous good news for employment, since new capability has historically meant new work. Some of that holds here. The orchestration layer needs people who can build and run it, the robots need operators, the data they produce needs interpreting, and the wind example shows the pattern at its most benign, with the dangerous task removed and the worker repositioned rather than dismissed.
New capability is not the same as new employment, though, and the new work rarely appears where the old work was. The cognitive engine is real but bounded. Kaplan’s own assessment is that AI, distributed resources, and virtual power plants do not replace wires or generation, but buy time and optionality, and that the industry needs precision at scale more than it needs more pilots. Its progress depends on data quality, on regulatory alignment, and on institutional trust, none of which arrives on a software schedule. The physical engine is bounded by the gap between seeing and fixing: the inspection is here, the in-place repair mostly is not. And where a robot removes a person from a hazard, it also removes the hours that person was paid for, even when no one would wish those particular hours back on a human being.
Part 3 of this series drew two lines worth recalling. One separated the moment a capability becomes possible from the much later moment institutions actually absorb it, since licensing, liability, labor agreements, and regulatory minimums all move slowly. The other separated the tasks a technology takes over from the jobs it eventually ends, since a role can shed its most dangerous tasks long before anyone counts a position as gone. The capabilities in this part bear on the first line in a particular way. Because so many of them do work no one was doing before, reaching where a person never could or coordinating more than a person ever tracked, there is no incumbent role to protect, no contract or license or staffing minimum standing in the way, and so the institutional resistance that slows displacement is weakest, almost by definition, exactly where the capability is newest. The harder questions, what institutions do as these tools mature and how the loss of tasks becomes the loss of roles, belong to the final part, which turns from what AI and robotics can do to what utilities, regulators, workers, and investors should do about it.
The capabilities described here are neither science fiction nor uniformly mature. A robot on an inspection route and a fleet of home batteries acting as a power plant are in service now, while in-place repair inside a turbine, a line laid by a burrowing robot, and the full orchestration of a dynamic grid are closer to the frontier than to the norm. That mixture is exactly what makes the trajectory easy to underestimate. Each piece looks like a pilot, a demonstration, a research grant. Put together, they describe a power system in which the limits once set by what a human body could survive, and by how many variables a human mind could track, are quietly being lifted. The decade ahead looks slow until it does not.
