Making AI work for utilities means treating technology as a partner, not a replacement

AI utilities technology: Grid operators gain 40% efficiency by treating AI as co-worker, not replacement

Sarah Mills | GlobalBeat

Southern California Edison cuts outage response time by 40% after training dispatchers to second-guess its machine learning models instead of obeying them.

The $2 billion pilot, reviewed Wednesday at the CIGRE power conference in Paris, marks the first large-scale proof that human override buttons turn AI from flashy lab demo into reliable grid gear.

Utilities globally lost $77 billion to weather-related outages last year while facing a 23% staffing gap in control rooms, forcing operators to automate faster than trust allows.

“The machine spots patterns my 22-year veteran eyes missed,” SCE senior dispatcher Carla Mendoza told reporters. “I still veto it about twice per shift. That marriage keeps the lights on.”

Mendoza’s team began using the AI co-pilot in March 2025. Since then, average outage minutes per customer dropped from 118 to 71 during heatwaves, according to data filed with California regulators. The software digests 15,000 sensor feeds every 10 seconds, recommending switch placements and voltage tweaks. Humans keep the final click.

European grid giants watched closely. RTE France announced it will copy the playbook this winter, embedding data scientists inside live control rooms rather than relegating algorithms to back-office analytics. “We learned the hard way that black-out risk rises when operators treat AI as gospel,” RTE innovation director Lucille Cheng said, referencing an August 2025 incident when automation wrongly shed 1,200 megawatts of solar across Provence.

Vendor earnings show the same story. Siemens Energy reported utility AI orders tripled to $1.4 billion in the past quarter, but 87% of contracts now demand explainable dashboards instead of black-box optimization. “Buyers want glass boxes, not black boxes,” product chief Jonas Rabe said. “If the model cannot show why it wants to open breaker B-42, the answer is no.”

Union pushback shaped the same trajectory. The International Brotherhood of Electrical Workers lobbied for contract language that bars unsupervised close-out of maintenance tickets. “We aren’t anti-tech,” union spokesperson Dana Hill said. “We are pro-accountability.” After six months of joint training at SCE, grievances tied to automation dropped 60%.

Money follows safety. Moody’s upgraded the outlook for three California utilities partly on evidence that AI-coordinated wildfire switching reduced liability exposure by an estimated $340 million this fire season. Investors once feared that smarter software merely accelerated failure; now they price in fewer sparks.

Grid regulators are rewriting rules to match the hybrid culture. The Federal Energy Regulatory Commission proposed in April that any AI tool controlling more than 100 megawatts must log every recommendation plus human response, creating a subpoena-ready audit trail. “We want the algorithmic chain of custody,” FERC chair Mark Christie said.

Smaller co-ops lack deep benches of data scientists. Arkansas-based Ozarks Electric found a workaround by pooling load data with 14 neighboring rural utilities in a cloud sandbox. Together they trained a shared forecasting model that cut peak-time purchases on the spot market by 8%, worth $11 million annually across the group. Each coop keeps local veto rights. “None of us could afford this alone,” Ozarks CEO Mitchell Johnson said.

Cybersecurity fears remain. A red-team exercise at Eesti Energia in Estonia showed that AI-assisted phishing emails breached operator workstations 42% faster than human-only attempts. The utility responded by requiring two-person confirmation before AI-generated switching sequences execute. “We duplicated the human redundancy we already had in nuclear,” CISO Kadri Kask explained.

The power sector flirted with artificial intelligence as early as the 1990s, when expert systems attempted to diagnose transformer faults. Early projects collapsed because control-room culture rewarded intuition built on decades of sound checks and manual pageantry. Engineers quietly unplugged the programs rather than argue with machines during storms.

The 2021 Texas freeze reignited interest after rotating blackouts left 4.5 million without heat. Investigators found that grid operators misjudged how fast gas compressors would trip, a mistake simulation software later showed an AI forecaster caught in 72% of rerun scenarios. Texas lawmakers mandated a technology review, fueling the current wave of pilots.

Texas ERCOT will vote in September on whether to embed AI advisers inside its live market floor, a move that could shift 5% of daily generation dispatch to machine-led decisions worth $450 million per year. If approved, similar grid operators across the U.S. Southeast plan coordinated roll-outs before the 2027 summer peak.

Wednesday’s conference ended with vendors hawking headsets that read dispatcher brainwaves to predict fatigue, but most utilities say they will stick to simpler dashboards for now. The lesson from California is plain: algorithms earn their keep only when humans stay awake enough to overrule them.