AI / ML · Data science & classical ML · AI agentEnergy & utilities
Utility Dive (Oct.
Source [1] Utility Dive · 22 Oct 2026 · “Utilities Turn to AI to Manage Rising Power Demand — from AI” ↗
Adoption stage: Pilot — A named organisation is trialling it on real data, bounded scope or time.
Summary
Utility Dive (Oct. 22, 2026) reports that U.S. utilities are a long way from letting large language models take actions that affect physical infrastructure; most AI there is narrow decision-support such as load forecasting and vegetation management, and National Grid Partners' Pradeep Tagare says agentic architecture has not been widely deployed on anything touching physical assets. A Guardrail Technologies review of federal filings found about 70% of energy and utility companies call AI a cybersecurity risk but only 23% of utilities and 9% of energy companies documented a process to manage it, while New York regulators have ordered AI-use inventories and Connecticut officials petitioned for a similar review on Oct. 13. On the deployment side, Portland General Electric says GridCARE's hourly demand modelling found more than 80 MW of data-centre connection capacity years earlier than expected, nearly 80% of utility leaders in NGP's US/UK/Europe survey report at least one fully deployed AI application for large-load planning, and machine-learning virtual power plants have been in use for about five years.
Kenny's take
In the grid operations this article describes, what actually runs is forecasting and demand modelling; the agents stop short of the physical assets. The one named case is PGE with GridCARE, and it sits at pilot. The sharper finding is the audit gap: if utilities cannot reconstruct what an AI system read and did, the open question is whether that record has to exist before agents move toward the assets. This take is wrong if New York's AI-use inventories, once public, show agents already directly controlling physical assets in operation.
AI draft · approved by Kenny
Analysis
Stage: pilot for the one named operator case — Portland General Electric reports a result from its work with GridCARE's hourly demand modelling, but the article does not say the model runs in PGE's day-to-day planning. The survey claim of "fully deployed" AI applications is unattributed to any named utility, so it cannot set a production stage on its own.
For grid-ml-decision-support in energy, the article's picture is that the AI in use is forecasting, demand modelling and machine-learning coordination of distributed resources, supporting human decisions; the agentic and LLM layer is described as not yet deployed on physical assets. That speaks directly to q-ds-vs-llm: in this article the classical ML is what runs, and the LLM is what executives say they are not yet letting act.
For ai-decision-audit-trail, the evidence is a gap, not an adoption: Guardrail's 23% / 9% documented-process figures and Marlin's statement that most utilities "could not reconstruct" an AI system's reads and actions. The regulatory moves named (New York inventories, the Connecticut petition) are about knowing what AI is running, which Render Networks' CEO says is "still unclear".
Facts from the source
- F1Utility Dive (Oct. 22, 2026) reports that "The utility industry is a long way from allowing large language models to take action affecting physical infrastructure, executives say", and that most AI deployments in the sector "are niche tools designed to support human decision-making on topics like load forecasting and vegetation management." [1]
- F2Pradeep Tagare, interim president of National Grid Partners: "[For] anything that touches the physical assets directly, we haven't seen widespread deployment of agentic architecture yet." [1]
- F3A review of federal regulatory filings by Guardrail Technologies found "roughly 70% of energy and utility companies identified AI as a cybersecurity risk — but just 23% of utilities and 9% of energy companies documented a process for managing that risk." Guardrail CEO T.J. Marlin: "Most utilities could not reconstruct what an AI system read, what it did and why." [1]
- F4New York regulators "recently directed utilities to file inventories of their AI use"; on Oct. 13 Connecticut's Attorney General and Consumer Counsel petitioned state regulators "to open a similar review." [1]
- F5Stephen Rose, CEO of Render Networks: "What's still unclear is how much AI is already running and what data it's running on." [1]
- F6GridCARE "uses hourly demand modeling to identify available grid capacity"; Portland General Electric "said its work with GridCARE identified more than 80 MW of capacity for data center connections years earlier than initially expected." [1]
- F7"Nearly 80% of utility leaders NGP surveyed this year" in the US, UK and Europe said their organisations had "fully deployed at least one AI application for large-load customer planning"; virtual power plants using machine-learning systems "have been in use for about five years". [1]
Organisations: Portland General Electric, GridCARE
Counterpoint — verification notes
Codex checked each sentence against the source and searched for counter-evidence. Verdict: Codex check · partly supported.
No counter-evidence was recorded.
5 sentence-level checks
- summary_enpartly-supported: U.S. utilities are a long way from letting LLMs act on physical infrastructure. — The article states this for the utility industry without limiting the observation to U.S. utilities.
- summary_kopartly-supported: 미국 전력회사들은 LLM이 물리 설비에 영향을 주는 조치를 하도록 허용하는 단계와는 거리가 멀다. — The Korean summary adds the same U.S.-only scope that the source does not specify.
- analysispartly-supported: The agentic and LLM layer is not yet deployed on physical assets. — The source says agentic deployment is not widespread, which does not establish that none exists.
- takepartly-supported: 에이전트는 물리 설비 앞에서 멈춰 있다. — This reads as a categorical boundary; the source says such deployment is not widespread.
- takepartly-supported: The agents stop short of the physical assets. — This reads as a categorical boundary; the source says such deployment is not widespread.
Sources
Keep reading
Stay in the loop.