Energy and utility operations: AI on the grid, on the meter

Utility companies — electric, gas, water — sit at a peculiar intersection of operational AI use cases. They run critical infrastructure that touches most of the population. They generate enormous volumes of operational data. They are heavily regulated. They have customer relationships that involve PII and billing data. They face increasing complexity in their operations as the energy transition accelerates.

AI has been adopted across the industry in the past few years. The bill is real, the operational data is uniquely sensitive, the regulatory frameworks are strict, and the case for moving inference local is straightforward.

The work

Utility AI workloads span operational, customer-facing, and regulatory categories.

Grid operations: monitoring grid health, identifying anomalies, predicting equipment failures, drafting incident reports. The data flows are continuous, the operational stakes are high, and the latency requirements are real.

Demand forecasting and load management: predicting demand patterns, optimizing dispatch, identifying conservation opportunities. The data is large-scale time series; the AI documentation and analysis layer sits on top.

Outage management: when something goes wrong, AI helps classify the issue, predict restoration time, draft customer communications, and produce incident reports for regulators.

Customer communications: utility bills are complex. Customer service interactions involve everything from billing questions to outage reports to service requests. AI augments these communications at scale.

Regulatory reporting: utilities operate under strict reporting requirements from state utility commissions, federal regulators, and (for nuclear) safety authorities. AI helps draft regulatory filings, monitor compliance, and prepare for rate cases.

Field operations: dispatching technicians, drafting work orders, capturing field reports, integrating with asset management systems. AI augments the documentation that flows between the field and the office.

The math

A representative midsize regional utility — say, a million customer accounts — generates a meaningful AI workload across all these functions.

Aggregate AI operations per year: somewhere between a hundred million and a billion tokens, depending on the depth of AI integration. At frontier pricing, the bill is somewhere in the low to mid six figures per year.

For larger utilities — investor-owned utilities with multiple states of operation, or large municipal systems — the bill scales to the high six figures or low seven figures per year. For the largest holding companies, well into seven figures.

These numbers exclude the specialized SCADA, OMS (outage management), and AMI (advanced metering) systems that generate the underlying data. The AI augmentation layer on top is the line item we're examining.

Why utilities are structurally a local-SLM case

The standard properties for local-SLM suitability are present, with several that are unusually strong in the utility context.

The work is narrow within each operational area. A model fine-tuned on the utility's own grid data, customer communication history, and regulatory filings outperforms a general model on the utility's specific work.

The work is repetitive. Grid anomalies follow predictable patterns. Customer service interactions cluster into a small number of categories. Regulatory filings follow predictable formats. Specialization compounds.

The privacy story is real and specific. Customer billing data, usage patterns, and service history are PII. Aggregated, this data reveals patterns about customers' lives — when they're home, what appliances they use, when their habits change. For regulated utilities, the privacy framework is explicit; sending this data to a third-party cloud LLM is a posture that the utility commission and the customers will both have opinions about.

The critical infrastructure designation is the under-discussed argument. Utilities are designated as critical infrastructure in most jurisdictions. The data they generate about grid operations is, in some cases, controlled under specific regulatory regimes (NERC CIP for electric utilities in North America, equivalent frameworks elsewhere). Sending operational data — load profiles, equipment status, vulnerability information — to a third-party cloud LLM creates national security questions that the utility's CISO and the relevant regulators take seriously.

The latency story matters in operational workflows. Grid anomalies need to be analyzed in seconds, not minutes. Outage classification needs to be near-real-time. Field operations need responsive AI on devices that may have poor connectivity.

What changes with local inference

A utility AI workflow on a local SLM looks like this.

A model is fine-tuned on the utility's operational, customer, and regulatory corpus. The fine-tuning respects the customer privacy framework and the critical infrastructure designation.

The model runs on infrastructure the utility controls — on-premises in operations centers, in private clouds meeting regulatory requirements, and at the edge for field operations. The deployment meets NERC CIP and equivalent requirements.

Operational data flows through the inference pipeline within the utility's controlled environment. The model produces classifications, summaries, drafts, and analyses. Customer communications and regulatory filings get drafted with the model's assistance.

The cost flips from per-operation to fixed. Operational complexity can grow without the AI bill spiking.

The critical infrastructure data stays inside. The customer privacy framework remains intact. The regulator's questions about data residency get easy answers.

The energy transition wrinkle

The energy transition is changing what utility operations look like, in ways that strengthen the local-SLM case.

Distributed energy resources (solar, batteries, EVs) are proliferating. Each one generates operational data. Each one needs to be coordinated with the broader grid. The complexity of grid management is increasing faster than the workforce that operates the grid.

The volume of data is exploding. Smart meters generate readings every fifteen minutes (or more frequently). Phasor measurement units generate data continuously. Distribution sensors are being deployed at scale.

The latency requirements are intensifying. As the grid becomes more dynamic, the operational AI has to keep up. Cloud-LLM round trips are increasingly incompatible with the speed at which grid decisions need to be made.

The energy transition makes the local-SLM case stronger over time. Utilities that build for local inference are also building for the operational reality of the next decade of grid management.

Where the cloud LLM is still acceptable

A few cases.

For research and planning workflows operating on aggregated, non-customer-identifying data. Some scenario analysis can be designed this way.

For customer-facing workflows that touch only the public-facing parts of the utility's operations — say, public-facing FAQs, public service announcement drafting.

For pilot deployments and early validation work where the infrastructure investment doesn't pay back and the utility has explicitly accepted the compliance risk.

For the bulk of utility AI work — the grid operations, the customer communications, the regulatory filings, the field operations — the local-SLM case is strong on cost, on privacy, on critical infrastructure compliance, and on latency.

The pattern, on the grid

Avery NXR is not a utility tool. It scaffolds Next.js applications. The architectural pattern repeats.

Utility AI is a narrow (within each function), repetitive, high-volume, privacy-sensitive, critical-infrastructure-relevant, latency-relevant workload. Every dimension favors local inference. The energy transition strengthens the case over time.

The utility AI vendors that build on local infrastructure — with appropriate fine-tuning across utility functions, NERC CIP-compliant deployment, and integration with the major utility software systems — will own the institutional utility AI market. The cloud-LLM-default products will struggle against the critical infrastructure framework as it tightens.

The pattern continues. Utilities are one of the workflows where the architectural shift is being driven by critical infrastructure and customer privacy as much as by cost — and where the energy transition makes the case stronger every year.