AI for utilities: where it works today, and what it needs

Most of what is written about AI for utilities is either a vendor promise or a warning. This is neither. It is a survey, from the practitioner's side, of where machine learning and language models are producing measurable results in utilities and energy management today, what each use case needs in order to work, and where the same tools fail in ways that matter to a regulated business. The International Energy Agency's own assessment is blunt on both counts: the energy sector can and should do more to seize the potential benefits of AI, but faces barriers including missing or inadequate access to data and digital infrastructure and skills, and persistent digital and physical security concerns, and the prevalence of AI-related skills is much lower in the energy sector than in other sectors.

The use cases below are ordered roughly by maturity. The pattern that runs through all of them is that the model is the easy part. The hard parts are clean historical data, a governance layer that knows what the model does not know, provenance for every output, and a human who reviews the ones that matter.

Where it is delivering

Document and bill data extraction

The most mature use case in energy management is reading documents: utility bills, interconnection agreements, tariff sheets, meter certificates. It is mature because the task is bounded, the ground truth is on the page, and the output can be checked arithmetically. Multimodal models that reason over a bill's layout, rather than matching a template, now read invoices from utilities they have never seen and return typed fields: periods, reads, consumption, demand, each rate line, taxes and adjustments. What makes it safe is the validation that follows: line items must sum to subtotals, reads must chain from the previous bill, and charges must recompute from the tariff. A field the model got wrong fails a check; it does not get paid. The comparison of manual entry, OCR and reasoning covers why template OCR does not scale to the thousands of bill layouts in North America.

In accounts payable more broadly, adoption is already the norm rather than the exception: Ardent Partners' 2025 research found that 75 percent of AP departments use some form of AI or automation. Utility invoices are the hard subset of that work because there is no purchase order to match against.

Meter data validation, estimation and editing

Utilities have run validation, estimation and editing (VEE) rules on meter data for decades. Machine learning adds two things: better estimates for gaps, because a model trained on the meter's own history and weather fills a missing interval more credibly than a straight-line rule, and anomaly detection that learns each meter's normal pattern instead of applying one fixed threshold to every account. On the customer side the same techniques catch stuck meters, wrong multipliers and estimated reads carried as fact. This works because interval data is plentiful, structured and numeric. It fails when the training history itself contains uncorrected errors, which the model then learns as normal.

Load, generation and price forecasting

Forecasting is where machine learning has the longest track record in utilities and the clearest economic case. The US Department of Energy's 2024 AI for Energy report lists improving variable renewable energy forecasting for solar, wind and run-of-river hydro, and improving demand forecasting using AI trained on historical weather, climate, economic and load data_043024.pdf) among the near-term grid opportunities. The IEA reaches the same conclusion, noting that AI can improve the forecasting and integration of variable renewable generation, reducing curtailment and emissions. For a facility or portfolio, the same models forecast next month's consumption and demand from history and weather, which is the basis of budgeting from bill data and of demand-charge management. The failure mode is drift: a model trained before an electrification project, a tenant change or a rate redesign forecasts the old building.

Outage and fault prediction

Grid operators are using classification models on outage history, asset condition, vegetation and weather to predict where faults will occur and how long restoration will take. The IEA reports that AI-based fault detection can help rapidly identify and precisely pinpoint grid faults, reducing outage durations by 30 to 50 percent, and that remote sensing with AI-based management could make up to 175 gigawatts of transmission capacity available without any new lines being built. The stakes are large: the DOE report notes that outages already cost American businesses $150 billion annually_043024.pdf). Peer-reviewed results are encouraging but dataset-specific; one study on a distribution utility's outage records reported a gradient-boosting model that predicted outage duration classes with 98.4 percent accuracy, against an 85.5 percent average for prior methods on different datasets. Read that as evidence the approach works on clean utility data, not as a number to expect on yours.

Customer service

Language models are answering billing, outage-status and program-eligibility questions in utility contact channels, and drafting agent responses. This is a real productivity gain with a specific risk: a fluent, wrong answer about a bill or a rate is a regulatory complaint. The UK regulator Ofgem published good-practice guidance on AI in the energy sector on 20 May 2025, updated in May 2026, and its consumer-interaction considerations focus on embedding explainability consistent with the risk profile and ensuring ongoing transparency to consumers about the use of AI. The practical rule is that a customer-facing model should answer from the customer's own account data and approved content, cite what it used, and hand off when its confidence is low.

Tariff, program and regulation question answering

Inside utilities, consultancies and large energy users, the same models are being pointed at tariff schedules, regulatory codes, incentive program rules and internal procedures so that staff can ask a question and get an answer with the source. This is where general-purpose models fail most expensively, because tariffs are precise, change several times a year, and vary by jurisdiction, and a model trained on the open internet will produce a confident answer that was true two rate orders ago. The approach that works is a governed, curated knowledge base with retrieval from the current documents, a confidence score, and a citation on every answer, which is the design described in trusted-domain AI for energy.

What every use case needs

Clean historical data. The DOE report is candid that for grid assets, records may not exist, undocumented maintenance during outages may make the data inaccurate, and available data may suffer from completeness or formatting issues that render it unusable_043024.pdf). The same is true of a building portfolio's utility history: if the baseline year is PDFs and a spreadsheet with month labels instead of periods, no model can forecast from it. Data quality is the project; the model is a week at the end of it.

Governance. Someone must own which model is used for what, how it was tested, how it is monitored and who can override it. The NIST AI Risk Management Framework, released on 26 January 2023, organizes this into four functions, Govern, Map, Measure and Manage, and its Generative AI Profile published on 26 July 2024 addresses the risks specific to language models. Canada's federal Directive on Automated Decision-Making, in force since 1 April 2019, requires departments to test the data and the underlying model for accuracy and unintended bias before production, monitor outcomes on a schedule, and document human overrides, with impact levels that determine how much human involvement a decision needs. Neither binds a private utility directly, but both are a ready-made structure for one.

Provenance. Every output should carry where it came from: the bill page for an extracted field, the meter and period for a forecast input, the document section for a tariff answer. Without it, a reviewer cannot check the output in less time than it took to produce.

Human review, proportionate to risk. The DOE report lists human-in-the-loop oversight, rigorously validated systems, and physics-informed, explainable models_043024.pdf) among the prerequisites for AI on the grid. In practice this means routing the low-confidence outputs to a person and letting the high-confidence ones flow, and measuring both.

The failure modes to plan for

Hallucinated tariff and regulatory answers

The DOE report puts it plainly: foundation models have stochastic outputs and can produce hallucinations, outputs that seem authoritative but are factually incorrect_043024.pdf). A rate class, a ratchet percentage or a program deadline is exactly the kind of fact a general model will invent. The mitigation is architectural: retrieve from the current approved documents, refuse when nothing relevant is found, and attach a citation and a confidence score so the reader knows which answers to check.

Model drift

Forecasting and anomaly models decay as the world changes. An electrification project, a new tenant, a tariff redesign or a pandemic year each shift the relationship the model learned. Drift is managed with monitoring, scheduled retraining, and a log of known change events so that non-routine adjustments can be made deliberately rather than discovered as an unexplained error.

Privacy and data residency

Interval data reveals occupancy and behavior; customer account data is personal information under Canadian and provincial privacy law. Sending either to a third-party model endpoint outside the organization's control is a decision, not a default. The IEA lists digital and physical security concerns that often trump potential efficiency gains among the sector's barriers, and for regulated utilities that caution is appropriate. Deployment in the customer's own environment, or a sovereign hosted one, with the same controls that apply to any customer utility data, removes most of the objection.

A governance checklist

1. Confidence scores on every output, with a threshold below which a human reviews before anything is acted on
2. Source citations: the bill page, meter and period, or document section behind each answer or extracted value
3. A curated, versioned knowledge base for tariff and regulatory questions, with a refusal path when no relevant source exists
4. Deterministic validation after extraction: arithmetic, continuity and tariff recomputation checks that do not depend on the model
5. On-premise or sovereign deployment for customer account data and interval data, with documented data flows
6. An audit trail of inputs, outputs, confidence, reviewer and override for every decision the system touched
7. Monitoring for drift, with scheduled retraining and a log of known change events
8. A named owner and a risk classification per use case, following NIST AI RMF or the Canadian directive's impact levels

How to evaluate a vendor

Give them your worst data, not their demo set. Three bills from utilities they have never seen. A meter with a known multiplier error. A tariff question whose answer changed in the last rate order. Then ask three things: what confidence did the system assign to each output, what source did it cite, and what happened to the ones it was unsure about. A vendor who cannot answer those questions is selling a model; a vendor who can is selling a governed system.