A state public utility commission staff attorney is cross-examining your expert witness. She points to a forecast number in your direct testimony and asks a simple question: where does this number come from? If the answer is "an AI tool generated it," and your witness cannot explain the methodology, the training data, the uncertainty band, and the verification steps your team ran before accepting it, you are about to have a bad day at the commission.

How a State PUC Works: The Forum You Are Filing Into

A state public utility commission (PUC) is the regulatory body responsible for overseeing the rates, service quality, and practices of electric utilities within the state's jurisdiction. In most states, the PUC has authority over retail rates charged to residential, commercial, and industrial customers, infrastructure investments the utility seeks to recover through rates, tariffs governing how different customer classes are served, and integrated resource plans that describe how the utility intends to meet future electricity demand.

A rate case is a formal proceeding before the commission in which the utility proposes changes to its revenue requirement and the rates it charges customers. Rate cases are adversarial proceedings: the utility files testimony and supporting data, other parties (consumer advocates, large industrial customers, environmental groups, state agencies) file responsive testimony and data requests, witnesses are cross-examined, and the commission staff independently analyzes the record before issuing a decision.

The rate case is the most rigorous scrutiny your utility's planning, forecasting, and operational data will ever face. Everything you file is public. Everything you claim can be challenged. Every number you present will be traced back to its source by opposing counsel. The evidentiary standard is not "this looks reasonable." It is: can you demonstrate that this is accurate, and can you explain how you know that?

Understanding this forum is essential context for understanding how AI-assisted filings work, and where they fail.

The AI-Assisted Forecast in a Rate Case: What the Commission Will Ask

Utilities are increasingly using AI-based load forecasting tools to produce the demand projections that underpin rate-case testimony on infrastructure investment, capacity needs, and revenue requirement. An AI-based forecast that achieves roughly 1 to 2 percent MAPE at the day-ahead horizon can produce materially better planning-horizon forecasts than traditional statistical models, particularly when large data-center step loads are in the picture. That accuracy advantage is real and can be presented in testimony.

But the commission will ask more than "what is your MAPE?" Here is the specific line of scrutiny you should anticipate:

What Model Produced This Forecast?

Commissions and intervenors will want to know: what type of model is it (neural network, ensemble, gradient boosting)? Who built it, and when? Has it been independently validated? What is its track record against actual outcomes in your service territory? A model purchased from a vendor and applied to your territory without local validation is a harder case to make than a model that has been producing forecasts for your territory for 18 months and whose accuracy has been documented.

This is not a theoretical concern. In recent rate cases across multiple states, intervenors have hired their own AI experts to challenge utility forecasting models, arguing that the model's training data did not include relevant local conditions, that the model overweighted recent data in a way that inflated the forward estimate, or that the uncertainty band around the forecast was not disclosed. Each of these challenges can be developed into a productive cross-examination that raises doubt about the validity of the entire forecast.

What Training Data Was Used?

A forecast is only as good as the data it was trained on. If your AI forecasting model was trained on national data or regional data that does not reflect your service territory's specific characteristics (climate, economic mix, industrial load profile, DER penetration), the commission will want to understand how that limitation was addressed. If the model was trained on data from a period before large data-center load existed in your territory, and you are now forecasting a period when data centers will represent a major portion of your load growth, the commission will want to understand how the model accounts for a regime change it has never seen.

This is the step-load problem expressed as a rate-case evidentiary challenge. The same forecasting deficiency that creates operational risk for the load forecaster creates legal risk for the rate-case team. The defense is to show your work: document how the model handles large-load additions, show that you have incorporated the FERC Large-Load Rule application pipeline as a conditional scenario, and demonstrate that you have tested the model against analogous step-load events from other territories if your own history does not include them.

Who Verified the Output?

This is the cardinal rule question. The commission does not care if an AI produced a number. It cares whether a qualified, accountable human reviewed, validated, and signed off on that number before it appeared in testimony. If your witness cannot testify that they personally reviewed the forecast methodology, ran the model's output against known benchmarks, and made a professional judgment that the forecast is reasonable, your testimony is on thin ice.

The phrase "the model produced this number" is never sufficient in a rate case. The correct phrasing is: "the model produced this output; our forecasting team then verified it against the following benchmarks and concluded it is a reasonable estimate of future demand." The verification steps need to be documented, the benchmarks need to be specified, and the person who performed the verification needs to be the one testifying to it.

The Data-Center Tariff Question: A New Kind of Rate-Case Battle

The arrival of large AI data centers as utility customers has created a new category of rate-case contest that will occupy state commissions for years: how should the costs of serving large, high-load-factor compute customers be allocated between those customers and other ratepayers?

The question has several dimensions:

Cost Causation and Rate Design

A hyperscale data center drawing 300 MW at a 95 percent load factor creates infrastructure costs that, under traditional utility cost-of-service ratemaking, should largely be borne by that customer. The principle is cost causation: a customer who drives infrastructure investment should pay for it. The challenge is that traditional rate design was developed for customers whose load patterns the system was already built to serve, not for customers whose load profile is qualitatively different from anything in the utility's history.

State commissions are actively working through several contested questions: Should data centers pay a standalone service rate that fully allocates the cost of the infrastructure serving them? Should they receive favorable rates to encourage economic development, with other ratepayers subsidizing part of the infrastructure cost? Should they be required to provide demand response or flexible load capability as a condition of the favorable rate structure? The answers vary by state, and they are being made in rate cases and tariff proceedings right now.

The AI-Assisted Filing for Tariff Proposals

Drafting a large-load tariff and defending it before a state commission is a document-intensive process: the tariff text itself, the supporting rate design study, the load research data, the cost-of-service allocations, and the expert testimony on each component. AI tools can accelerate the drafting of all of these components. A large language model can produce a first draft of tariff language, a cost-of-service study narrative, and expert testimony outlines in a fraction of the time traditional drafting takes.

The risk is in the specific numbers. An AI tool that drafts testimony has no inherent knowledge of your utility's cost-of-service structure, your commission's historical positions on rate design, or the specific data-request responses from intervenors in your current proceeding. It will produce plausible-sounding text that may contain numbers, citations, and characterizations that are simply wrong for your situation. Submitting AI-drafted testimony without systematic verification of every specific claim, number, and regulatory citation is one of the fastest ways to damage your utility's credibility before the commission.

The correct use of AI in this context is as a drafting accelerator, not as a substantive analyst. The AI produces the structure and the boilerplate. Your rate-case team verifies every number against source data, every regulatory citation against the actual proceeding, and every characterization against your utility's factual record. The verification step is not optional and cannot be delegated to the AI.

Worked Example: The AI Forecast That Nearly Failed

A Midwestern IOU is before its state commission seeking a rate increase driven primarily by a large transmission capital program. The commission's primary concern is whether the utility's load forecast justifies the investment. The utility's direct testimony presents an AI-based five-year peak demand forecast showing 15 percent growth, driven largely by data-center additions in the service territory.

The consumer advocate's expert files testimony challenging the forecast. She identifies three specific problems: first, the AI model was trained on national utility data and has not been locally validated against the utility's historical actuals; second, the data-center load projections are based on announced facility plans, not on actual interconnection applications, meaning the forecast includes speculative load that has not been formally filed; third, the uncertainty band around the forecast was never disclosed, meaning the commission cannot assess the range of outcomes around the 15 percent growth number.

None of these are fatal objections. They are correctable. But they were correctable before direct testimony was filed, not after. A witness who cannot answer "when was this model validated against our territory's actuals?" under cross-examination cannot rehabilitate the testimony in rebuttal. The rate case is delayed while the utility submits supplemental data responses, and the commission ultimately conditions the capital program approval on the utility conducting a local model validation study before its next rate case.

Here is how an informed rate-case team would have handled this differently. Before filing direct testimony, they would have: obtained the AI model's validation report, checking whether it covers the utility's specific service territory or a comparable territory; reviewed the data-center load projections, confirming that each facility above a threshold is supported by an interconnection application or a signed LOI, and labeling speculative load explicitly; and run the forecast at three confidence levels (low, base, high) and reported the uncertainty band in the testimony along with the assumption set for each scenario. The AI tool still produces the forecast. The rate-case team does the preparation work that makes the forecast defensible.

In a rate case, a forecast is not a number. It is a number plus a methodology, plus a validation record, plus an uncertainty band, plus the name of the engineer who reviewed it. Strip any of those elements away and the number becomes an opinion that opposing counsel will dismantle.

What to Show a Commission: Building the Evidentiary Record

When you are presenting AI-assisted work product in a rate case, the evidentiary record you need to build has five components:

• Model description: A plain-language description of what type of model was used, who built it, when, and what data it was trained on. This does not need to be a technical paper; it needs to answer the questions a non-technical commission staff member will ask.
• Validation record: Evidence that the model has been tested against actual outcomes in a comparable territory or time period, showing its accuracy (MAPE or similar metric) on a known dataset. If the model has been validated locally, present that validation. If it has not, explain what comparable validation exists and acknowledge the limitation.
• Input data documentation: A clear description of the input data used to produce this specific forecast, including sources, dates, and any adjustments made to the raw data. Large-load additions should be sourced to interconnection applications or formal customer commitments, not to press releases or announced plans.
• Uncertainty band: A presentation of the forecast at multiple confidence levels or scenarios, showing the commission the range of possible outcomes and the assumptions that drive each. A commission that approves a capital program is approving an investment that will be in service for decades; they need to understand the planning uncertainty they are accepting.
• Human verification statement: Testimony from the responsible professional that they personally reviewed the model's output, verified the inputs, and concluded the forecast is a reasonable basis for the proposed investment. This is the signature on the work product that transforms an AI output into defensible expert testimony.

Building the Evidentiary Record Before You Need It

The single most common mistake regulatory affairs teams make with AI-assisted filings is treating the evidentiary record as a post-filing task. The model runs, the testimony draft appears, and then someone asks: do we have documentation for this? At that point, the filing is days away, the discovery deadline is weeks away, and the answer is often: not in a form we can produce in a data request.

The evidentiary record needs to be built as a parallel track to the model development, not retrospectively. That means: documenting the model's architecture and training data at the time the model is deployed, not when the first data request arrives; logging the validation results in a retrievable format with the timestamp and testing methodology attached; documenting the input data sources for each specific forecast run, so that the source for every major input variable is traceable; and creating the human verification record as a contemporaneous note, not a reconstruction written the night before the hearing.

These are process changes, not technology changes. They require the load forecasting team and the regulatory affairs team to coordinate their workflows before the rate case is filed. Utilities that build this coordination into their standard rate-case preparation cycle will be in a structurally stronger evidentiary position than those that treat AI governance as an ad hoc response to regulatory pressure. The commission does not see the difference between a utility that has governed AI forecasting rigorously for three years and one that assembled its documentation after the intervenors asked for it, but the utility's witnesses certainly feel the difference under cross-examination.

Key Takeaways

• State public utility commissions are adversarial forums where every number in a utility's testimony is subject to cross-examination, data requests, and challenge by intervenors with their own technical experts.
• AI-assisted load forecasts can be used in rate-case testimony but must be accompanied by model description, local validation record, documented input data sources, disclosed uncertainty bands, and personal verification by a named expert witness.
• The data-center tariff question (how to allocate costs between large compute customers and other ratepayers) is an active contested issue in state rate cases and tariff proceedings, with answers varying by state and evolving through individual proceedings.
• AI drafting tools can accelerate the production of tariff text, cost-of-service narratives, and testimony outlines, but every specific number, regulatory citation, and factual characterization must be verified against source data before filing.
• The most common AI-assisted filing failure is submitting AI-drafted testimony that contains plausible-sounding but wrong numbers, unsupported citations, or characterizations that do not match your utility's factual record.
• The phrase "the model produced this number" is never sufficient in testimony. The required statement is: "the model produced this output; our team verified it against these benchmarks and I personally conclude it is a reasonable estimate."
• Building the evidentiary record before filing direct testimony, not after intervenors challenge it, is the difference between a smooth rate case and a delayed, conditioned decision that requires supplemental filings and damages the utility's credibility.