The rate-case docket is open, the deadline for your direct testimony is in 72 hours, and opposing counsel just served a 47-item data request at 5 p.m. This is the moment when AI either becomes your most useful drafting partner or your most expensive liability, depending entirely on how you use it.

The Rate Case as a Document Factory

A general rate case at a mid-size investor-owned utility is one of the most document-intensive processes in American regulated industry. The filing package alone can exceed 10,000 pages: cost-of-service studies, rate design schedules, workpapers, policy statements, and several volumes of testimony from the utility's witnesses covering topics from revenue requirement to depreciation to reliability investment. Then comes discovery. Opposing parties, which include the state commission's staff, the attorney general's consumer advocate, large industrial customers, and environmental intervenors, issue rounds of data requests. A contested rate case may generate 1,000 or more data requests, each requiring a factual, documented response within a regulatory deadline that rarely exceeds 10 business days.

Regulatory affairs teams at utilities have traditionally handled this volume through sheer labor: staff attorneys, rate analysts, and subject-matter experts pulling all-nighters to produce testimony drafts and data-request responses. The quality is often uneven, because the person who knows the revenue requirement calculation cold may write a technically precise but unreadable response, while the person who writes clearly may not understand the calculation well enough to verify the numbers.

Generative AI enters this environment as a genuine productivity tool, with one absolute condition: every number in an AI-produced filing must be flagged for source verification before the document leaves the building. A rate case is a sworn proceeding. Testimony is presented under oath or under penalty of perjury. A data-request response that misquotes a number from the cost-of-service study, even if the error was the AI's and the human assumed it was correct, is still a false statement in a regulatory proceeding. The professional and legal consequences are real, and "the AI produced the number" is not a mitigating factor. It is an aggravating one, because it suggests the entity filed without verification.

What AI Does Well in Rate-Case Work

AI has genuine strengths in the rate-case environment that justify deploying it systematically. The key is knowing which tasks are safe to delegate to AI and which require the AI output to serve only as a starting point for expert review. A useful mental model divides rate-case document tasks into two categories: tasks that require domain knowledge of the specific numbers (where AI is a useful starting point only if the number-flagging discipline is applied), and tasks that require structural and language competence (where AI is a genuine accelerant). Testimony structuring, discovery response formatting, cross-testimony matrices, and commission correspondence are primarily structural tasks. Revenue requirement calculations, cost-allocation decisions, and rate design tradeoffs are substantive expert tasks where AI contributes language and organization, not analysis.

Drafting Testimony Summaries

Direct testimony in a rate case is dense and highly technical. A utility typically files testimony from 10 to 20 witnesses covering topics as varied as demand forecasting methodology, depreciation curves, vegetation management cost trends, and rate design equity. Summarizing these testimonies into clear, consistent executive summaries, commission briefings, or cross-testimony matrices (a table showing how each witness's testimony relates to the overall revenue requirement case) is time-consuming and requires reading comprehension more than specialized expertise.

AI excels here. Given the full testimony text, the model can produce a structured summary that captures each witness's key assertions, the analytical basis for each, and the exhibit references. The verification step is straightforward: the rate analyst reads the summary against the original testimony to confirm no assertions were dropped, overstated, or softened, and that every exhibit citation is accurate.

Structuring Data-Request Responses

Data requests in a rate case follow a standard format: the party issues a numbered request, and the responding party provides a factual answer, often with attached supporting documents. The structure is formulaic: restate the request, provide the responsive information, cite the workpaper or exhibit from which the answer is drawn, and note any objections to scope.

AI can produce the structural skeleton of a data-request response at high speed: restating the question in compliant format, drafting the boilerplate preambles and objections where applicable, and organizing the response around the data points the subject-matter expert provides. The expert then fills in the actual numbers and document citations, the AI produces a coherent response document, and the attorney reviews for privilege and accuracy before filing.

Pre-Filing Consistency Checking

One of the most underused AI applications in rate-case work is consistency checking: comparing numbers across exhibits, testimony, and workpapers to confirm they are internally consistent. A cost-of-service study that shows total distribution O&M of $347 million but has a witness testimony citing $348 million is an example of the kind of discrepancy that creates discovery ammunition for opposing counsel. AI can scan documents for numerical citations and flag apparent inconsistencies for human review.

The Number-Flagging Discipline: Every Number, Every Time

The most important operational rule for AI-assisted rate-case work is absolute and admits no exceptions: every number in an AI-produced draft must be flagged for source verification before the document is filed. Not most numbers. Every number.

This rule exists because AI's most dangerous failure mode in document drafting is the confident, plausible-looking numerical error. The model has read thousands of rate cases, regulatory filings, and financial documents in its training data. It knows what a revenue requirement discussion looks like, it knows the structure of a depreciation study, and it can produce a paragraph about return on equity that sounds authoritative and cites numbers that are in the right ballpark for the industry. But the specific numbers in your rate case are derived from your utility's specific cost data, and those numbers are in your workpapers, not in the model's training data. Any number the model produces without being given the exact workpaper data is either interpolated from training data or invented. Either way, it must be verified.

A practical implementation of the number-flagging discipline uses a markup protocol. When the AI produces a draft, every numerical value is automatically highlighted in a different color, and the draft is routed to the subject-matter expert responsible for that section with a cover note: "These highlighted values require source citation and verification before this document may be filed." The expert opens the relevant workpaper, confirms each number, and either approves it or corrects it. The signed verification form is kept in the filing management system as part of the document's audit trail.

A rate case is a sworn proceeding. Every number in a filing carries the weight of testimony. Verify every number before it leaves the building, regardless of whether AI or a human produced the first draft.

The FERC Large-Load Rule and AI-Assisted Drafting

The FERC large-load rulemaking, which addresses how loads over 20 MW connect to the transmission grid, is the most significant new regulatory development for utility rate cases in a generation. FERC committed in April 2026 to issue the rule by the end of June 2026 (Docket RM26-4-000), resetting the interconnection policy framework that governs how data centers and other large electricity consumers take transmission service. This rulemaking is generating an enormous volume of comment filings, reply comments, and technical exhibits, and it is directly affecting how utilities structure their rate cases for large commercial and industrial customers.

For regulatory affairs professionals, the FERC large-load rulemaking creates several specific AI-drafting opportunities and pitfalls:

• Opportunity: comment drafting assistance. The FERC comment record is publicly available, and AI can be used to summarize prior comments, identify arguments made by similar utilities, and draft the structural outline of a comment filing. The attorney then reviews for accuracy and legal strategy, and engineers verify any technical claims.
• Pitfall: AI summarizing the rule incorrectly. The FERC large-load rulemaking is evolving rapidly in 2026. An AI tool trained before the final rule is issued will not know the final rule's terms and may produce a summary based on the proposed rule, an earlier notice of proposed rulemaking, or commentary about the proceeding that does not accurately reflect FERC's actual order. Before any filing references the FERC large-load rule, the rate analyst must read the actual FERC order and confirm that the citations are to the final text, not to earlier-stage documents.
• Opportunity: data-center tariff analysis. Utilities designing tariff provisions for large load customers, including data centers that may become Computational Load Entities under the NERC framework, are developing novel tariff language for which there is limited precedent. AI can be used to survey existing tariff provisions at other utilities, identify common structures, and draft initial language for attorney and rate-design expert review. The drafting is a starting point for expert work, not a final product.

Testimony Drafting: A Worked Example

Consider a utility preparing direct testimony from its revenue requirement witness for a general rate case before a state public utilities commission. The witness oversees the cost-of-service study and will testify to the utility's overall revenue requirement, the allocation of costs between customer classes, and the proposed rate design. The testimony is expected to run 80 to 120 pages.

The rate analyst uses an AI drafting tool to produce the initial draft. They provide the AI with: the cost-of-service study output tables (as structured data), the prior rate case testimony from the same witness (for voice and structure continuity), the commission's procedural schedule, and a prompt specifying the witness's name, title, topic scope, and a specific instruction: "For every numerical value in the testimony, insert [SOURCE: cite workpaper table and line number]. Do not insert any numerical value without this citation marker."

The AI produces an 85-page testimony draft in approximately 20 minutes. The draft has consistent structure, appropriate regulatory language, and the witness's established tone. And every number in the document is followed by a [SOURCE] marker.

The rate analyst then routes the draft to the cost-of-service study team. Their job: replace every [SOURCE] marker with the actual workpaper reference, confirm the number matches the workpaper, and flag any number that does not match for correction. They find 12 discrepancies. Seven are rounding differences (the AI used unrounded figures from training-data-pattern estimates rather than the rounded figures in the workpaper). Four are actual errors where the AI interpolated a number from the cost-of-service output in a way that misread the table structure. One is a transcription error from the prior testimony that existed in the source material.

The analyst corrects all 12 discrepancies. The witness reviews the full draft for voice accuracy and adjusts two explanatory paragraphs that do not capture the nuance of the commission's specific cost-allocation policy history. The attorney reviews for legal sufficiency. The testimony is filed on time, with every number sourced to the workpaper of record.

This is the model that works. Without the [SOURCE] marker discipline, a rate analyst under deadline pressure might not catch the 12 discrepancies. Without the four numerical errors, the testimony would have stated incorrect cost allocations under oath. The AI saved 15 hours of initial drafting time; the discipline saved the utility from a potentially significant evidentiary error.

Building a Repeatable Rate-Case AI Workflow

The utilities that get the most value from AI in rate-case work treat it as a system with defined roles and handoffs, not as an individual productivity shortcut. The system looks like this:

Document library and version control: All rate-case source materials (workpapers, cost studies, prior testimony, tariff provisions) are organized in a central document library before AI drafting begins. The AI is only permitted to reference materials in this library. This prevents the model from drawing on external sources that may contain competitor rate-case data or outdated regulatory language.

Role-differentiated access to AI output: AI drafts are marked as "draft pending verification" and are not accessible to external parties or filing systems until the verification workflow is complete. The rate analyst has access to the draft for number verification. The attorney has access for legal review. The witness has access for voice and accuracy review. Nobody files without all three reviews being documented as complete.

Standardized prompt templates for common filing types: Direct testimony, data-request responses, rebuttal summaries, and comment filings each have a prompt template that includes the number-flagging instruction, the citation format required by the commission's procedural rules, and the instruction to use only the source documents provided in the current session. Templates are updated when the commission changes its procedural rules or when a new rate case cycle begins.

Post-filing review: After each filing cycle, the rate team conducts a brief review of the AI drafting workflow: how many numerical discrepancies were caught, what types of errors appeared most frequently, and whether any near-misses reached the attorney review stage without being caught at the analyst stage. This feedback loop improves the prompt templates and the analyst training over time.

The FERC large-load rulemaking and the emergence of large compute load customers as a significant rate-case topic mean that utilities will be filing more novel, complex testimony over the next several years than at any previous point. Novel filings are exactly where AI's confidence in pattern-based drafting is most dangerous, because there are no prior precedent patterns to fall back on. The discipline of source verification is more important for novel topics than for established ones, not less.

Key Takeaways

• Rate-case testimony and data-request responses are sworn or penalty-of-perjury documents. Every number in an AI-produced draft must be flagged for source verification before filing. "The AI produced it" is not a defense; it is an aggravating factor in a false-statement allegation.
• AI genuinely accelerates testimony summary drafting, data-request response structuring, and pre-filing consistency checking. The productivity gains are real, but they are only safe when the number-flagging discipline is implemented without exception.
• The [SOURCE] marker protocol, where every numerical value in an AI draft carries a required workpaper citation that a human analyst must resolve, is the operational implementation of the number-flagging rule. It converts the drafting workflow into a verification workflow, not an approval-without-review workflow.
• The FERC large-load rulemaking (FERC committed in April 2026 to issue the rule by end of June 2026) creates new testimony and comment-filing demands. AI can accelerate comment drafting and tariff language research, but any citation to FERC orders must be verified against the actual final order text, not the AI's potentially outdated summary.
• A standardized, version-controlled prompt template library for common rate-case filing types (direct testimony, data requests, rebuttal summaries, comment filings) makes AI drafting systematic rather than ad-hoc and ensures the verification discipline is consistently applied across filing teams.
• The NERC Computational Load Entity framework will drive new tariff and rate-case work related to large data center customers. AI can help utilities survey tariff precedents and draft initial language, but novel tariff provisions require attorney review and commission-practice expertise that AI cannot substitute.
• A post-filing review of AI drafting errors, conducted after each rate-case cycle, creates a feedback loop that improves prompt templates and analyst training. Utilities that treat rate-case AI as a system rather than a shortcut will produce better filings with fewer errors over time.