A senior interconnection engineer sits down to write the methodology section for the twenty-third power flow study her team has completed this year. The physics are different from the previous twenty-two. The engineering judgment is different. But the words she types in the first four pages are nearly identical to the words she typed last month, and the month before that: the same citations to the same NERC reliability standards, the same description of the N-1 contingency analysis framework, the same boilerplate explanation of how the power flow model was constructed. If you asked her what percentage of a typical interconnection study report requires genuine engineering judgment versus careful assembly of standard language, she would tell you: maybe 20 percent needs judgment. The rest is architecture you have built a hundred times. AI-assisted study narrative drafting is built for that 80 percent.

Anatomy of an Interconnection Study Report: Finding the 80 Percent

To understand where AI-assisted drafting adds value, you need to understand what a typical interconnection study report actually contains. A large generator interconnection study for a project in a major ISO typically runs between 50 and 200 pages, depending on the project's complexity and the number of contingencies studied. The sections are predictable, because the applicable tariff specifies what the report must include and the NERC reliability standards specify what the analysis must cover.

A representative report structure includes: an executive summary describing the project and the study results; a project description section identifying the proposed generator, its technology type, nameplate capacity, point of interconnection, and proposed commercial operation date; a study scope section citing the applicable tariff provisions and reliability standards that govern the analysis; a methodology section describing how the power flow model was constructed, which contingency scenarios were studied, and what reliability criteria were applied; results sections presenting the findings for each study type (power flow, short-circuit, stability) with tabulated violations and proposed mitigations; a network upgrade section specifying the required improvements and their estimated costs; and an appendix with model diagrams and data tables.

Of these sections, the executive summary narrative, the project description, and the tariff citations in the study scope are almost entirely formulaic: the project's own application data fills the variable fields, and the surrounding prose follows a template that changes only when the tariff changes. The methodology section requires more tailoring but is largely standardized at the ISO level: the same analytical methods are used across all projects in the same queue, and the description of those methods is consistent from report to report. The results presentation requires accurately reporting the model outputs, which is a formatting and transcription task, not a drafting task. What requires genuine engineering judgment is the interpretation of the results: recognizing a voltage violation that the model flagged but that the engineer knows is a model artifact rather than a real-world concern, determining the appropriate mitigation for a short-circuit duty exceedance, assessing whether a proposed network upgrade is technically adequate and cost-reasonable.

The 80/20 framing is a practical heuristic, not a precise measurement. In some studies, the judgment-intensive content is closer to 30 percent; in simpler studies it may be 10 percent. But the directional point holds: a significant fraction of every interconnection study report is assembly work that follows established patterns, and AI-assisted drafting tools can handle that assembly work faster and more consistently than manual typing.

What AI Drafting Produces and What It Does Not

It is worth being explicit about what AI-assisted study narrative drafting actually generates and what kinds of content it should not be trusted to produce without careful engineering review.

AI drafting tools perform well at: populating template sections with project-specific variables extracted from the application and study inputs; assembling the citations to applicable tariff provisions and NERC reliability standards from a curated source document; generating the boilerplate narrative that contextualizes the study methodology in terms consistent with the ISO's standard language; and formatting results tables from structured model outputs. These tasks are well-suited to AI because they are largely pattern-following operations: the AI is filling slots in known structures, drawing on a defined set of source documents, not making analytical inferences.

AI drafting tools are less reliable at: interpreting model results and drawing engineering conclusions (the AI does not understand power systems; it is pattern-matching against text it has seen); generating technically accurate descriptions of novel mitigation solutions that have not appeared in the AI's training or reference documents; or correctly reporting numerical results when the model outputs are in non-standard formats or when unit conversions are required. These are the categories where engineering review is not optional but essential.

The AI draft is a structural scaffold, not a finished report. The engineer reads it with the same critical eye they would apply to work from a junior engineer on their first solo study: correct the structure if it is wrong, verify every number against the source, and own everything that goes out under the company's signature.

The distinction matters practically because it defines the review workflow. When AI generates a citations block that lists the applicable NERC standards and tariff provisions, the engineer's review task is: are these the correct and current standards for this study, and are they cited accurately? This is a targetable check that takes minutes. When AI generates a narrative about what the power flow results mean for the project's required network upgrades, the review task is more demanding: the engineer must read the narrative against the actual power flow outputs, verify that the conclusions are technically supported, and rewrite any sections where the AI's narrative does not accurately reflect the engineering analysis. The second review task cannot be shortcut; the first review task can be made efficient.

Building the Source Documents: Grounding the Draft in Your Materials

The quality of AI-assisted study drafting is almost entirely determined by the quality and organization of the source documents you give the AI to work from. An AI tool operating from well-organized, current source materials will produce a draft that is mostly correct and needs targeted editing. An AI tool operating from poorly organized materials, outdated templates, or inconsistent previous reports will produce a draft that requires extensive revision and may be more work to fix than starting from scratch.

The source document package for interconnection study drafting typically includes: the project's interconnection application (for project description variables), the ISO's current standard methodology document (for the methodology section template), the applicable tariff provisions cited in the study (for the study scope section), the current NERC reliability standards applicable to the study type (for reliability criteria citations), the model run outputs from the power flow, short-circuit, and stability simulations (for the results sections), the network upgrade specification from the engineering analysis (for the upgrade section), and previous reports for similar project types in the same ISO (as style and structure references).

Each of these documents must be current. Using an outdated tariff provision citation in an interconnection study report is not a minor formatting error; it is a potential tariff compliance issue if the cited provision has been superseded. The AI will use whatever source documents you provide, and if those documents are outdated, the draft will cite outdated provisions. Maintaining a current library of source documents is an operational discipline that the AI tool cannot substitute for: the tool is only as current as the materials it is given.

Prompt Structure for Study Narrative Sections

When using a generative AI tool to draft specific study report sections, the prompt structure significantly affects the quality of the output. A prompt that provides the AI with the project's specific data, the section's template structure, the relevant source documents, and a clear instruction about what the section must accomplish will produce a more usable draft than a prompt that asks the AI to "write a methodology section for an interconnection study."

A well-structured prompt for the project description section provides: the project name and identifier, the developer name, the technology type and nameplate capacity from the application, the proposed point of interconnection, the proposed COD, and an instruction to follow the ISO's standard project description template (provided in the source documents). The AI's output for this prompt should be a complete project description section that needs only a factual verification pass: are all the project data fields correct, does the technology description match the application exactly, is the POI identified consistently with the network model.

A well-structured prompt for the reliability criteria section provides: the applicable NERC reliability standards by designation (for example, FAC-002, TPL-001 through TPL-004), the study type being described (for example, transmission planning analysis for a 200 MW generating facility), and an instruction to cite the standards accurately with their current effective dates and to describe the specific reliability criteria applied in this study. The review task for this section is to verify each standard designation and effective date against the current NERC standards library, confirm that the criteria description accurately reflects the standards as currently written, and confirm that the study actually applied those criteria as described.

Worked Example: Drafting a Thermal Violation Narrative

Walk through a specific case to see both what AI drafting does well and where the engineer's judgment is irreplaceable.

The study team has completed a power flow analysis for a 350 MW wind project. The model outputs show a thermal overload condition on a 138 kV line segment during the worst-case N-1 contingency: the loss of a parallel 138 kV line puts 112 percent of normal-peak-emergency rating on the remaining segment during the project's full-output export scenario. The engineering team has determined that the required mitigation is a reconductoring of 4.3 miles of the affected segment with a higher-rated conductor, which the cost estimating team has priced at $3.8 million.

The prompt to the AI drafting tool for the thermal violation narrative section provides: the power flow results summary (line name, rating, contingency name, loading percentage, scenario description), the mitigation determination (reconductoring description, mileage, conductor type, cost estimate), and the ISO's standard narrative template for thermal violation findings.

The AI produces a draft that reads substantially as follows: "Power flow analysis identified a thermal overload condition on the [Line Name] 138 kV facility during the [Contingency Name] contingency under the project's normal operating scenario. The post-contingency loading of 112 percent of the normal-peak-emergency rating exceeds the NERC TPL-001 planning event criteria for P1 contingencies. The required network upgrade to address this violation is reconductoring of 4.3 miles of [Line Name] with [Conductor Type] conductor, at an estimated cost of $3.8 million. This upgrade must be completed prior to the project's commercial operation date."

The engineer reviews this draft against the actual model outputs and makes three specific checks: does the 112 percent loading figure match the model output table exactly; is the NERC TPL-001 P1 classification correct for this type of contingency and project size; and is the "prior to commercial operation" completion requirement consistent with the tariff's network upgrade timing provisions. Two of the three check out immediately. The NERC classification requires a brief review of the current TPL-001 standard to confirm that this contingency scenario falls under P1 rather than a different planning event category. After confirming, the engineer approves the section with a single word change (correcting a minor tense issue in the last sentence) and moves to the next section.

Total engineer time for that section: approximately eight minutes, versus the fifteen to twenty minutes it would have taken to draft the section from scratch. The efficiency gain is meaningful but modest for a single section. Across a 150-page report with forty such sections, the aggregate time saving becomes substantial.

Now watch the failure case. The same prompt, applied to a different project, produces a thermal violation narrative that states the loading is "within the normal-peak-emergency rating" when the model output actually shows a 103 percent overload, a genuine violation. The AI appears to have matched a narrative pattern from a previous report where the result was a passing condition rather than a violation. The engineer catches this in review because they have the model output table open and are comparing it line by line. If the engineer had not performed this specific review step, the report would have characterized a genuine violation as a non-violation, a potentially serious technical and tariff compliance error.

This failure case illustrates the non-negotiable review discipline: every numerical result in the AI draft must be verified against the source model output, not assumed correct because the surrounding prose reads well. The AI is very good at producing text that sounds authoritative and consistent. It is not good at ensuring that the numbers in that text are accurate. That verification is the engineer's job.

Version Control and Traceability in AI-Assisted Reports

A practical operational question for teams using AI-assisted drafting is how to maintain version control and traceability in the study report documents: how do you know which parts of a report were AI-generated, which were human-drafted, and which represent the engineer's verified conclusions?

The most common approach in production environments is to treat the AI draft as a working document in the same version-controlled system used for other study documents. The AI generates the initial draft, which is tagged as an AI-generated draft in the document management system. The engineer's edits, corrections, and verified content are tracked in revision history. The final signed report reflects the engineer's reviewed and approved content, regardless of whether sections originated in an AI draft or were written from scratch.

Traceability to source documents matters particularly for citations and numerical results. The study management system should record which model run produced the results reported in the study (by model version, scenario name, and run timestamp), so that if a question arises later about where a specific loading percentage came from, the audit trail leads back to the specific simulation output. AI-generated drafts that include result values should include a source citation for each value (for example, "per power flow run [ID], contingency [Name], scenario [Description]") so that the reviewing engineer can verify each value directly against its source.

This source citation discipline also protects against a subtle but important failure mode: the AI that synthesizes results across multiple runs and produces a composite narrative that is internally consistent but does not accurately represent any single run's results. Power flow studies often involve multiple scenarios (peak, light, off-peak, future case) and multiple contingency categories. An AI draft that blends results across scenarios without flagging which scenario produced which result can produce a narrative that is technically incorrect even though every individual number it cited was taken from a real model output. Requiring the AI to cite sources for every result forces the output to be traceable in a way that a synthesized summary narrative is not.

Key Takeaways

• A typical interconnection study report is roughly 80 percent assembly work (methodology boilerplate, tariff citations, results table formatting, standard narrative templates) and roughly 20 percent engineering judgment (interpreting model results, specifying mitigations, assessing technical adequacy). AI-assisted drafting addresses the 80 percent, freeing engineering time for the 20 percent.
• AI drafting tools produce reliable output for pattern-following tasks: populating templates with project-specific variables, assembling tariff and standards citations from curated source documents, and formatting results from structured model outputs. They are not reliable for interpreting model results or generating technically accurate conclusions about novel situations.
• The quality of AI drafting output is almost entirely determined by the quality and currency of the source documents. Outdated tariff provisions or NERC standards in the source library will produce citations that are wrong and potentially a compliance issue. Maintaining a current source document library is an operational discipline that AI cannot substitute for.
• Prompt structure matters. A prompt that provides project data, the section template, and relevant source documents will produce a more usable draft than a generic prompt. The more specific and structured the input, the more targeted the review task that follows.
• Every numerical result in an AI draft must be verified against the actual model output, not assumed correct because the surrounding prose reads well. The worked example failure case illustrates exactly why: the AI matched a narrative pattern from a passing result and applied it to a failing result, producing a technically incorrect characterization that only review against source outputs would catch.
• Version control and source citation in AI-assisted reports are not administrative formalities. They are the technical controls that make an AI-assisted report auditable: knowing which model run produced which result, which template version the methodology section draws from, and which engineer verified each section is the information that makes a study report defensible in a tariff dispute or regulatory review.
• The engineer who signs off on an interconnection study report authored with AI assistance is fully accountable for its accuracy and tariff compliance. The AI draft reduces drafting labor; it does not reduce accountability. Every section that goes out under the company's signature has been reviewed and approved by a responsible engineer, regardless of how it was drafted.