The interconnection study report lands on the engineer's desk with a table showing that the proposed 400 MW solar project causes a 114 percent post-contingency thermal overload on a 345 kV line. The AI drafting tool has already generated the violation finding narrative and proposed a 14-mile reconductoring upgrade at an estimated cost of $7.2 million. Everything looks coherent. The prose reads well. The number is sourced from the correct model run. And the engineer, under pressure to issue the report before the end of the study period, is tempted to sign. This lesson is about why that temptation is the most dangerous moment in AI-assisted interconnection work, and what the professional discipline of verification looks like when the stakes are this high.

Why Interconnection Numbers Get Human Sign-Off Every Time

The interconnection study report is not a document that merely describes what happened in a model. It is a legal and commercial instrument that establishes rights and obligations. The violation findings in the report determine which network upgrades are required. The network upgrade specifications determine what those upgrades cost. The upgrade cost allocation determines how many millions of dollars the developer must pay as a condition of interconnecting. The timeline for upgrade completion determines the project's commercial operation date. And all of this flows from whether a number in a model run, mediated by an AI-generated narrative, accurately represents a real-world reliability concern that the applicable tariff and NERC standards require to be addressed.

The financial stakes of an incorrect violation finding are not abstract. A false positive (a model artifact flagged as a real violation) can add millions of dollars in upgrade costs to a project that does not actually cause a reliability problem. A false negative (a real violation that is not identified or is incorrectly characterized as within limits) can result in a project connecting to the grid with an inadequate network, creating a reliability risk that the transmission provider is obligated to prevent. Both error types have consequences that cannot be corrected by noting that the model produced an unexpected output.

The phrase "the model recommended it" is not a defense in a tariff dispute or a FERC complaint proceeding. When a developer challenges a violation finding, the transmission provider must demonstrate that the finding is technically supported by a correctly configured model run, interpreted according to the applicable reliability standards, and reviewed by a qualified engineer who exercised professional judgment about its accuracy. The AI's role in producing the draft narrative does not change any part of this accountability structure. The engineer who signs the report owns the content of that report.

The Verification Framework: What to Check and Why

Verification of interconnection engineering output is not a single review step. It is a structured sequence of checks, each targeting a specific type of potential error, performed by a qualified engineer before the study report is issued. Understanding the structure of the verification framework requires understanding what categories of error are possible in an AI-assisted study workflow.

Category One: Numerical Accuracy

The most consequential category of error is a mismatch between the number reported in the study and the number produced by the model. This can occur because the AI extracted the wrong value from the model output (for example, reading a pre-contingency loading instead of a post-contingency loading), because the model run used for the report was not the final validated run, or because the AI applied a unit conversion incorrectly. For each numerical result in the study, the verification step is to locate the specific model output table that produced the value, confirm the value against that table, and confirm that the table is from the correct run (identified by run ID) and the correct scenario.

This verification step cannot be shortcut by plausibility assessment. A 114 percent post-contingency loading sounds plausible for a 400 MW solar project at a 345 kV node. So does 109 percent, or 118 percent. The specific number matters because the violation finding and the upgrade specification derive from it. The engineer's job is not to assess whether the number is in a reasonable range; it is to confirm that the number is exactly correct.

Category Two: Citation Accuracy

Every regulatory citation in the study report must reflect the current, applicable version of the relevant standard or tariff provision. A citation to NERC FAC-002-3 when the applicable version is FAC-002-4 is not a minor formatting error; it may mean the study applied the wrong reliability criteria. A citation to a tariff section that has been superseded by a subsequent tariff filing means the study's legal framework may not be the framework that currently governs the interconnection. Citation verification requires the engineer to check each cited standard designation and version against the current NERC standards library and each tariff citation against the effective version of the tariff as of the study date.

Category Three: Engineering Judgment on Model Artifacts

Some violations that appear in model results are model artifacts rather than real-world reliability concerns: numerical instabilities in the power flow solution, violations that occur only in scenarios that are physically implausible, or results from a model that is known to have a configuration issue that the study team is aware of but has not yet corrected. An AI drafting tool will report these artifacts as violations because the model output says they are violations. The engineer's verification step is to assess each violation finding for whether it represents a real-world reliability concern or a model artifact, and to document that assessment in the study record.

This is the highest-judgment category of verification, and it is the one that most clearly cannot be delegated to an AI tool. Recognizing a model artifact requires knowledge of the specific network model, the model's known configuration issues, the physical plausibility of the violation scenario, and the applicable reliability standards' treatment of different scenario types. An experienced interconnection engineer has this knowledge. An AI tool pattern-matching from study report text does not.

Signing a study report is a professional act. It means: I have reviewed this content, I have verified the numbers against their sources, I have applied my engineering judgment to the findings, and I stand behind this document as an accurate and reliable representation of the study results. The AI helped produce it. The accountability is entirely mine.

Category Four: Completeness of Findings

An AI drafting tool works from the violation flags in the model output. If the model output contains a violation that is not prominently flagged but is visible in the results tables, the AI may not include it in the violation narrative. The engineer's verification step is to review the complete model results to confirm that every violation identified by the model is reflected in the study report, not just the violations that the AI's drafting tool surfaced. This completeness check is particularly important for short-circuit duty exceedances, which are often reported in separate tables from thermal overloads and may be overlooked in a review focused on the power flow results.

Documenting Verification: Building the Audit Trail

Verification is only useful if it is documented. A review performed without documentation leaves no record that it occurred, provides no basis for defending the study report if it is challenged, and creates no institutional memory that future engineers reviewing the file can draw on. The documentation standard for interconnection study verification must be sufficient to reconstruct, from the study file, exactly what was checked, who checked it, when, and what was found.

The practical implementation is a study report sign-off checklist that maps each verification step to the specific elements in the report it covers. The checklist includes, at minimum: confirmation that each result value was verified against a specific model run (run ID documented); confirmation that each regulatory citation was verified against the current applicable version (version date documented); confirmation that model artifact assessment was performed for each violation finding (assessment outcome documented); confirmation that the complete model results were reviewed for violations not surfaced by the AI's drafting tool; and the engineer's signature and date.

The sign-off checklist is not a cover sheet with a single signature. It is an itemized record of specific verification steps. In a tariff dispute or FERC complaint, the checklist is evidence that the review actually happened, not just that someone signed a document. Without the itemized checklist, the signature on the study report demonstrates only that the report was released, not that it was reviewed to any specific standard.

For studies involving large upgrade cost allocations (say, above $5 million or above a threshold defined in the transmission provider's quality program), a second-level review by a senior engineer who was not involved in the study production is a best practice. The second-level reviewer performs an independent verification of the highest-consequence findings: the violation that drives the largest upgrade, the cost estimate for that upgrade, and the tariff basis for the cost allocation. Two engineers independently verifying the same critical finding is the strongest available protection against an error becoming a signed, issued study report.

Failure Walkthrough: The Cost of a Missed Verification

A concrete failure case makes the consequences tangible. A 200 MW wind project is under interconnection study. The power flow analysis identifies a potential violation on a 138 kV line during an N-1 contingency. The model shows 107 percent post-contingency loading. The AI drafting tool generates a violation narrative and proposes a 3-mile line reconductoring at an estimated $2.1 million.

The study engineer, pressed for time, reviews the AI narrative and confirms that the violation finding is present. They do not verify the specific loading percentage against the model output. They do not check whether the contingency is correctly classified under NERC TPL-001 (it is actually an N-2 event, not an N-1, which changes the applicable reliability criterion). They sign the report.

The developer executes the interconnection agreement, budgeting the $2.1 million upgrade cost into their project finance model. Eighteen months later, during final engineering, the developer's consultant reviews the original study and discovers the N-2 misclassification. The N-2 reliability criterion for that system, under TPL-001, allows the post-contingency loading condition that was flagged as a violation; there is no actual violation, and no upgrade is required. The developer has been carrying $2.1 million in their project cost for 18 months, delaying investor commitment to the project, based on a violation finding that should not have been in the report.

The developer files a tariff dispute. The transmission provider must issue a corrected study report, amend the interconnection agreement, and defend their study process. The engineer who signed the original report must explain what review was performed and why the N-2 misclassification was not caught. The answer is that it was not caught because the NERC contingency classification was not verified. That is the specific verification step that was skipped, and it had a specific, traceable financial consequence for the developer and a significant tariff compliance consequence for the transmission provider.

This scenario is not hypothetical. Contingency classification errors, model run version errors, and upgrade cost estimate discrepancies appear in real interconnection dispute proceedings. They are the specific failure modes that a structured verification framework is designed to prevent.

Verification in the Age of AI-Scaled Study Production

As AI tools make it possible to produce study reports faster, there is a risk that the productivity gain is extracted at the expense of verification time. If a team that previously took three weeks to produce a study report now takes ten days, the natural pressure is to use the saved time to start the next study faster, not to perform more rigorous verification. This is the wrong trade.

The appropriate use of AI's productivity gain is to maintain or increase the time and rigor of the verification step, not to reduce it. If the assembly work that previously took two weeks now takes four days, those freed days should be allocated to a more thorough verification pass, including second-level review on high-cost studies, not just to starting the next study sooner. The verification step is what justifies the engineer's signature; it should not become the part of the workflow that gets compressed when the calendar tightens.

This requires explicit organizational commitment to a verification time standard. The study production workflow should allocate specific time for the verification step as a non-negotiable phase, not as something that happens in the margins of other work. In a team that has adopted AI-assisted drafting, the project management for study production should show a verification phase with a defined duration, regardless of how quickly the AI produces the initial draft.

There is also a more subtle risk: as AI-assisted drafting becomes routine, the reviewing engineer may gradually shift from active verification to passive reading. The first few times an engineer reviews an AI-drafted study report, they check every number carefully because the AI is new and unfamiliar. After reviewing a hundred AI-drafted sections that were all correct, the reviewing engineer may start to read the AI's prose for consistency and plausibility rather than verifying each claim against its source. This is the point at which the review loses its reliability protection. Maintaining active verification over time requires deliberate professional discipline: the review is source verification, not a reading comprehension exercise.

Key Takeaways

• Interconnection study reports are legal and commercial instruments that establish upgrade obligations, cost allocations, and commercial operation dates. Numerical errors in these reports have direct financial consequences, and the engineer who signs the report is accountable for every number in it, regardless of how it was produced.
• The phrase "the model recommended it" is not a defense. When a violation finding is challenged, the transmission provider must demonstrate technical support from a correctly configured model run, applied to the correct reliability standards, and reviewed by a qualified engineer. AI drafting does not change this accountability structure.
• Verification is a structured sequence of four check types: numerical accuracy (every result against its model source), citation accuracy (every standard and tariff citation against the current applicable version), engineering judgment on model artifacts (assessing whether each violation is a real-world concern or a model configuration issue), and completeness (confirming that all model violations are reflected in the report, not just those surfaced by the AI).
• Verification must be documented to be defensible. An itemized sign-off checklist that records what was checked, what source was referenced, and who performed the verification is the difference between a review that happened and a review that can be demonstrated. A single signature on the cover page does not substitute for an itemized verification record.
• The failure walkthrough illustrates the specific consequence of skipping NERC contingency classification verification: an N-2 event misclassified as N-1 produced a violation finding under the wrong reliability criterion, resulting in an unnecessary $2.1 million upgrade specification that the developer carried in their project cost for 18 months before a dispute corrected it.
• AI's productivity gain should not be extracted from the verification step. The time freed by AI-assisted drafting should be allocated to thorough verification, not to starting the next study faster. The verification step is what justifies the engineer's signature; it must be protected as the non-negotiable phase of the study production workflow.
• Maintaining active verification over time requires deliberate professional discipline. The risk that reviewing engineers gradually shift from source verification to plausibility reading is real, and it is greatest in teams that have used AI-assisted drafting long enough for the AI's correct outputs to feel routine. The review discipline must be maintained regardless of confidence in the AI tool's recent track record.