The interconnection queue has become the single largest bottleneck to grid modernization: more than 2,060 gigawatts of generation and storage projects were waiting for study results at the end of 2025, the median time from request submission to commercial operation date has doubled to over four years, and the majority of applicants withdraw before they ever reach an executed agreement. The study process itself is not broken because engineers are slow or careless; it is broken because the same painstaking boilerplate work gets repeated thousands of times a year, absorbing the finite hours of people whose judgment is genuinely irreplaceable. AI does not solve that problem by replacing engineers. It solves it by clearing the throughput-limiting boilerplate so engineers can spend their time on the 20 percent of work that actually requires them.

The Queue Crisis in Numbers

To understand why AI matters so acutely for interconnection, you have to hold the scale of the crisis in mind. The 2,060-plus gigawatts sitting in queues at the end of 2025 represents roughly twice the installed generating capacity of the United States. These are not all real projects: analysts estimate that a substantial fraction of queue capacity consists of speculative filings, placeholder positions, and projects that face insurmountable upgrade costs. But the study process cannot distinguish the viable from the speculative without going through the same study steps for each one, because FERC Order 2023 and the transmission tariffs underneath it require procedural consistency.

The four-year median request-to-COD timeline breaks down roughly as follows. An applicant submits an interconnection request, often with incomplete or ambiguous documentation. The transmission provider spends weeks in an intake review asking for corrections. Once the request clears intake, it joins a queue position and waits for a cluster study to begin, often for a year or more while earlier clusters complete. The cluster study itself takes six to eighteen months. The applicant receives the study results, negotiates upgrade costs and timelines, and then either signs an interconnection agreement or withdraws. At each step there is documentation to generate, review, and version. The boilerplate alone, the standard language about study assumptions, N-1 contingency definitions, facility cost allocations, and agreement terms, runs to dozens of pages per project. Multiply that by thousands of active studies and the arithmetic of the throughput problem becomes obvious.

The consequence is not just delay. Delay has a market cost. Every month a solar or storage project sits in queue is a month it cannot contribute to resource adequacy, cannot reduce curtailment from other projects, and cannot earn a return on capital. Utilities that are transmission providers feel the pressure from both sides: developers demand faster studies, and regulators increasingly expect faster interconnection timelines as a precondition for clean-energy deployment targets.

Where AI Fits in the Workflow

The full interconnection workflow runs from intake through completeness verification, study drafting, engineering review, and finally to the executed agreement. Each stage has a different ratio of judgment-intensive work to procedural, pattern-matching work. AI is effective precisely where that ratio tilts heavily toward the procedural.

Stage 1: Intake and Completeness

When a developer submits an interconnection request, the transmission provider must verify that the application contains every required element: point of interconnection, project type, MW capacity, technology, site control documentation, application fee evidence, and a set of project-specific technical parameters. Missing any one of these triggers a deficiency notice, which starts a clock on the developer and adds days or weeks to the process.

An AI-assisted completeness checker reads the submitted package against a checklist drawn from the applicable tariff and interconnection procedures. It flags missing fields, identifies ambiguous entries (a developer listing "battery storage" without specifying charge and discharge capacity, for example), and drafts the deficiency notice in the standard format. What once took a study coordinator two to four hours per application now takes minutes, with the coordinator's role shifting to reviewing and approving the AI-generated deficiency notice rather than composing it from scratch. This is not a trivial gain: a busy transmission provider processing hundreds of applications per year can recover thousands of staff-hours annually from intake automation alone.

The critical guard here is that the checklist must be grounded in the actual tariff, not in the AI's general knowledge of interconnection procedures. Different RTOs and utilities have different requirements, and the AI system must be anchored to the governing document for each jurisdiction. If it isn't, the completeness check will be systematically wrong in ways that do not announce themselves.

Stage 2: Study Draft Generation

Once a project has cleared intake and been assigned a queue position, the study process begins. Interconnection studies, whether Phase 1 (feasibility), Phase 2 (system impact), or Phase 3 (facilities), share a substantial common structure: a description of the project and study assumptions, a presentation of the power flow model used, a description of the contingencies evaluated, a table of constraint violations or upgrade needs, and a section explaining how costs are allocated.

All of that structure is boilerplate in the sense that it follows a template, references standard definitions, and draws from inputs that are already in structured data systems. An AI drafting tool can assemble the standard sections from those inputs, pull the correct definitions for N-1 contingencies and facility-cost allocation from the governing documents, and produce a readable first draft in a fraction of the time a study engineer would spend on the same task. The engineer then reviews the draft, checks the numbers against the power flow results, adds the professional judgment sections that require interpretation, and signs off.

This division of labor, AI drafts the frame, engineers fill the judgment, is where the productivity gains are most substantial. In programs where this approach has been trialed, study report drafting time has been cut significantly, though published figures vary and every number should be treated as a benchmark to verify against your own program's context, not a guarantee.

Stage 3: Engineering Review and Agreement

The engineering review stage is where the human-AI boundary is most critical. The power flow results, the contingency analysis, the thermal and voltage violation assessments: these are not AI outputs. They come from validated power flow software run by engineers who understand the grid model and can defend the study assumptions. AI's role at this stage is narrower but still valuable: it can check that the narrative in the study report is consistent with the numerical results, flag discrepancies between the draft text and the tables, and draft the standard cover letter and transmittal documentation.

Once the developer accepts the study results and moves toward an interconnection agreement, AI can assist with drafting the standard agreement sections, flagging where the template requires project-specific values to be inserted, and version-tracking the document through negotiation. The executed agreement is a legal document, and every clause must be reviewed by counsel. AI accelerates the assembly of the starting draft; it does not replace the review.

The Throughput Math

It is worth thinking carefully about where the time savings actually accumulate. Consider a hypothetical program processing 500 interconnection requests per year. Each request requires, on average, two hours of intake review, eight hours of study drafting work per study phase (with three study phases), and four hours of agreement assembly. That is roughly 7,500 staff-hours per year in purely procedural work, before any engineering judgment is applied.

If AI tools cut that procedural work by half, the program recovers 3,750 staff-hours annually. At a fully-loaded cost of $150 per hour (a conservative figure for experienced interconnection engineering staff), that is over $500,000 in recovered capacity per year, capacity that can either be redeployed to handle more studies or allow existing staff to give more careful attention to the technically complex cases.

The second-order effect is often more valuable than the first-order savings. When engineers are no longer buried in boilerplate, they can do the genuinely hard work better: catch edge cases in the power flow results, engage more carefully with developers on technically unusual projects, and write the judgment sections of study reports with more depth and precision. The queue does not just move faster; the studies that emerge from it are better.

Failure Modes and Guard Rails

The worst outcomes in AI-assisted interconnection work do not come from obvious errors. They come from confidently wrong outputs that pass through review without triggering skepticism. Consider the scenario where an AI drafting tool, grounded in a slightly outdated version of the tariff, generates a completeness checklist that omits a new requirement added in a recent FERC filing. Every application that passes that checklist will have the same deficiency, and the transmission provider will not discover this until a developer disputes a deficiency notice or, worse, until an executed agreement is challenged.

This is why grounding discipline is non-negotiable. The AI system must be anchored to dated, version-controlled copies of the applicable tariff and procedures, and those source documents must be updated whenever the governing rules change. A checklist that is accurate today and wrong in six months is worse than no checklist at all, because it generates false confidence.

A second failure mode involves the study narrative. An AI-generated draft that uses slightly imprecise language about contingency definitions, upgrade cost allocation methodologies, or interconnection agreement terms can be technically wrong in ways that matter legally and financially. Engineers reviewing AI-generated drafts must have enough background in the governing documents to catch these errors, which means the review cannot be delegated to junior staff who have less experience with the tariff.

The model drafts the frame. The engineer owns the content. These are not interchangeable, and the document that goes to the developer must reflect the engineer's judgment, not the AI's first pass.

A third failure mode is version confusion: if AI-generated drafts are stored alongside manually-created drafts without clear provenance, a reviewer may not know which document was AI-assisted and which was fully manually produced. Version control and metadata tagging are not optional; they are how you maintain the audit trail that a FERC compliance review or a developer dispute will require.

Designing a Defensible Workflow

A defensible AI-assisted interconnection workflow has five structural properties. First, it distinguishes clearly between AI-drafted sections and engineer-reviewed sections. The final study report must show who reviewed each section, when, and using what verification against the underlying power flow results and governing documents.

Second, the AI system's source documents are version-controlled and auditable. If a developer later disputes a study result or an agreement term, the transmission provider must be able to show exactly which version of the tariff the completeness checker or drafting tool was grounded in at the time the study was produced.

Third, the workflow includes a structured human review checkpoint at every document boundary: intake checklist to deficiency notice, study inputs to study draft, study draft to transmittal, agreement draft to executed agreement. Each checkpoint is logged, not just completed. The log is the evidence.

Fourth, the AI tools are never given write access to the official study record or the executed agreement. They produce drafts. Engineers approve, modify, and file the official documents. The distinction between draft and official document is enforced by the system, not just by policy.

Fifth, the program tracks whether AI assistance is changing the quality of outcomes over time. Metrics like deficiency notice accuracy, developer dispute rates, and study result challenge rates should be monitored both before and after AI tools are introduced. If the dispute rate goes up after AI assistance is deployed, something is wrong, and the program needs to investigate before assuming the tools are working as intended.

Worked Example: A Solar Intake Done Right

Imagine a 200 MW solar-plus-storage project submitted to a regional transmission organization's interconnection portal. The application arrives as a PDF package: a completed application form, a site control letter, an environmental assessment, and a single-line diagram. The study coordinator uploads it to the AI-assisted intake system.

The system parses the application against the applicable Large Generator Interconnection Procedures. It identifies five issues: the application form lists the AC output capacity but not the inverter loading ratio; the site control letter is dated eighteen months ago and may not meet the currency requirement; the single-line diagram does not show the point of interconnection clearly; the storage component's round-trip efficiency assumption is absent; and the application fee confirmation number is missing from the package. The system drafts a deficiency notice in the standard format, listing each deficiency with the specific tariff section requiring the missing information.

The study coordinator reviews the draft. Four of the five deficiencies are correct. The fifth, the one about inverter loading ratio, reflects a requirement that was modified in a tariff amendment six weeks ago, and the AI system had not been updated with the amendment. The coordinator catches this, removes the erroneous deficiency, and sends the corrected notice. The developer responds in ten days with a complete package, and the application clears intake in three weeks rather than the six to eight weeks the same process took before AI assistance was introduced.

The coordinator's intervention on the fifth deficiency is not a system failure; it is the system working correctly. The coordinator caught an error that the AI made because the source documents were stale. The program now has a documented reason to update the source documents immediately when tariff amendments are filed. That feedback loop, AI drafts, human catches errors, program improves, is what a well-designed AI-assisted workflow produces over time.

Key Takeaways

• The 2,060-plus GW interconnection queue and four-year median study time exist not because engineers are slow but because the same boilerplate work is repeated thousands of times; AI's highest-leverage role is clearing that boilerplate so engineers can spend time on judgment-intensive work.
• The interconnection workflow has five stages (intake, completeness verification, study drafting, engineering review, and agreement execution), and the AI-to-human ratio of work is different at each stage; intake and study drafting are the highest-leverage targets.
• Grounding discipline is the most important technical requirement: every AI completeness checker or drafting tool must be anchored to version-controlled, dated copies of the applicable tariff and procedures, and those documents must be updated when governing rules change.
• The most dangerous failure mode is not an obvious error but a confidently wrong output from a stale source document, catching these errors requires reviewers with enough tariff knowledge to recognize when the AI's output does not match the current governing document.
• A defensible AI-assisted workflow distinguishes AI-drafted from engineer-reviewed sections, logs every review checkpoint, keeps AI out of write access to official documents, and tracks outcome metrics before and after deployment.
• The second-order benefit of AI-assisted throughput, engineers doing better engineering on hard cases, is often more valuable than the first-order time savings, because the studies that come out of a well-designed program carry more defensible engineering judgment.
• Every ROI number cited for AI-assisted interconnection work should be treated as a benchmark to verify against your program's context, not a guaranteed outcome; the variables of queue volume, staffing, tariff complexity, and tool maturity differ significantly across programs.