You finish this lesson today. Ninety days from now, you will have a shippable AI-assisted workflow running in your real job, a portfolio artifact you can show in a job interview or a department presentation, and the verified-AI-output discipline that separates an informed energy professional from someone who just has access to the same tools. This lesson is the plan to get there.

Why a Sequenced Plan Matters

The most common failure mode for energy professionals trying to develop AI skills is starting in the wrong place. They open an AI tool, ask it a grid question, get a confident-sounding answer, and either trust it without verification (dangerous) or distrust it without engagement (wasteful). Neither builds skill. The skill that matters in a reliability-critical environment is not the ability to use AI tools; it is the ability to evaluate their outputs and deploy them responsibly within a domain you already understand well.

That evaluation skill builds in sequence. You have to understand what the tool is doing before you can verify that it is doing it correctly. You have to verify correctly before you can use the output to support a real decision. You have to use outputs in real decisions before you can design a workflow that another professional can follow and trust. Each phase builds on the previous one, and skipping phases produces fragile results: a workflow that works until the first edge case, or a portfolio artifact that impresses in a presentation but cannot survive a skeptical engineering review.

This 90-day plan is sequenced exactly that way. Each phase ends in a concrete, shippable artifact: something you produce, that carries your name, that demonstrates the skill in a form a hiring manager or a department head can evaluate. The artifacts are not homework assignments; they are deliverables you would produce in the course of doing your job. The plan is designed so that following it also improves your current role, not just your career prospects.

The plan also bridges directly into L2 of this program. The last artifact of the 90 days is the starting context for the first applied skill in L2: prompting for utility work. When you enter L2, you will already have a real workflow, real AI-generated outputs you have evaluated, and real verification decisions you have made. L2 builds the technical depth on the foundation this plan creates.

Days 1 Through 30: Read AI Outputs Skeptically

The first phase has one objective: learn to read AI outputs as a trained energy professional, not as a general user. This means developing the habit of asking four questions every time you see an AI output in a grid context.

First: what data did the model use? A day-ahead load forecast produced by an ML model is only as good as its training data and its current input data. If the weather inputs are stale, if a large-load customer notification is missing from the feature set, if recent DER output data was not refreshed from the AMI feed, the forecast can be confidently wrong in ways that look like model error but are actually data pipeline failures. Asking about data provenance is not paranoia; it is the first verification step for any model output in a regulated, audited environment. The forecaster who cannot answer "when was this model last retrained and what data did it train on?" cannot defend the forecast in a rate case or explain an operational deviation to a reliability coordinator.

Second: does this output match what I would expect from physical and operational reality? If an AI-generated switching recommendation says to close a specific breaker when that breaker is currently tagged out for a maintenance outage, the recommendation is wrong regardless of how sophisticated the model is. If a restoration sequencing tool routes a crew to a segment that you know was restored three hours ago by a mutual-aid crew whose completion was not yet logged in the OMS, the recommendation is working from a stale picture. Physical-reality checking is the most basic form of domain-expert verification, and it catches many of the highest-consequence AI errors before they drive a field crew to the wrong location or send an operator to execute an unsafe switching sequence.

Third: is this output within the model's validated operating range? Every model has conditions under which its performance has been validated and conditions it has never seen. A load-forecasting model trained on historical data from 2015 through 2022 has never seen a 300 MW overnight step-load from a hyperscale data center, and it has never seen the net-load duck curve that emerges on a mild spring day with 800 MW of distributed solar on the feeder network. Asking whether the current situation resembles the model's training distribution is not a technical question; it is a domain judgment that any experienced forecaster can make by looking at historical load shapes and asking "have we ever seen this before?"

Fourth: what is the consequence of acting on this output if it is wrong? A wrong AI-generated report draft costs time to correct. A wrong AI-generated peak-day forecast costs capacity procurement dollars and potentially a NERC reliability notice. A wrong AI-generated switching recommendation in a real-time contingency has safety and reliability consequences that extend across the EMS footprint. The verification rigor should scale explicitly with the consequence of error. An energy professional who applies the same level of caution to every AI output, regardless of its operational stakes, will be too slow on low-stakes drafts and, more dangerously, will have normalized a verification habit that is insufficient for high-stakes decisions.

The Phase 1 artifact is a written evaluation of one AI output you encounter in your work during this month. It does not have to be dramatic. It can be an AI-generated day-ahead load forecast narrative, a regulatory filing draft produced by a generative AI tool, an asset-health risk score from a predictive maintenance system, or a chatbot answer to a question about a NERC standard. The evaluation answers the four questions above, in writing, in two to three pages. The discipline of writing the evaluation, rather than just thinking through it, is what builds the verification habit. This document is not for public consumption; it is your own development record, and it is the evidence base you draw on when a manager or a regulator asks you to describe your AI output verification methodology.

What to Read in Month One

Alongside the written evaluation, spend the first thirty days reading the ground-truth material that anchors the rest of this program. NERC's May 2026 Level 3 Alert on computational load growth is publicly available and gives you a direct view into the reliability organization's current concerns at the highest level of urgency. FERC's 2026 large-load interconnection rulemaking materials are available in FERC's docket management system and are worth reading in summary form for the regulatory context they establish for any utility with data-center customers in its territory. EPRI's workforce analysis publications, available to members, provide the data behind the 30-plus percent digital-role growth projection that establishes the career opportunity this program addresses. These are not casual reads; they are each one to two hours of substantive primary-source review. But they ground your AI literacy in the actual regulatory and workforce context of 2026, making the tools you use more legible and the outputs you evaluate more interpretable against the current professional standard.

Days 31 Through 60: Run One Real Workflow

The second phase moves from evaluation to application. The objective is to identify one task in your current role that you can partially complete with AI assistance, run it for a month, and produce a documented verification record. The emphasis on "partially" is intentional: you are not trying to automate a task. You are trying to use AI to accelerate the boilerplate component of a task so you can spend more time on the judgment component that requires your grid expertise.

Good candidates for the first AI-assisted workflow depend on your role. If you are a load forecaster, a good starting task is having an AI tool draft the narrative explanation of your day-ahead forecast, which you then edit and verify before it goes to the operations center. The AI drafts the temperature-load correlation explanation, the comparison to the analogous historical period (for example, noting that the current forecast resembles conditions from the same week in the prior year with a 2-degree temperature offset), and the caveats about uncertainty related to any pending large-load interconnections. You verify every number against the model's actual output, check that the cited historical comparison period is genuinely comparable in weather and load-composition terms, and rewrite any section where the AI's language understates the forecast uncertainty or overstates the model's precision relative to your actual holdout MAPE. The value is not that the AI writes the narrative instead of you; it is that it produces a structural first draft that you refine in 20 minutes rather than building from a blank page in 60 minutes.

If you are an interconnection engineer, a good starting task is using an AI document review tool to screen incoming interconnection applications for completeness before they enter your formal study queue. The AI checks whether all required exhibits are present, whether the stated point of interconnection is within the service territory boundaries in your GIS, whether the proposed capacity is consistent across the application form and the one-line diagram, and whether the required interconnection request form fields have been completed according to your tariff's specifications. You verify the tool's completeness determination on a sample of ten applications, document any false positives (applications the tool flagged as complete that were missing something) and false negatives (applications the tool cleared that actually had a deficiency), and set a threshold for which completeness issues require human review versus which the tool can clear automatically. The value is not that the AI screens instead of you; it is that it handles the rote checking on 200 applications so your engineering time concentrates on the 30 that have substantive technical issues.

If you work in NERC compliance, a good starting task is having an AI tool organize your evidence files for a specific reliability standard, such as FAC-002 or BAL-001, into a structured summary that maps each piece of evidence to the corresponding requirement element and sub-element. You verify that the mapping is accurate by cross-referencing the AI's requirement citations against the current effective version of the standard, flag any gaps between your evidence inventory and the full requirement structure, and use the summary as the basis for drafting your compliance narrative. The value is that the AI produces a navigable evidence map in 30 minutes that would have taken you several hours to build manually from a folder of documents, and you spend your time on the substantive gaps and the narrative rather than the organizational work.

The Phase 2 artifact is your verification log from the month: a record of each AI-assisted task, what the tool produced, what you checked, what you corrected, and why. Structure the log with a row for each task: date, tool used, task description, AI output summary, verification steps taken, corrections made, and a one-sentence quality assessment. The log should be honest about where the tool failed and where it helped. This log is the operational evidence of your verification discipline, and it is also a genuine contribution to your organization's understanding of how AI tools behave in your specific operational context. If you are building institutional knowledge as this chapter recommends, your verification log is a foundational piece of that knowledge base.

Managing the Failure Cases in Month Two

You will encounter AI failure modes during Phase 2. The most common in utility contexts are: confidently wrong numbers (the model states a specific MW figure, cost value, or MAPE percentage that is demonstrably incorrect), invented citations (the model references a NERC standard requirement number, a tariff attachment, or a commission order that does not exist or is misrepresented), and scope creep (the model addresses a broader question than you asked, introducing material about adjacent regulatory provisions or market rules that were not in your prompt and may not be accurate for your jurisdiction). Each of these is a verification catch, and each one should be logged. The failure cases are not discouraging; they are the evidence that your verification discipline is working. An AI tool that never produces errors you catch is a tool you are not checking rigorously enough, not a tool that is perfectly reliable.

Pace yourself through Phase 2. The workflow should take no more than thirty minutes per day of additional time above what you already spend on the underlying task. If it is taking longer than that, the workflow is too complex for a first application. Scale it back to a narrower task, build the discipline at that narrower level, and expand later. The 90-day plan is specifically designed to avoid the common failure mode of an ambitious pilot that becomes unsustainable by week three because the verification overhead exceeded the drafting savings.

Days 61 Through 90: Ship an Artifact

The third phase produces the shippable artifact that marks your completion of L1 and your readiness for L2. The artifact is a workflow documentation that describes the AI-assisted process you ran in Phase 2: what task it supports, what the AI tool does, what the human verification step covers, what the override criteria are, and what the output looks like. The document should be specific enough that a colleague in your department could follow it and reproduce your results. It should be written in the register of an operational procedure or a technical memo, not in the register of a case study or a presentation.

The workflow documentation has a specific structural form:

Title and scope: Name the task, the tool, and the operational context (e.g., "AI-Assisted Day-Ahead Forecast Narrative Review, [Utility Name] Load Forecasting, Using [Tool Category]").

Data inputs: What data the AI tool uses, where it comes from, and what freshness or completeness requirements apply before the tool runs.

AI task description: What specifically the tool does, in plain language. Not how the model works internally; what it produces as output and in what format.

Verification checklist: The specific checks the engineer applies to the AI output before it is used. This is the core of the document and should be specific to your domain: not "check for accuracy" but "verify that the cited historical comparison period had similar temperature profiles to the forecast period" and "confirm that any large-load customer with a notification in the past thirty days is reflected in the forecast adjustment."

Override criteria: The specific conditions under which the engineer overrides the AI output and the documentation required. "If the model's forecast deviates from the planner's estimate by more than X percent and the planner has identified a specific reason for the deviation (such as a large-load commissioning event not in the model's data), the planner documents the reason and submits the adjusted forecast as the official position."

Performance record: A summary of the verification log from Phase 2: how many times the AI output was used as-is, how many times it was modified, and what the most common modification categories were. This section turns your Phase 2 log into organizational evidence of the tool's reliability in your specific context.

Presenting the Artifact

The final week of Phase 3 is for presenting the workflow documentation to one other person: a manager, a colleague, or a mentor. The presentation does not have to be formal. It can be a thirty-minute walkthrough of the document with someone who will ask hard questions about the verification methodology. The goal is to test whether the documentation is clear and specific enough to survive scrutiny, which is also the test it would face in a rate case or a compliance audit.

If the presentation surfaces gaps or weaknesses in the documentation, revise it. A workflow document that has been questioned and improved is more valuable than one that has never been tested. The revised document is your Phase 3 artifact: a verified, stress-tested workflow procedure that demonstrates your AI literacy in a concrete, operational form.

This artifact is also your entry point into L2. The L2 curriculum builds on exactly this foundation: prompting techniques that improve the quality of the AI outputs your workflow depends on, verification methods that extend the checklist you developed in Phase 2, and structured output formats that make AI-generated content more directly usable in grid contexts. When you reach L2, you will not be starting from theory. You will be applying technical depth to a workflow you have already run and a verification discipline you have already built.

Maintaining Momentum Across the 90 Days

Three practices help maintain momentum through a 90-day skill-building plan in a demanding operational environment.

First, protect thirty minutes per week for deliberate AI practice. This is not the time you spend using AI tools in the course of your work; it is time specifically set aside to evaluate what you learned, update your verification log, and identify one improvement to your workflow. Thirty minutes per week is sustainable even in high-pressure periods, and the consistency compounds over twelve weeks into a genuine skill.

Second, find one colleague who is also working through this program or developing AI skills, and share your verification log entries with them monthly. Peer review of verification decisions, even informal, surfaces assumptions and blind spots that solo practice does not. It also creates a social accountability structure that makes the plan more likely to survive the first busy week.

Third, keep a running list of AI output failures you observe in your work or in industry news. Not as a collection of horror stories, but as a calibration resource. When a well-resourced organization's AI tool produces a confidently wrong output in a high-profile context, it is a reminder that the verification discipline you are building is not overcautious; it is the standard that the industry is converging toward. The list also gives you concrete examples for the job interview or the department presentation where you are asked to explain why AI outputs need human verification in a grid context.

The Bridge to L2

L2 of this program is titled "AI-Assisted Analyst/Operator" and its first chapter is "First AI Conversations for Energy Work." The skills L2 builds are the technical complement to the operational discipline you built in L1: how to prompt AI tools to get more reliable outputs, how to force citation discipline in AI-generated regulatory content, and how to recognize bad AI output before it gets into your verification step.

The specific L2 skill that your Phase 3 artifact positions you for is the prompting basics lesson. When you know what your AI tool produces in a real workflow and what its common failure modes are in your specific domain, you have the context you need to write effective prompts. A prompt that locks in your jurisdiction, units, standard citations, and the "show your sources" rule is more effective when you already know, from Phase 2 experience, which failure modes the prompt needs to prevent.

The 90-day plan is designed so that the energy you spend on it builds your current role's value simultaneously with your career options. If you are already a grid data analyst or a forecaster, the workflow artifact is a genuine operational contribution that improves your department's AI governance. If you are aiming for a new role, the artifact is a portfolio demonstration of exactly the skill the market cannot find enough of. Both outcomes come from the same work. That is intentional. The best AI skill-building for energy professionals happens in the context of real grid problems, not in a sandbox. Start where you are, use what you have access to, and verify everything.

Key Takeaways

• The 90-day plan builds AI skill in the sequence it must be learned: read skeptically first, then apply to a real workflow, then ship a documented artifact. Skipping phases produces fragile results that fail at the first edge case.
• Phase 1 (Days 1-30) develops the four-question verification habit: data provenance, physical-reality check, validated operating range, and consequence-scaled rigor. The artifact is a written evaluation of one real AI output.
• Phase 2 (Days 31-60) applies the habit to one real workflow: AI assists the boilerplate, the human professional verifies and retains accountability. The artifact is a monthly verification log. Failure cases in the log are evidence the discipline is working, not evidence to abandon the workflow.
• Phase 3 (Days 61-90) produces a shippable workflow documentation: title, data inputs, AI task description, verification checklist, override criteria, and performance record. This document is the operational artifact of L1 and the entry point into L2.
• The 30-minutes-per-week deliberate practice cadence, peer verification-log review, and failure-case catalog are the maintenance practices that make the plan sustainable in an operational environment.
• The Phase 3 artifact bridges directly into L2's first chapter, which builds prompting technique on the foundation of a real workflow and real verification decisions you will have already made.
• Building AI skills in the context of a real grid problem is more valuable than any abstract training. Start where you are, use tools you have access to, verify every output, and document your reasoning. That discipline, consistently applied, is what the market cannot find enough of.