A load forecaster at a mid-size investor-owned utility once told me that her team had "integrated AI into forecasting." When I asked how, she said the model produced a number every morning, and they used it. No step-by-step map of what the model touched, no explicit list of what a human verified, no documented handoff. That is not an AI-integrated workflow. That is an AI-dependent one, and the difference will matter the first time a 400 MW data-center interconnection lands in the service territory overnight and the model has never seen that pattern before.

Why Process Mapping Is the First Practitioner Skill

Before you automate anything, before you pick a vendor tool, before you write a prompt, you need a clear map of the work. Process mapping for AI integration is not a project-management formality. It is the document that tells you which steps are genuinely AI-ready today, which steps require a human decision and accountability signature, and where the two hand off to each other. Without that map, you cannot govern the workflow, you cannot audit it, and you cannot defend it to a NERC reviewer or a state commission.

The energy industry already knows how to document processes. NERC Operations and Planning standards depend on documented procedures. Transmission planning studies follow defined methodologies. Restoration switching orders follow a written sequence. The discipline is native. What is new is adding a column to that map that says: "Does an AI model produce, draft, or influence this step, and if so, what is the boundary of that influence?"

Three grid workflows appear repeatedly across utilities as candidates for AI integration: load forecasting, interconnection study processing, and outage restoration. Each is complex enough to illustrate the mapping method fully, and each has a different mix of AI-ready steps and human-only steps. This lesson walks through all three, builds the mapping discipline, and ends with a reusable framework you can apply to any process in your organization.

Decomposing a Workflow Into Its Atomic Steps

The first move in process mapping is decomposition: breaking a high-level workflow down into its smallest meaningful actions. "Produce the day-ahead load forecast" is not a step. It is a process. The steps inside it include: ingest the weather feed, ingest yesterday's metered load by zone, apply the model, review the output for anomalies, compare against the previous seven days, apply any known override (a large customer shutdown, a planned outage), document the rationale, and release the forecast to the EMS and to the day-ahead scheduling team.

That is nine distinct actions, and they are not all the same from an AI perspective. Ingesting a weather feed is data plumbing. Applying the model is where AI does its core pattern-matching work. Reviewing the output for anomalies is human judgment. Applying an override is a human decision that requires accountability. Documenting the rationale is a task where AI can assist drafting but a human must confirm. Releasing the forecast is a commitment with consequences in the real-time energy market.

The discipline is: for each step, ask three questions. First, what are the inputs and outputs? Second, does success at this step require judgment about conditions the model was not trained on, accountability for a reliability decision, or interpretation of a regulatory requirement? Third, if this step fails silently, what is the consequence?

Steps that fail silently with major consequences are not AI-only steps. They require a human checkpoint, a logged review, and a documented sign-off. This is not a limitation on AI capability; it is a recognition that in a regulated, safety-critical system, accountability cannot be delegated to a model.

The AI-Ready vs. Human-Only Classification

Once you have decomposed a workflow into atomic steps, classify each step. The classification is not permanent and not binary; it is a starting point for design. A step that is human-only today may become AI-assisted in two years when better training data exists. A step that is AI-ready today may need a human checkpoint added when a new load class enters the service territory.

What Makes a Step AI-Ready

A step is AI-ready when: the relevant data is available in structured, consistent form; the task is pattern-based or involves processing large volumes of similar inputs; errors at this step are detectable before they propagate; the outputs feed a downstream human review rather than going directly to a reliability-critical action; and the performance of the AI at this step can be monitored quantitatively (MAPE, precision, recall, latency).

In load forecasting, the step "apply the gradient-boosting or transformer-based model to produce a 24-hour load shape" is AI-ready by these criteria. The model has been trained on millions of historical weather-load pairs. Its output is a number that a human will review. Its accuracy can be tracked as MAPE over time. If the model drifts, a monitoring dashboard catches it before it drives a procurement decision.

In interconnection study processing, the step "check the incoming application for completeness against the standardized information requirements" is AI-ready. The checklist is well-defined in the tariff and the pro forma interconnection agreement. The AI reads the submitted documents and flags missing fields. A study engineer reviews the flag list before returning the application. The AI's error is bounded: it might miss a subtlety in a non-standard document, but a human reviews before any commitment is made.

What Makes a Step Human-Only

A step is human-only when: the decision carries NERC or FERC accountability; the inputs include conditions the model has not been trained on (a novel load type, a new regulatory interpretation, a first-of-its-kind contingency); the consequences of a silent failure are irreversible or safety-critical; or the output will be filed with a regulatory body or cited in a compliance record.

In load forecasting, the step "apply or decline a human override before releasing the day-ahead forecast" is human-only. The forecaster is making a judgment call that will influence how much energy the utility buys in the day-ahead market. A MAPE of 2% on a 5,000 MW peak means a 100 MW error. At $50/MWh real-time imbalance price, that is $5,000 per hour. The forecaster's name is on that decision, and "the model did it" is not an acceptable defense in an audit or a rate case.

In restoration, the step "authorize the switching sequence that re-energizes a 115 kV bus" is human-only. An AI can generate a proposed switching order and flag it against the switching procedure library. But the transmission operator signs the switching order. The N-1 contingency check is the operator's responsibility under FAC-001 and FAC-002 principles. The model advises; the operator decides and documents.

The AI-Assisted Middle Ground

Many steps fall in a third category: AI-assisted, meaning the AI produces a draft or a recommendation that a human reviews, modifies if needed, and confirms. This is the most common category in well-designed utility workflows, and it is where you get the highest leverage from AI without sacrificing accountability.

The key design rule for AI-assisted steps is: the human's review must be substantive, not ceremonial. If the human is rubber-stamping AI output without actually checking it, the step is functionally AI-only, regardless of what the workflow diagram says. Good workflow design makes the human review easy to do and hard to skip: the system presents the AI draft alongside the source data, highlights the fields the AI filled in, flags any values outside the expected range, and requires a typed confirmation that the reviewer has checked specific items.

Mapping the Three Canonical Workflows

Day-Ahead Load Forecasting

Let us walk through the day-ahead load forecasting workflow with a full step-by-step map. A typical utility runs this process by 10 AM each day for the following day's operating period.

Step 1: Ingest weather forecast data from the National Weather Service and any supplemental commercial feeds (classification: automated data plumbing, AI-ready). Step 2: Ingest prior-day metered load by zone from the EMS historian (automated, AI-ready). Step 3: Check both feeds for data quality flags (missing intervals, sensor outages) (AI-assisted: the model flags anomalies, a data engineer confirms before the run). Step 4: Execute the load forecasting model (AI-ready: this is the model's core function). Step 5: Compare model output against recent actuals and the seasonal baseline (AI-assisted: the model generates a deviation summary, the forecaster reviews it). Step 6: Check for known events not in the training data: a large industrial load coming back online, a planned transmission outage that alters the zone shape, a confirmed data-center energization (human-only: the forecaster applies judgment and, if an override is applied, documents the reason). Step 7: Release the forecast to the EMS and to the day-ahead scheduling desk (human-only: requires a sign-off). Step 8: Log the run: model version, input timestamps, any overrides, the releasing forecaster's ID (AI-assisted: the system auto-populates most fields, the forecaster confirms and submits).

The result is a map that is explicit about where the model touches the forecast and where a human owns the outcome. This map is also your audit trail template. Every field in Step 8's log corresponds to something a commission or a NERC auditor might ask about if the forecast is later scrutinized.

Interconnection Study Processing

The interconnection queue held over 2,060 GW of requested capacity at the end of 2025. Median request-to-completion times have stretched to over four years. The throughput problem is not primarily computational; it is administrative. Studies are delayed by incomplete applications, back-and-forth on data requests, and the time it takes a small study team to read and compare documents against tariff requirements.

AI integration addresses the administrative steps, not the engineering judgment steps. The map looks like this. Step 1: Receive interconnection application (automated intake into the queue management system, AI-ready). Step 2: Parse the application for required fields against the tariff checklist (AI-ready: the model reads the documents and populates a completeness matrix). Step 3: Review the completeness matrix and issue a deficiency notice if items are missing (AI-assisted: the engineer reviews the matrix, confirms or corrects it, issues the notice with their name on it). Step 4: Assign the application to the appropriate study cluster (AI-assisted: the model suggests a cluster assignment based on location and requested MW; the study engineer confirms). Step 5: Run the power flow cases (human-supervised computational step: the study engineer defines the contingencies and reviews the solver outputs). Step 6: Draft the study report boilerplate (AI-assisted: AI drafts sections 1-4 of the report from the application data and the tariff template; the engineer writes the engineering-judgment sections). Step 7: Review and sign the report (human-only: the responsible engineer signs; this is a tariff commitment). Step 8: File with the transmission provider and notify the applicant (automated with logged timestamp).

The efficiency gain is concentrated in steps 2, 3, and 6, where AI reduces the time a study engineer spends on structured document review and boilerplate drafting. The accountability steps, 5 and 7, remain fully human. The map makes this explicit so that if the process is ever audited, the utility can show exactly where AI was and was not in the decision chain.

Outage Restoration

Restoration is where the stakes are highest and where the human-only boundary must be the most sharply drawn. The scenario: a transmission contingency has caused an N-1 isolation of a 115/34.5 kV substation, and the distribution system below it is unpowered. The Energy Management System has alarmed. The control room operator has 90 seconds to assess and begin the restoration sequence.

Step 1: Detect and classify the outage (AI-ready: the EMS/ADMS automatic fault locating runs instantly; the AI classifies the outage type and affected zones). Step 2: Generate proposed restoration options (AI-assisted: the ADMS switching advisor generates 2-4 restoration paths, ranked by restoration time and load served; the operator reviews). Step 3: Check each proposed path against current switching restrictions, tagged-out equipment, and the N-1 reliability standard (human-only: the operator verifies each path against the switching log and the current system state; the model does not know which breakers a lineman has padlocked for maintenance unless the lockout is properly recorded in the OMS). Step 4: Select the restoration path and issue the switching order (human-only: the operator authorizes; this is a reliability accountability decision). Step 5: Execute and track each switching step (AI-assisted: the ADMS tracks step completion and re-runs power flow as the topology changes; the operator monitors and can halt the sequence). Step 6: Confirm restoration and clear the alarm (human-confirmed: the operator verifies restoration on the SCADA display before clearing). Step 7: Log the event: cause, sequence, times, operator ID, any deviations from the proposed path (AI-assisted: the system auto-generates the event log from SCADA timestamps; the operator reviews and signs it before filing).

The 90-second clock is not an argument for removing the human. It is an argument for making the AI's advisory output fast, clear, and easy to verify. A well-designed ADMS switching advisor shows the operator the proposed sequence in a visual one-page format with the estimated restoration time and the N-1 consequence of each step. The operator can approve in 15 seconds when the output is well-designed. The signature takes five seconds. The accountability is preserved.

Building Your Own Process Map: The Framework

You can apply this framework to any grid workflow. The output is a table with six columns: Step Number, Step Description, Inputs, Outputs, Classification (AI-ready / AI-assisted / Human-only), and Accountability Owner. A seventh column, Audit Evidence, documents what record is created at each step.

When you build this table for a real workflow at your utility, you will discover several things. You will find steps that are currently undocumented, which means they have no accountability owner today and adding AI without fixing that is a governance problem, not an AI problem. You will find steps where the classification is contested, which means your team has different mental models of where AI is appropriate; those are the conversations to have before you buy a tool. You will find steps where the AI can take over a task that currently consumes 40% of a skilled engineer's week, which is where the business case for investment lies.

A note on drift: the classification of a step is not static. When a data-center load class joins your service territory, step 6 in the forecasting map (the human override check) gets harder and more important, not easier. When a new transmission element is added, the ADMS model needs to be updated before the switching advisor can be trusted. The process map is a living document. Schedule a formal review whenever a material change occurs in the system being modeled.

A process map that shows where AI touches your workflow is also the document that shows an auditor, a commissioner, or a NERC reviewer that a human was accountable at every decision point that matters.

The NERC-Survivable Audit Trail

Every step on your process map that produces a regulatory or reliability-relevant output needs to leave a record that survives a NERC audit or a commission inquiry. The standard is not perfection; it is reconstruction. A reviewer should be able to look at the record and answer: what did the AI produce, what did the human review, what decision was made, and who made it?

The minimum fields for an AI-assisted step's log record are: timestamp (to the second, UTC), model version and configuration (so the same inputs can be re-run and the output verified), input snapshot (or a hash of the input data sufficient to confirm nothing changed), AI output (the draft, recommendation, or classification the model produced), human reviewer ID, the human's disposition (accepted as-is, modified, or rejected, with the reason if modified or rejected), and the final value released downstream.

This record does not need to live in a special AI governance system. It can live in the same workflow management system your team uses today. What matters is that it is immutable once created, accessible to auditors, and linked to the step in your process map. If your EMS historian already captures tagged-operator switching actions with timestamps, you may need to add only the AI-output field and the model-version field to have a complete record.

Under NERC CIP-003-9 (enforceable April 1, 2026), low-impact BES assets now have documented cybersecurity management obligations. Under CIP-012-2, real-time data communicated between control centers must be protected. Neither standard prohibits AI assistance in the control room, but both require that you can account for what software touched what data. Your process map and its audit log are the evidence that you can.

Key Takeaways

• Process mapping for AI integration is not optional overhead; it is the governance foundation that makes an AI-assisted workflow defensible to NERC, a state commission, and your own management chain.
• Decompose every workflow into atomic steps, then classify each as AI-ready, AI-assisted, or human-only using the criteria: is the task pattern-based with verifiable outputs, or does it require accountability for a reliability or regulatory decision?
• In the three canonical workflows (day-ahead forecasting, interconnection study processing, and outage restoration), the highest-value AI steps are completeness checking, draft generation, anomaly flagging, and option generation; the human-only steps are forecast override authorization, engineering sign-off, and switching order execution.
• AI-assisted steps are only as good as the human review behind them; design the review interface so that the human's check is substantive, not ceremonial, by surfacing the source data alongside the AI draft and requiring explicit confirmation of specific fields.
• Every step in your process map that produces a reliability or regulatory output needs an immutable audit record: timestamp, model version, AI output, human reviewer ID, disposition, and final value released.
• Treat the process map as a living document; reclassify steps whenever a material change enters the system (a new load class, a new transmission element, a regulatory reinterpretation) because the step that was AI-ready last year may need a human checkpoint added today.
• NERC CIP-003-9 and CIP-012-2 do not prohibit AI in grid operations, but they do require that you can account for what software touched what data; your process map is the primary evidence of that accountability.