The board slide reads: "AI saved us $12 million last year." The commission staff attorney reads the same number and asks: "Saved from what, compared to what, and where is the source document?" One sentence. Three questions. The utility that cannot answer all three will not recover that $12 million in its rate case. This lesson teaches you to build the board deck and the rate-case exhibit as a single, internally consistent document chain.

The Two Audiences That Define Your ROI Story

Reporting AI return on investment in a utility is more complex than in any other industry, because the same program must satisfy two audiences with fundamentally different accountability frameworks and fundamentally different time horizons. Your board cares about financial return, strategic positioning, and risk management over a three-to-five-year planning horizon. Your state commission cares about ratepayer benefit, regulatory compliance, and the evidentiary record it will use to approve or deny cost recovery. These two audiences overlap but are not the same.

The board deck mistake most utilities make is treating AI ROI as a technology story. Slides full of accuracy improvements and processing speed gains do not land in a boardroom where directors are thinking about the capital plan, the credit rating, and the regulatory compact. The correct framing for the board is: AI investment is a capital allocation decision with a measurable return, and the return is denominated in deferred infrastructure spend, reduced operating cost, and improved regulatory outcomes. Those are categories the board already knows how to evaluate.

The rate-case exhibit mistake is the mirror image: utilities try to use their board presentation as supporting evidence for cost recovery. Board presentations are internal management documents designed to convey confidence. Rate-case exhibits are evidentiary records designed to withstand adversarial cross-examination. They have completely different documentation standards, and conflating them creates problems in both directions. The board needs a compelling narrative; the commission needs a traceable calculation.

The solution is to build both from a single underlying dataset with a shared methodology, then present that dataset appropriately for each audience. This lesson walks through how to do that.

Building the Board Deck: What Directors Actually Need to Hear

A board presentation on AI ROI should have exactly four questions it answers, in this order. What did we invest? What did we get? What is the risk? What is the plan?

What We Invested

Investment in utility AI is rarely a single clean line item. It typically spans multiple budget categories: capital expenditure for servers, software licenses, and integration work; operating expense for training, ongoing vendor contracts, and internal data-science staff time; and opportunity cost of engineering hours spent on implementation rather than other projects. The board deserves a consolidated view of total program spend, by year, with the split between CAPEX and OPEX clearly marked. CAPEX matters for the rate-case cost-recovery question; OPEX affects the income statement differently. A number that blends both without explanation will confuse both the board and the commission.

For a typical mid-sized utility AI program in 2026, total three-year investment for a multi-function AI deployment (load forecasting, queue automation, and one operational use case) might run $8 to $18 million, varying significantly with system size, data integration complexity, and whether the utility builds, buys, or partners. Be honest about the full number including internal staff time, which is often excluded and which commission staff will add back in when they evaluate the cost-benefit ratio.

What We Got

The return side of the board calculation should contain three categories: CAPEX deferral, OPEX reduction, and regulatory value.

CAPEX deferral is the largest and most defensible category. As covered in the previous lesson, it requires a project-level accounting: each deferred project, the deferred dollar amount, the AI-enabled action responsible, and the present-value calculation at the utility's weighted average cost of capital. For the board, summarize this as a total number with the project count behind it. For 2026, a utility that has been operating an AI grid optimization program for two to three years should be able to document between $3 million and $30 million in cumulative CAPEX deferral, depending on system size and the leverage points the AI addressed. Do not present a single top-line number without the project-level breakout available as a backup slide, because the CFO will ask for it.

OPEX reduction is harder to quantify but real. AI-assisted interconnection queue automation typically reduces staff time on administrative and documentation tasks, potentially freeing 30 to 50 percent of engineering time on those steps. If you can quantify the FTE hours saved and apply a fully-loaded labor rate, you get a dollar number. Be conservative: only include time savings that are verifiable from timesheet or project-tracking records, not estimated. OPEX savings from better forecasting (fewer emergency energy purchases, better capacity market positioning) are also real but require access to your actual procurement economics to calculate accurately.

Regulatory value is the category boards undervalue and commissions oversee. A utility with strong AI governance and documented performance metrics is in a better position in every regulatory interaction: rate cases, service reliability proceedings, interconnection policy discussions. This is difficult to put a dollar number on, but the board should understand it as a risk-management benefit. An adverse rate-case outcome or a NERC compliance finding for an AI-related event can cost far more than the AI investment itself.

What Is the Risk

Board presentations on new technology often omit the risk section or bury it in small print. For AI in a regulated utility, risk deserves its own slide with three categories: performance risk (model drift, MAPE degradation, a system that stops delivering the projected benefits), regulatory risk (a commission that disallows cost recovery, a NERC finding related to AI operations), and cyber risk (an AI system that touches OT infrastructure is subject to CIP obligations, and a breach is a headline event). None of these risks is a reason not to invest; all of them are reasons to govern the investment carefully.

What Is the Plan

End the board presentation with the forward roadmap: what AI capabilities are being added in the next 18 months, what the incremental investment is, and what the projected incremental return is. Frame this as portfolio management, not a series of disconnected technology projects. The board should leave the presentation understanding that AI capability is being built systematically, with metrics, with governance, and with a clear line of sight from investment to regulatory outcome.

Building the Rate-Case Exhibit: The Evidentiary Standard

A rate-case exhibit on AI investment cost recovery has to survive four specific challenges: a data-request barrage from opposing parties, cross-examination of your witnesses by an expert hired by consumer advocates, review by commission staff who may be skeptical of technology claims, and, in some states, formal intervenor hearings. Each challenge requires something different from your documentation.

The first requirement is a clear investment line. Your rate-case filing must identify the specific capital projects being sought for cost recovery, with the vintage year, the account classification under FERC Uniform System of Accounts (usually Account 303, 303.1, or similar depending on jurisdiction), and the in-service date. If AI software is capitalized as an intangible asset, cite the applicable accounting treatment and its regulatory precedent in your jurisdiction. Do not assume the commission staff knows how your utility accounts for software assets; explain it in the filing.

The second requirement is a quantified benefit calculation. This is the rate-case version of the "what we got" section of the board deck, but it has to be traceable to source documents at every step. The calculation chain is: (1) AI system produced X improvement in metric M, (2) metric M improvement translated to Y operational outcome, (3) operational outcome Y avoided or deferred Z dollars of cost, (4) Z dollars of cost flows to ratepayers as W benefit per customer per month. Every arrow in that chain requires a source document in the exhibit. The commission staff will follow the chain and ask for the source document at each step.

Third, you need a benchmark comparison. Commission staff will ask whether the improvement you are claiming required AI, or whether it could have been achieved with a simpler, cheaper tool. Have an answer. Specifically: compare the AI system's performance to what a best-available non-AI alternative would have produced on the same task. For load forecasting, that is a modern statistical ensemble model. For queue automation, it is a structured workflow software that does not use AI for screening. The AI should win the comparison for the use cases you are deploying it in, and if it does not, you should not be deploying it there.

Fourth, you need a full cost disclosure, including costs that did not make it into the CAPEX line. Commission staff will identify and add back internal staff costs, training costs, and any costs paid to third parties that were not capitalized. Proactively disclosing these costs and explaining their regulatory treatment puts you in a better position than having staff discover them.

The CAPEX Deferral Exhibit in Detail

CAPEX deferral is the most financially significant AI benefit claim in most utility rate cases, and it is also the one most likely to draw challenge. Here is how to build it to hold up.

The exhibit should be organized as a table with one row per project. Each row contains: project name and account code, original capital budget, original planned in-service date, the AI system and specific operational action responsible for deferral, SCADA-verified loading data showing the circuit stayed below the trigger threshold during the deferral period, the revised planned in-service date, the years of deferral, the present value of deferral at disclosed WACC, and a reference to the supporting operational log.

The operational log reference is critical. It links the exhibit to the ADMS or EMS record that shows, for each project, the specific switching events or dispatch actions that kept loading below the trigger. This is the document chain that proves the deferral was caused by AI operations, not by load growth that came in below the original forecast.

A common challenge is: "The project was deferred because you over-invested in the original peak-load forecast, not because of AI." Your defense is to show the actual metered peak load on that circuit and compare it to the original forecast. If actual peak load was at or above the original forecast but the circuit stayed below the trigger threshold because AI-enabled topology switching redistributed load, the deferral is operationally caused. If actual peak load came in significantly below the original forecast, the commission may split the benefit: some to AI operations, some to forecast conservatism. Be prepared for that negotiation and know your numbers well enough to defend a partial attribution.

Include at the bottom of the CAPEX deferral table a disclosure of projects that were originally identified as potential deferral candidates but were not ultimately deferred. This transparency is important. Commissions expect that not every candidate defers, and showing both the successes and the non-starters demonstrates that the utility is not cherry-picking its evidence.

MAPE Improvement as a Rate-Case Benefit

Translating forecast accuracy improvement into a dollar number for a rate case requires connecting MAPE to procurement economics, which most utility planners can do but which needs to be done explicitly and transparently in the filing.

The calculation sequence: (1) identify the MAPE improvement in percentage points for the relevant horizon and season; (2) convert to MW of average forecast error at peak load (a 1 percentage-point improvement at 800 MW peak is approximately 8 MW of average error reduction); (3) identify the unit cost of procuring that MW as a reserve margin, using actual market prices or bilateral contract costs from your procurement records; (4) multiply by the number of peak operating hours to get an annual dollar figure. This number should be presented with sensitivity analysis: if MAPE improvement was 1.2 percentage points instead of the claimed 1.5, what does that do to the benefit? If reserve market prices were 20 percent lower, what does that do? Sensitivity analysis shows the commission that the benefit is real across a range of assumptions, not dependent on a single favorable parameter choice.

One nuance that frequently comes up in cross-examination: MAPE improvement does not directly translate to avoided energy cost, because reserves are typically procured in advance based on the forecast before the day-ahead market clears. The benefit of better forecasting flows through better reserve positioning, not through real-time energy dispatch. Your procurement witnesses need to be prepared to explain this mechanism, because commission staff and intervenors sometimes conflate the two.

The Regulatory Value Argument: Making the Qualitative Quantifiable

Some of the most significant benefits of utility AI are difficult to quantify precisely but are real and important: better regulatory relationships, reduced risk of NERC compliance findings, improved service quality, and organizational capability that positions the utility for the next technology wave. Regulators are skeptical of qualitative claims, but there are ways to make these arguments more concrete.

On service quality: present your CAIDI trend relative to peer utilities and relative to your historical average. If AI-assisted restoration is keeping CAIDI flat or improving while your peer group is seeing degradation (common in the current environment of aging infrastructure and more frequent weather events), that comparison is powerful evidence of customer benefit. Commission staff and consumer advocates understand CAIDI. A five-minute improvement in CAIDI across a 400,000-customer service territory has a calculable value using the commission's own Value of Service criteria from your state's reliability proceeding.

On compliance posture: document the AI governance framework, the CIP-compliance review that cleared each AI system for operational use, and the audit trail that supports your NERC self-certification. A utility that can show the commission its AI governance binder demonstrates responsible stewardship of ratepayer investment, which is a regulatory value in itself. It is harder to quantify than CAPEX deferral, but it belongs in the qualitative section of your testimony.

On competitive positioning: with 2,060 GW in interconnection queues and FERC acting on large-load interconnection policy in 2026, utilities that can process interconnection studies faster and more accurately are better positioned to support economic development in their service territories. That is an economic development argument the commission understands and values, especially in states where data-center attraction is a policy priority.

Worked Example: The Complete Exhibit Chain

Imagine a regional investor-owned utility filing a rate case in 2026 covering three years of AI investment totaling $11.4 million. The utility deploys AI in three areas: day-ahead load forecasting, interconnection queue automation, and distribution topology optimization. Here is how each benefit category flows from board deck to rate-case exhibit.

The board deck presents total investment of $11.4M, total attributed benefit of $19.2M NPV over five years, yielding a net benefit of $7.8M and a benefit-cost ratio of 1.68. That is the board slide. Behind it is a three-tab spreadsheet, one tab per program, showing the calculation for each benefit category. The board presentation is supported by the spreadsheet; the spreadsheet is supported by source documents; the source documents are the rate-case exhibits.

For load forecasting, the rate-case exhibit shows: baseline MAPE of 3.3% (measured for 24 months pre-deployment on the same holdout methodology), current MAPE of 1.7% (measured for 30 months post-deployment), improvement of 1.6 percentage points, equivalent to 12.8 MW average error reduction at 800 MW peak, equivalent to $840,000 per year in avoided reserve procurement costs at the utility's actual reserve market economics. Exhibit Attachment A contains the raw MAPE data by month. Exhibit Attachment B contains the reserve procurement methodology explanation. Exhibit Attachment C contains three years of reserve market pricing used in the calculation.

For queue automation, the rate-case exhibit shows: baseline throughput of 4.1 studies per engineer per month (measured for 18 months pre-deployment), current throughput of 6.3 studies per engineer per month (measured for 30 months post-deployment, normalized for staffing), improvement of 2.2 studies per engineer per month, equivalent to approximately 0.4 FTE of recaptured engineering capacity at $165,000 per FTE fully-loaded, yielding $66,000 per year in labor efficiency. The rate-case argument for queue automation is not primarily about cost savings (the number is modest), but about regulatory capability: the utility can now process the increasing volume of large-load interconnection requests triggered by FERC's 2026 large-load rulemaking without adding staff, which benefits both developers and ratepayers.

For topology optimization, the rate-case exhibit is the CAPEX deferral table described above: seven projects, $2.1M to $6.8M each, deferred one to three years, total PV of deferral $4.3M. Each project row has its SCADA reference and operational log attached as a sub-exhibit. The two projects that were originally identified as potential deferrals but were not deferred are disclosed with explanation.

Total attributed benefit in the rate-case exhibits: $840K per year from forecasting, $66K per year from queue automation, $4.3M PV from topology optimization. The board's $19.2M five-year NPV is the sum of these streams discounted at the utility's WACC, which matches the exhibit totals to the dollar. This internal consistency, the board number traces exactly to the exhibit numbers, is what protects the utility in cross-examination. When the consumer advocate's expert asks where the $19.2M comes from, the answer is three clean exhibit chains, each with source documents.

Key Takeaways

• Build the board deck and the rate-case exhibit from a single underlying dataset and methodology. Internal consistency between the two documents is your primary defense against cross-examination.
• The board needs four questions answered: what we invested, what we got, what is the risk, and what is the plan. Frame AI return in terms directors understand: CAPEX deferral, OPEX reduction, and regulatory value.
• The rate-case exhibit requires a traceable calculation chain: AI improvement to operational outcome to avoided cost to ratepayer benefit, with a source document at every link in the chain.
• CAPEX deferral is the largest and most defensible AI benefit category. Each deferral claim needs a specific project, a documented AI-enabled operational action, SCADA-verified loading data, and a present-value calculation using disclosed WACC.
• MAPE improvement translates to rate-case dollars through reserve procurement economics, not real-time energy dispatch. Your procurement witnesses need to understand and explain this mechanism.
• Proactively disclose AI investment costs that were not capitalized (internal staff time, training, integration overhead) and projects that did not achieve deferral. Transparency on the full picture is more credible than a selectively positive presentation.
• Peer-utility CAIDI comparison is a powerful qualitative benefit argument when commissions already have Value of Service criteria. A five-minute CAIDI improvement across 400,000 customers has a calculable dollar value using commission-established metrics.