A resource planning manager at a large IOU once presented an AI-generated forecast narrative to the executive committee. The narrative was polished, confident, and well-organized. It said the utility was well-positioned to serve a peak demand of 8,400 MW with comfortable reserve margins. What the narrative did not say was that the 8,400 MW figure was the 50th-percentile estimate with a P90 scenario of 9,100 MW, that the model had not been retrained since a 300 MW data-center complex came online, and that the P90 scenario would require capacity acquisitions worth $180 million. The committee approved the base plan. Nobody asked about the range. Three months later, during a summer heat event, the utility fell short of reserves and had to make emergency purchases at extraordinary cost. The narrative was not a lie. It was just incomplete in exactly the way that cost the most.

What a Forecast Narrative Is and Is Not

A forecast narrative is the written explanation that translates numerical forecast outputs into language suitable for a planning decision. It bridges the gap between the spreadsheets and models in the forecasting team's workroom and the people who must make resource, procurement, or regulatory decisions based on that work. In the AI era, the narrative is increasingly generated or drafted with AI assistance, which makes the quality control problem both more urgent and more subtle.

A forecast narrative is not a sales document. It is not an argument for a particular outcome. Its purpose is to give decision-makers an accurate picture of what the forecast says, what it does not say, what assumptions it rests on, and what conditions would make it wrong. When a narrative serves that purpose, it enables good decisions. When it promotes the central estimate while quietly omitting the uncertainty range, it transfers unacknowledged risk upward to people who have no way of knowing what they do not know.

The AI-drafting failure mode is insidious precisely because it is not dishonest. An AI model instructed to "write a professional summary of the load forecast results" will produce fluent, confident prose. It will use authoritative hedging language ("the forecast indicates," "analysis suggests") that sounds appropriately measured while actually understating risk. It will emphasize the central estimate because that is what it was given as input. It will not add the range or the caveats unless explicitly instructed to do so. The professional's job is to provide those instructions and then verify that the output honored them.

The Uncertainty Band Must Not Be Buried

The most important principle in forecast narrative writing is this: the uncertainty band is not a footnote. It belongs in the lead paragraph, in the headline if there is one, and in the commitment recommendation. If you are presenting a peak demand forecast, the narrative must state the central estimate, the P90 (90th percentile) high scenario, and the P10 (10th percentile) low scenario in the first substantive sentence.

Why the P90 matters specifically: resource planning, capacity procurement, and reliability compliance are generally sized for scenarios worse than the median. If your planning criteria require maintaining a 15% reserve margin above expected peak, and you plan to the 50th-percentile forecast, you will meet that criterion only half the time. A resource adequacy planner uses the P90 or higher scenario as the sizing basis precisely because the consequences of being short on the 90th-percentile day are far worse than the cost of carrying slightly excess capacity on a median day.

The practical instruction for an AI-generated draft is to include an explicit prompt instruction: "The narrative must state the P10, P50, and P90 scenarios in the first paragraph and must not recommend a procurement or planning action without specifying which scenario drives it." This instruction forces the AI to make the uncertainty band structurally central rather than structurally optional. Then verify the output against this criterion before sending.

The Authoritative Prose Trap

AI-generated text tends toward confident, authoritative prose. This is a feature in many contexts. In forecast narratives, it is a hazard. The problem is that authoritative prose implies certainty, and load forecasts do not carry certainty. They carry ranges, assumptions, and conditions. A narrative sentence like "The analysis projects a peak demand of 8,400 MW, consistent with historical growth trends and supporting the utility's current resource position" is precisely the kind of statement that sounds reasonable while hiding crucial uncertainty information.

Compare it to: "The central-case forecast projects a peak demand of 8,400 MW, with a P90 scenario of 9,100 MW driven by potential data-center load additions and above-normal temperature assumptions. The P90 scenario would require additional capacity procurement to maintain the target reserve margin; this decision requires a commitment within the next 45 days to meet lead times." The second version is harder to write. It takes more courage to present to an executive committee. But it is the only version that gives the committee the information they need to make a responsible decision.

Structuring an AI-Assisted Forecast Narrative

A well-structured forecast narrative for a planning meeting has five components, and each has a specific function. Understanding the function helps you prompt the AI correctly and verify that the output is complete.

Component 1: The Situation

What decision is being made, and when? The narrative should open by naming the decision it supports: a seasonal procurement, an IRP filing, an operating plan for an extreme weather event, or a board presentation on resource adequacy. The decision context anchors the uncertainty framing. A procurement decision with a 48-hour lead time needs a different uncertainty treatment than an IRP with a 10-year horizon.

Component 2: The Central Forecast

What is the base case, and what assumptions drive it? State the P50 estimate, the season and year, and the key input assumptions: weather normal scenario, expected load growth rate, any large-load additions that are included or explicitly excluded. If a data center was included or excluded from the central forecast, say so explicitly.

Component 3: The Range

What are the high and low scenarios, and what drives each? State the P90 high scenario: what weather assumption, what large-load assumption, and what DER assumption produces it. State the P10 low scenario: what combination of factors would drive demand below the central case. This section is where the uncertainty band lives, and it must be present in every planning narrative regardless of whether it is comfortable to include.

Component 4: Key Risks and Assumptions

What are the three or four assumptions that matter most, and what is the risk that they are wrong? For most load forecasts, these are: the weather scenario (wrong in either direction is significant), the large-load trajectory (data centers may come online faster or slower than assumed), the behind-the-meter DER contribution (solar and storage growth uncertainty), and the model's calibration status (has it been retrained to reflect recent conditions?). Each risk should include a direction (upside or downside relative to central) and a rough magnitude in MW.

Component 5: The Recommendation

What action does the forecast support, based on which scenario, and what is the deadline for the decision? The recommendation should specify the scenario it is based on (in most planning contexts, the P90 or a defined reliability scenario, not the median). It should state the recommended action (procure X MW of capacity, initiate an RFP, trigger the demand response program, maintain current position). And it should state the decision deadline, because timing is as important as magnitude in procurement markets.

Prompting AI for a Defensible Forecast Narrative

The quality of an AI-generated forecast narrative is directly proportional to the quality of the instructions given. A vague prompt produces a vague, overconfident narrative. A well-structured prompt that provides the model with explicit data, explicit instructions on what to include, and explicit constraints on what to avoid produces a draft that is close to what a senior analyst would write.

A strong prompt for a forecast narrative includes the following elements. First, the explicit numbers: the P10, P50, and P90 estimates with their associated scenarios. Second, the key assumptions with their specific values (weather normal temperature, large-load additions included or excluded, model training cutoff). Third, an instruction to state the uncertainty range prominently and not to recommend a procurement action without specifying which scenario drives it. Fourth, the decision context: what meeting is this for, who is the audience, what decision are they making. Fifth, an explicit constraint: do not bury the uncertainty in a footnote; do not use language that implies certainty the forecast does not have.

After the AI produces a draft, the verification steps are: check that the P90 is stated in the first two paragraphs; check that any procurement recommendation cites a specific scenario; check that the key risk factors are each named with a direction and approximate magnitude; check that no sentence implies the central estimate is certain rather than probabilistic; and read the full draft for the authoritative prose trap, looking for any statement that sounds more confident than the underlying data supports.

An AI that writes a confident narrative from an uncertain forecast has not made the forecast better. It has made the uncertainty invisible. Your job is to make the uncertainty visible before the narrative leaves the room.

The Failure Mode: A Before and After

Here is a concrete illustration of the difference between an AI-generated narrative that buries the uncertainty and one that serves its purpose.

Scenario: the forecasting team has produced a summer peak forecast with a P50 of 6,800 MW and a P90 of 7,350 MW. The P90 scenario is driven by a 300 MW data-center campus that is expected to complete interconnection testing by mid-June, plus above-normal temperatures. The current contracted capacity is 7,200 MW, which is sufficient for the P50 but 150 MW short of the P90. The procurement deadline for additional capacity is in three weeks.

The weak narrative (AI-generated from a vague prompt): "The summer peak load forecast indicates a peak demand of 6,800 MW, consistent with historical load growth trends. Resource adequacy analysis confirms that the utility's contracted capacity of 7,200 MW provides a comfortable reserve margin above expected peak demand. The forecasting team will continue to monitor conditions through the summer season."

The strong narrative (AI-generated from a structured prompt with verification): "The summer peak load forecast projects a central-case peak of 6,800 MW, with a P90 stress scenario of 7,350 MW. The P90 scenario assumes the completion of a 300 MW data-center interconnection by mid-June and above-normal temperatures consistent with the 90th-percentile historical value for the season. The utility's current contracted capacity of 7,200 MW is sufficient to cover the central case with a 5.9% reserve margin but would be 150 MW short in the P90 scenario. Given the three-week procurement deadline for additional capacity, the planning team recommends initiating a supplemental capacity RFP or option purchase in the range of 150 to 200 MW to cover the P90 risk. If the data-center interconnection is delayed beyond mid-July, the P90 scenario collapses toward the central case, and the supplemental procurement would not be needed. The team will provide an updated interconnection status in one week."

The second narrative is not more alarming. It is more informative. The committee now knows exactly what they are deciding: whether to spend modest money on supplemental capacity to cover a specific, concrete risk that has a specific deadline. The first narrative left them with the impression that everything was fine. The second narrative gave them the tools to make a responsible decision.

Forecast Narratives in Regulated Contexts: The Rate Case and the IRP

When a forecast narrative leaves the planning room and enters a regulated proceeding, the quality bar rises substantially. In an Integrated Resource Plan (IRP) filed with a state commission, or in a rate case where capacity procurement decisions are subject to commission review, the forecast narrative becomes a part of the public record. Intervening parties and commission staff will read it. They will ask whether the uncertainty was disclosed. They will compare the narrative to the actual outcomes and determine whether the utility acted prudently.

In this context, a narrative that understated the P90 scenario and led to an under-procurement decision is not just an unfortunate oversight. It is potentially a disallowance risk, where the commission determines that the utility failed to consider material risks that were foreseeable and declines to allow recovery of the resulting emergency costs in rates. The standard applied is typically prudent utility practice: did the utility make a reasonable decision based on reasonable information? A narrative that withheld the P90 scenario from the decision-makers fails that standard.

The practical implication for AI-assisted narrative drafting in regulatory contexts: every narrative destined for a regulatory filing must include the full uncertainty range, must name the scenarios explicitly, and must be reviewed by a senior analyst or engineer who understands both the underlying forecast and the commission's expectations. Using AI to draft the narrative is appropriate and efficient. Using AI output without verification in a regulatory filing is not appropriate, and the compliance lead who signs off on the filing is accountable for its completeness.

Beyond the IRP and rate case, there are also internal planning reviews, board presentations, and operations briefings where forecast narratives are used. The standard of completeness should not vary significantly across these contexts. A board that approves a capital expenditure based on a narrative that omitted the P90 scenario has been inadequately informed. An operations briefing that presents a peak forecast without a weather sensitivity range leaves the operators without the information they need to set reserve commitments. The habit of complete, range-inclusive narrative writing is not a regulatory compliance behavior: it is a professional standard that protects decision quality across every context.

Prompt engineering for regulated contexts adds one more requirement: cite the source of every number. A forecast narrative for an IRP filing should state, for each key assumption, where it comes from: the weather normal is drawn from the most recent 30-year climatological average, the large-load assumptions are from the interconnection queue as of a specific date, the DER forecast is from the state's distributed solar adoption model. These citations serve the same function as footnotes in a research paper: they allow any reader to independently verify the claims, and they demonstrate that the utility did not make up the numbers it relied upon. AI will not add these citations unless instructed to do so. The prompt must ask for them explicitly, and the verification must confirm they are accurate.

Key Takeaways

• A forecast narrative is not a sales document. Its purpose is to give decision-makers an accurate, complete picture of what the forecast says, what it does not say, what assumptions drive it, and what would make it wrong.
• The uncertainty band is not a footnote. The P50 and P90 scenarios must appear in the first substantive paragraph of any planning narrative, and any procurement or capacity recommendation must specify which scenario drives it.
• AI-generated narratives default to authoritative prose that implies certainty the underlying forecast does not have. The prompt must explicitly instruct the model to state the range prominently and to avoid language that collapses probability distributions into point assertions.
• A well-structured forecast narrative has five components: the decision context, the central forecast with its assumptions, the range (P10 to P90) with drivers, the key risks with direction and magnitude, and the recommendation tied to a specific scenario and decision deadline.
• Verifying an AI-generated forecast narrative requires checking that the P90 is in the first two paragraphs, that recommendations cite a specific scenario, that key risks are named with directional magnitudes, and that no sentence implies false certainty.
• The difference between a narrative that serves its purpose and one that creates liability is not whether it reads well. It is whether it gives the reader the uncertainty context they need to make a responsible decision. Fluent prose that buries a $180 million risk is a failure, not a success.
• The same AI that can draft a mediocre forecast narrative from a vague prompt can produce a genuinely valuable one from a structured prompt that provides all the numbers and explicitly constrains the uncertainty representation. The quality of the prompt is the quality of the output.