# Prompt 10 — Translate a Stakeholder Question into a Defensible Artifact

> Adapted from Chapter 13. The cross-surface pattern that applies regardless of which Copilot surface the question lands on.
>
> **Verified against**: Power BI Copilot, June 2026 surface (applies across report-author, AI visuals, and Copilot for Power BI surfaces)

## The pattern

The most expensive class of error in AI-directed analytics is **answering the causal question with a what answer that looks like a why answer**. The pattern below converts the question to one the data can actually answer before the artifact gets built — so the eventual *why* interpretation belongs to the analyst, not to the data or the visual.

## Crystallize — reframe causal questions into *what* questions before asking

> Original stakeholder question: *"Why did denials rise in March?"*

> Reframed question for the data: *"Which denial reasons contributed most to the March rise?"*

The reframed question is answerable from data. The original causal interpretation belongs to the analyst's reading of the data, not to the data's own claim.

## Scope

Confirm the reframe with the audience if possible. **The analyst should not silently swap the question for the audience** — the swap is a defensible move only when it is named.

If the swap cannot be confirmed in advance (the question came in via email; the meeting starts in five minutes), proceed and surface the reframe explicitly at delivery time: *"You asked why; I've prepared the what. Here are the denial reasons that drove the rise. The why behind those reasons is my reading, not the data's claim."*

## Assemble

Generate the *what* artifact. In the denials example: the chart shows top denial reasons ranked by contribution to the March rise.

## Refine (corrections named by the chapter's pattern catalog)

- **Pattern 1 (hallucinated insights)**: Confirm each top reason against the source-of-truth denial log. If the visual shows a denial reason that doesn't exist in the operational system, the visual is hallucinating — fix at the data layer before letting the chart ship.
- **Pattern 4 (confidently wrong narratives)**: Read the Smart Narrative against the BRD's takeaway. If the narrative implies causation the data does not support, **rewrite or remove the narrative entirely**. Copilot's natural language tends toward causal phrasing even when the underlying analysis is purely descriptive.

See the [companion to Chapter 13](/ch/13/) for the full eight-pattern failure-mode catalog.

## Ship decision

Ship the *what* artifact alongside a one-sentence note that **names the causal interpretation as the analyst's reading of the data rather than the data's own claim**. The note is the difference between a defensible artifact and one that quietly attributes the analyst's interpretation to the chart.

---

**When to use this prompt**: The stakeholder asked a causal question (*why*, *what's the reason for*, *what caused*). This pattern applies regardless of whether the eventual artifact is a Copilot-generated report (Ch 9), an AI visual (Ch 10), a Q&A answer (Ch 11), or a hand-built chart (Part I). It is the most cross-cutting prompt in the library.