# Prompt 7 — Ask a Question of the Data

> Adapted from Chapter 11 (Mei worked example, revenue-cycle model, before and after Prep for AI). The pattern for the natural-language Q&A surface — and the demonstration of why Prep for AI matters.
>
> **Verified against**: Power BI Copilot, June 2026 surface (Q&A successor in Copilot for Power BI)

## Crystallize, first pass (before Prep for AI)

> "What was net patient revenue by service line last quarter?"

## Assemble, first pass

Copilot returns a number grouped by an **internal join-key field name** rather than the human-readable service-line name. The answer is technically correct but operationally unusable — the audience cannot map the join-key labels to the service lines they know.

## Refine, first pass

The failure mode is **field-name ambiguity**. The model has multiple fields that could match "service line," and without disambiguation, Copilot picks the wrong one. Fix the failure at the model layer, not in the prompt:

1. **AI data schema**: deselect ambiguous fields (the internal join-key column should not be exposed to Copilot at all)
2. **Verified answers**: pin five canonical answers (revenue by service line, revenue by month, etc.) so Copilot's first answer to a common question is the verified one
3. **AI instructions brief**: write the brief that tells Copilot how to name service-line groupings (human-readable, not join-key)

See the [companion to Chapter 11](/ch/11/) for the three configurable surfaces this Prep-for-AI pass touches.

## Crystallize, second pass (after Prep for AI)

> "What was net patient revenue by service line last quarter?" *(same prompt as first pass)*

## Assemble, second pass

Copilot returns the number grouped by the **human-readable service-line name**, using the right measure, matching the finance close to within rounding.

## Confirmation turn (the chapter's verification move)

> "Confirm what measure you used and what grain you returned the answer at."

Copilot restates the measure as `[Net Patient Revenue]` and the grain as service line, which matches the BRD. The confirmation turn is cheap and load-bearing: it converts an implicit answer into an explicit, auditable claim.

## Ship decision

Ship when the confirmation-turn restatement matches the BRD AND the answer reconciles against the finance close. Either alone is insufficient.

---

**When to use this prompt**: You want Copilot to answer a natural-language question against your semantic model, and you want the answer to be defensible — which means the model has to be prepped first. **The Prep-for-AI pass is the prompt** as much as the natural-language question itself.

**Dataset**: [mei-revenue-cycle](/datasets/mei-revenue-cycle/) — download CSV + JSON to reproduce the before/after Prep-for-AI Q&A pass end-to-end.