# Mei's Revenue-Cycle Prep for AI — Chapter 11 Worked Example

> The five-CFO-question before-and-after trace behind Mei’s Prep-for-AI pass on the revenue-cycle model in Chapter 11 of *The Defensible Decision*. Two work cleanly; three break in three different ways; the Prep pass fixes two and correctly refuses the fourth.

**Wholly synthetic.** Any resemblance to a specific real healthcare-network revenue-cycle model is coincidental; the five-question failure distribution is drawn from typical practitioner-observed patterns on Copilot-enabled semantic models.

## Files

- [`mei-revenue-cycle.csv`](mei-revenue-cycle.csv) — flat rows: one per CFO question with before verdict/notes, after verdict/notes, and the Prep component that closed each gap
- [`mei-revenue-cycle.json`](mei-revenue-cycle.json) — canonical scenario contract with source hash

## What it represents

Five CFO-shaped questions Mei asked Copilot against a healthy hand-built revenue-cycle semantic model, in one session. Before Prep-for-AI: two questions work cleanly, one returns a number Mei cannot defend (definition ambiguity), one confidently returns a wrong number (denial-rate vs denial-count confusion), one returns a hedged non-answer even though the data is there. After running the three Prep components (schema simplification + verified answers + AI instructions) in a fresh session: four questions work cleanly; the fifth (a causal «why did denials rise?» question) correctly refuses per the AI-instructions directive on causal questions.

## Invariants (locked)

The scenario carries executable invariants that Chapter 11’s release contract enforces:

- Exactly 5 questions with ids 1..5 and no gaps.
- Every before/after verdict is from `{worked, ambiguous, wrong, refused}`.
- Before-state distribution: 2 worked, 3 ambiguous-or-wrong.
- After-state distribution: 4 worked, 1 correctly refused.
- The refused-after question is the causal («why...») shape.
- All three Prep components are present in the contract (`ai_data_schema`, `verified_answers`, `ai_instructions`).
- Schema simplification deselects at least one table and multiple columns.
- Verified-answers pins at least 5 trigger-phrase / verified-answer pairs.
- AI-instructions declare `includes_causal_refusal: true`.
- Before-after summary counts reconcile with per-question verdicts (no drift).

## Why the invariants matter

Chapter 11’s pedagogical point is that a model that is good for hand-built reporting is not automatically ready for Copilot; the three Prep components each fix a specific *class* of failure that hand-built discipline never had to address. If any of the three components were skipped, one of the four fixed questions would still fail. If the causal question quietly returned an answer post-Prep, the «refuse causal questions» discipline would collapse into «maybe answer them.»

## What it does NOT represent

- Actual healthcare revenue-cycle data from any real payer or provider.
- The specific measure definitions (the dataset names the ambiguity; the resolution lives in the AI-instructions brief).
- Payer-name detail (the fourth question is about denial-rate change by payer; specific payer identities are out of scope).

## Reproducibility

The dataset ships from the book source at `AlexBooks/applied-ai/ai-directed-viz/data/mei-revenue-cycle.json` and is re-emitted here by `AlexBooks/applied-ai/ai-directed-viz/scripts/generate-mei-revenue-cycle.mjs --site-root <site-root>`.

## License

Released for teaching use alongside *The Defensible Decision*. Use freely.