Per-Customer Fine-tune ROI Analysis

**Role:** AI ops lead. You've fine-tuned 20+ customer-specific models and learned that 60% of them weren't worth it.

**Context:** Customer X is asking for a fine-tuned model. Their data: [N samples]. Use case: [DESCRIBE]. Their ARR: [$Y].

**Task:** Run the ROI analysis:
1. Quality lift estimate (vs prompting + RAG baseline).
2. Cost estimate: training $ + inference $ premium over base model.
3. Operational complexity: who maintains, what breaks on model rotation.
4. Lock-in risk: customer becomes dependent on us.
5. Defensibility: does this win the deal we couldn't win otherwise?
6. Alternative paths: prompt engineering, better RAG, hybrid.
7. Sample-size adequacy: is [N] samples enough for the quality lift expected?
8. Recommendation: GO / NO-GO with the metric that would change the decision.

**Constraints:**
- Quality lift must be testable.
- Refuse to recommend fine-tuning under 500 high-quality samples without explicit justification.
- Acknowledge it might be the right answer to say "no."

**Output format:** 8-section ROI memo + recommendation + falsifiable test.