Six complete builds, each run end to end in the same order: situation, numbers, options, decision, how it is measured, and what would falsify it. These assume everything earlier in the course and exist to show the decisions connected rather than separated into topics.
| # | Lesson | What it works through |
|---|---|---|
| 01 | An In Product Copilot | Acceptance rate, trust budget, and cost per accepted suggestion |
| 02 | An AI Support Assistant | Containment against satisfaction, and the escalation design |
| 03 | An AI Analyst Feature | Grounding on a warehouse, citations, and what happens when the number is wrong |
| 04 | An Agentic Workflow That Acts | Approval gates, blast radius, and the reliability maths over ten steps |
| 05 | A Document Extraction Pipeline | High volume, cheap verification, and where the humans go |
| 06 | A Search and Ranking Rebuild | The oldest AI product problem, and why it is still the largest |
| 07 | Pricing an AI Feature | From cost per task to a value metric to a published price |
| 08 | A Model Upgrade Gone Wrong | A real regression, found late, and the postmortem |
| 09 | Killing an AI Feature | How to decide, and how to tell the customers |
| 10 | The First Ninety Days on an AI Product | What to measure and fix, in order, when you inherit one |
| 11 | The AI Product Review | The checklist, runnable against any AI feature |
| 12 | A Year in Review | One product, four quarters, and what the readouts actually said |