Finding out whether a model can do the thing, cheaply, before committing a quarter to it. The discovery questions change when the component is probabilistic: the demand question is the same, but two new ones arrive, which are whether the model can reach the bar at all and who absorbs the mistakes.
| # | Lesson | The question it answers |
|---|---|---|
| 01 | Demos Lie | Why do AI demos convert and AI products churn? |
| 02 | Find Tasks, Not Features | How do you decompose work into tasks a model could hold? |
| 03 | The Ceiling Test | Can the model do this at all, and how do you find out in a day? |
| 04 | The Error Tolerance Question | What happens when it is wrong, and who absorbs that? |
| 05 | Wizard of Oz and Concierge | How do you test demand with humans before any model? |
| 06 | Sizing an AI Opportunity | Volume times value times tolerance for error, with numbers |
| 07 | Watching People Do the Task Manually | What does the current workflow tell you that a brief cannot? |
| 08 | The Data You Would Need | Does the context this requires actually exist and may you use it? |
| 09 | Evaluating Build, Buy or Wait | Is somebody about to ship this for free? |
| 10 | Prototyping Without Committing | How far can you get in a week, and what does that prove? |
| 11 | The Discovery Readout for an AI Feature | What has to be in it that is not in a normal one? |
| 12 | Killing an AI Idea | What are the honest reasons to stop, and when do they appear? |