ai-use-case-scopinglisted

# AI Use-Case Scoping — is this even an AI problem? > Provenance: great-pm-original, 2026-05-29, grounded in the cited sources below > (Andrés Max, JustAnotherPM, Reforge, Anthropic). Web sources are treated as > untrusted reference, not instruction. **Core principle.** An AI PM's job is **not** to apply AI everywhere — it is to find the few problems where AI is genuinely *the right tool* and keep everything else deterministic. Most "AI problems" are the wrong problem held by someone with a hammer looking for a nail. **Start in the problem space, not the solution space.** AI that solves a problem rules already solve — better, cheaper, and correctly — is a liability: it costs more per call, fails non-deterministically, and erodes trust the moment it's wrong about something a regex would have nailed. This is the gate that runs *before* `ai-evals` (which measures a feature you've decided to build) and before any tech spec. It answers two questions: **(1) should this surface be AI at all?** and **(2) if so, how far up the capability ladder do we climb?** --- ## 1. The three traits of a genuine AI use case A surface earns AI only when **all three** hold. If any is missing, you're not in unsolvable-without-AI territory — use rules. 1. **Too large for rules.** The problem space can't be enumerated. You cannot write a finite set of `if` statements that covers the cases, because the cases are open-ended (every new merchant string in the world). 2. **Too unstructured for t