Skip to main content
Back
35 XP

Lesson 1 · 9 min

How AI features are different from regular features

If you spec an AI feature like a regular feature, it ships broken. The four ways AI features change the PM playbook — and what to do about each.

Four differences

  1. Output is probabilistic, not deterministic. Two identical inputs can produce different outputs. The spec needs an acceptable-output range, not a single golden case.
  2. The model can be wrong, confidently. Hallucination is real. The PM needs to spec the recovery path (refusal, escalation, citation) as carefully as the success path.
  3. Cost scales with usage in a way regular features don't. A successful AI feature has a usage-cost curve that turns into a CFO conversation around month 4. The PM needs cost in the spec, not as a back-end concern.
  4. The capability ceiling moves under your feet. A feature that needed Sonnet last quarter may run on Haiku this quarter. The PM needs a quarterly review built into the lifecycle.