The Cost–Latency–Quality Triangle

Every AI feature lives in a triangle: cost (tokens in and out, per user action), latency (time-to-first-token and total response), quality (frontier model, longer reasoning, richer context). You can max two: frontier quality fast → expensive; cheap and fast → weaker model; quality and cheap → async or batch with slow UX.

The PM's job isn't choosing once — it's setting the operating point per surface. Autocomplete needs cheap+fast; the weekly report tolerates slow+quality.