Before any concept, the mindset. Four things change, and ignoring them is how AI features fail.
ONE. Output is probabilistic, not deterministic
In traditional software, the same input gives the same output. add(2, 3) always returns 5. In AI, the same input may give a different output every time, and even the correct output varies in tone, length, and wording. This breaks several PM habits.
- QA pass/fail doesn't apply. You can't write "the system returns X" as an acceptance criterion. You write "the system returns a summary scoring ≥4/5 on faithfulness in 95% of cases."
- Single-sample testing lies. A demo that works once may fail 20% of the time. You need test sets, not anecdotes.
- Bugs become statistical. A dropped quote is either one-off variance or a regression. Telling them apart needs evaluation infrastructure.
TWO. The triangle: quality, cost, latency
Every AI feature decision reduces to a position on this triangle. Larger models are smarter but slower and more expensive. Smaller models are fast and cheap but miss nuance. Caching, routing, and retrieval are all moves on this triangle.
"Use the best model" is not a strategy. It is avoiding the decision.
THREE. Evaluation is the new core competency
Traditional PM skills stay: discovery, prioritization, stakeholder management, GTM. What gets added is evaluation, and it becomes load-bearing. You define the rubric. You curate the golden set. You decide what "good enough" means. You run the eval before every prompt change, every model upgrade, every retrieval tweak. Lesson 14 introduces it; Lesson 16 turns it into a full operating manual.
FOUR. The model is a vendor surface that updates on you
Anthropic ships a new Claude version. Your prompt that used to work now misbehaves. OpenAI deprecates a model. Your fine-tune is suddenly orphaned. The model is not "the engine," it is a vendor surface that updates whether you ask for it or not. PMs who think of the LLM as static infrastructure get blindsided. The job is closer to managing a third-party SDK that occasionally rewrites itself.
What the PM owns from Day 1
The eval rubric. The cost ceiling per request. The latency budget. The fallback behavior when the model fails. The trust UX (citations, confidence, "I don't know"). The escalation path when something goes wrong in production. These six decisions are yours from Day 1.
The case
Before writing a single line of the PRD, Jack commits to owning three decisions even though the numbers won't be final on Day 1.
- The rubric. "A summary is acceptable if it captures all named entities, doesn't add facts, and is under 60 words." To be refined with editors over two weeks.
- The latency target. "Editors shouldn't feel they're waiting." Translates to roughly 4 seconds, validated later against the desk's actual tolerance.
- The cost discipline. Estimate per-request cost at each model tier (Lesson 17), pick the cheapest that clears the rubric, propose a ceiling with 30% headroom. The ₹0.20/summary ceiling arrives three weeks later, after prototyping and business approval.
What matters Day 1 isn't that he knows the numbers; it's that he owns these three decisions instead of letting them happen by default.
Further reading
- Anthropic · Evaluating AI Systems
- Hamel Husain · Your AI Product Needs Evals