AI PM Ready

Discovery & requirements

The upstream work that determines whether you build the right thing — before a single sprint begins.

The four discovery questions

Every AI PM interview will probe these in some form. Answer them in this order — do not skip to question 3.

01Is this a real problem?

How often does it occur? Who experiences it? What is the quantified cost — time, revenue, error rate? Anecdotes are not requirements.

AI NoteFor AI: can you observe the problem in data? If the answer is no, you cannot train a model to solve it. Discovery must surface the data signal before scoping begins.
02Is it worth solving?

Does solving this problem move a metric the business cares about? What is the addressable impact if you execute well?

AI NoteFor AI: unit economics must close. Cost-per-inference multiplied by decision volume must be less than the value of one correct decision. Run this calculation before any model scoping.
03Can AI solve it?

Is there a learnable pattern in historical data? Is there a feedback mechanism for improvement? Are the failure modes acceptable for this use case?

AI NoteFor AI: this is where most discoveries fail. A problem that is real and high-value may still be unsuitable for AI if training data does not exist or if false positive cost is too high.
04What does success look like?

Define measurable outcomes before solution design begins. Name the metric, the target value, and the measurement method.

AI NoteFor AI: define two layers — model success (precision ≥ 0.85, latency < 200ms) and business success (20% reduction in manual review time). Both must be defined before sprint 1.

The kill criteria principle

The sharpest CTO interview question: "At what point during discovery would you have killed this project?" Candidates who cannot answer have not done real discovery.

Data gateIf labeled examples are fewer than 50k after a 2-week audit — stop. Scope a labeling pipeline first.
Error rate ceilingIf precision cannot exceed 80% on the required prediction horizon with available features — stop.
Latency constraintIf inference P99 exceeds your serving budget (e.g., 200ms for a page-load feature) — stop and re-evaluate model architecture.
No feedback loopIf there is no mechanism to capture corrections or signals post-launch — the model cannot improve and will degrade silently.
Senior PM signal
Can name specific kill criteria before any sprint is planned.
Junior PM mistake
Discovers that AI is unsuitable after sprint 3. Kill criteria catch this in week 2.
AI vs traditional
Traditional discovery validates desirability and feasibility. AI discovery adds a third gate: data feasibility.
Interview framing
Falsification is not pessimism — it is how you protect engineering from building the wrong thing.
Was this page useful?