Validate an AI Product
A useful product begins with a specific job and evidence that people can complete it more safely or effectively—not with a revenue screenshot, a model wrapper, or a fixed launch calendar.
Define the smallest useful job
Interview consenting target users about a recent instance of the problem. Ask what triggered it, what they tried, what failed, and what a wrong answer would cost. Do not send private interview material to a model without explicit permission and an approved data environment.
User: one clear role and context
Job: one observable task, not “use AI”
Current alternative: what happens today
Risk: harm from wrong, late, leaked, or inaccessible output
Evidence: consented interviews, observed workflow, or product telemetry with a documented purpose
Prototype with boundaries
- Use synthetic examples to establish a basic task-specific evaluation set.
- Add representative, adversarial, multilingual, and accessibility cases.
- Separate model output from policy and permissions; enforce controls in code.
- Show uncertainty and route high-impact decisions to a qualified human.
- Log only what is necessary, protect tenant boundaries, and test deletion.
- Provide a usable fallback when the model, vendor, or network fails.
Use a zero-default economics sheet
Monthly contribution = collected revenue (0) − model usage (0) − hosting (0) − payment fees (0) − refunds (0) − support and review labor (0) − compliance and incident cost (0). Include taxes and local obligations with qualified advice. Do not treat founder time or organic distribution as free.
Test willingness to pay with an honest offer and cancellation terms. Never fabricate demand, testimonials, waitlists, scarcity, or “spots.” A pre-order needs clear delivery risk, refund terms, and a realistic description of what exists today.
Stage decisions by evidence
Explore
Proceed when repeated, consented observations show the same costly or frustrating job. Stop or reframe when evidence is weak.
Prototype
Proceed when the prototype clears predefined quality and safety thresholds on a versioned evaluation set—not when a demo merely looks impressive.
Pilot
Proceed when a small group can use it with support, informed consent, reliable fallback, and acceptable incidents.
Operate
Expand only after retention, support load, accessibility, privacy, and contribution economics remain acceptable across more than one observation window.
Rights and trust checklist
Ship a learning loop, not a promise. Keep the first version reversible.
The goal is a defensible decision: continue, change direction, or stop. All three are valid outcomes when they follow transparent evidence.