Answer in one sentence
A prototype is worth running when it tests a valuable, uncertain assumption with real users, defined evidence, explicit guardrails, and a decision that can change.
Teams rarely suffer from a shortage of AI ideas. They suffer from a shortage of honest selection criteria and experiments shaped around decisions rather than demonstrations.
The five-question filter
1. Whose problem is this?
Name the person, the workflow, the moment of friction, and the current cost. “The business” and “productivity” are not specific enough.
2. What must we learn?
Write the riskiest assumption as a statement that evidence could change. If no likely result would stop or reshape the work, the exercise is advocacy—not experimentation.
3. What is the smallest useful experience?
Do not minimize features mechanically. Build just enough of the real interaction to observe whether it creates value, confusion, failure, or risk.
4. What stays human-owned?
Assign accountability for inputs, outputs, exceptions, privacy, quality, escalation, and the decision to use the result. “Human in the loop” is only useful when the loop is specified.
5. What decision follows?
Define the evidence threshold and the available moves: scale, reshape, integrate, park, or stop. A pilot without a decision date tends to become permanent limbo.
The minimum evidence trail
| Artifact | What it makes visible |
|---|---|
| Opportunity brief | User, problem, value hypothesis, constraints, owner |
| Assumption map | What must be true and which unknown matters most |
| Test plan | Experience, participants, measures, timeframe, guardrails |
| Risk log | Known harms, detection, response, escalation owner |
| Decision readout | Evidence, limits, learning, and the explicit next move |
When stopping is a successful result
Stop when the problem is not valuable enough, the workflow will not support adoption, the quality or risk cannot meet the threshold, a simpler non-AI approach works better, or the evidence contradicts the value hypothesis.
A stopped experiment can be a high-value deliverable when it prevents a low-value system from becoming infrastructure.
Method note: This field guide is designed as a decision aid. Scoring should support judgment, not replace accountable review.