Use this before committing any engineering time. Kill bad ideas here, not after a prototype.
Next: if you decide to pursue or prototype, write the AI PRD. If the idea is not coherent across user, AI job, human control, evals, risk, and cost, stop here and sharpen it before building.
Problem
- Target user:
- Workflow where the problem happens:
- Current workaround:
- Cost of the problem today: time, money, errors, missed revenue, risk
Current workflow
Why AI, and why not deterministic software
Smallest useful version
AI job draft
The AI [does what] using [inputs] to produce [outputs] for [user] inside [workflow], subject to [review or safety rule].
- Inputs:
- Outputs:
- Explicitly out of scope:
Human control
| AI action | Review mode | Reviewer or monitor | Why this mode |
|---|---|---|---|
| in / on / after / none |
Data, eval, and operating assumptions
- Data sources:
- Data quality concerns:
- First eval cases to create:
- Cost per task target:
- Observability signals needed before pilot:
Alternatives considered
| Alternative | Pros | Cons | Why not (or why maybe) |
|---|---|---|---|
What happens if we do nothing
Risks already visible
| Risk | Scenario | Likelihood | Severity | Mitigation or next action | Owner |
|---|---|---|---|---|---|
Highest-severity risk
Legal, security, privacy, or compliance concern:
What would make this a "do not launch" decision?
Prototype success criteria
| Question the prototype must answer | Proceed if | Stop if |
|---|---|---|
Decision
- Pursue: commit to a PRD and eval plan
- Prototype: time-boxed spike to answer the questions above
- Defer: revisit in timeframe
- Reject: not worth pursuing because reason
Decided by: name Date: YYYY-MM-DD