Use this for internal pilots, betas, or anything customer-facing.
Can the core task be completed without AI? If not, document what happens when the model fails.
Do you have at least five real inputs from actual users — not demo scripts?
Can you state in one sentence what data the model receives on each request?
Is there a human path when the answer is empty or low-confidence?
Has someone outside the build team used it and reported a dead end?
If the answer is wrong, how will the user know — and what should they do next?
Mark every “no” before launch. Those items are blockers, not backlog.