Challenge

AI generated protocol content even when underlying decisions were incomplete or not yet made. In a high-stakes, regulated context, partially correct or assumption-based outputs could mislead users and undermine trust. This created a fundamental dilemma between speed and correctness: generating everything made progress feel fast but unreliable, while generating nothing preserved accuracy but reduced perceived value.

Diagram showing the trade-off between no AI generation and full generation, highlighting the tension between speed and correctness in protocol content creation.

Decision

AI should not generate by default. Its behavior must be constrained by the certainty of underlying information. Based on this, generation was gated into three modes: auto-generate for confirmed information, suggest with user validation for high-confidence insights, and no generation when information is incomplete. This makes AI behavior predictable and preserves trust in high-stakes use.

Three-level model of AI text generation based on information certainty, defining when to generate, suggest, or not generate content.
AI generation is constrained by the certainty of underlying information.
Protocol generation panel showing how AI output changes based on information certainty.
Mapped the confidence-based model into the interface, making generation status and next actions visible to users.