Why Knowing When to Stay Silent Is the Hardest Problem in Proactive AI

Most engineering effort in proactive AI systems focuses on generating insights, but the more consequential challenge lies in deciding when not to surface them. A poorly timed or unwanted notification can permanently erode user trust, since attention trained away from a channel rarely returns. The author describes a common design failure where a single numeric threshold tries to compress many unrelated suppression reasons — such as time of day, user history, and budget — making none of them individually tunable or debuggable. In their system LILA, the decision of whether an insight is worth saying is separated entirely from the policy of when and to whom it should be delivered. All candidate notifications pass through a single 12-step delivery gate that logs the specific reason for any suppression, enabling the restraint logic — often the least visible part of such systems — to be measured and improved over time.