Interfaces
Let me finish this first.
A voice assistant needs a way to defer useful information and let a person recover it.
A suggestion at the wrong moment
Imagine two people discussing a scheduling problem. An assistant hears enough to identify a conflict and speaks while one person is explaining why they can’t move the meeting. The suggestion might be useful a moment later. At this moment, it breaks the explanation and makes everyone turn toward another speaker.
Getting the words right is only part of helping in conversation. Timing changes the cost of an otherwise good contribution. A voice interface has to consider the person’s current activity and whether the information will lose value if it waits.
Waiting needs somewhere to put the thought
For ordinary suggestions, the assistant could hold the point until the person asks or until the conversation visibly shifts. Give the person a simple way to retrieve it. “What were you going to say?” is an interaction they already understand, provided the answer is still tied to the right moment.
Urgent warnings are different. Define the circumstances in which speaking immediately is appropriate and let people understand and change the rule. In a test, separate ordinary suggestions from time-sensitive warnings so the results don’t hide what either class needs.
Evaluate the interruption and the recovery
Compare speaking as soon as the suggestion is ready with deferring it until an agreed cue. Ask whether people finish their task, notice the useful point and retrieve anything the system held back. Also ask when a pause has become long enough that the suggestion is no longer relevant.
A good result would preserve the conversation and make the useful information available while the person can still use it.