Voice input for an APEX assistant fits neatly if the region lets you edit the transcript before it reaches the agent. Does the recognition happen in the browser or do you pass audio to a service, because that choice decides whether it works on the locked-down instances most of us deploy to. And when someone dictates something like "create a task for the Miller account", how much of that arrives as structured intent versus free text the agent has to guess at. What did you settle on?