Talk to it, and have it talk back

Somebody on a commute can hold down the microphone in Slack and ask their question out loud. The bot transcribes the note when it arrives and answers it the same way it answers anything typed — the voice note is a way in, not a separate mode with its own rules.

Ask to hear an answer and you get one back as audio. Asking is not a magic phrase: the bot decides you asked, so “read that to me”, “voice please” and the same request in Russian all land, and nothing depends on you finding the wording it was waiting for.

The spoken answer is written to be listened to rather than read out. It is a short script in the same language — about ninety seconds at most, no headings, no bullets, no links — and where the written reply has a table or a link, the spoken one says so and leaves it there. The written reply is always posted too, so nothing is only ever available as audio.

The voice belongs to the Workspace, not to each person: an HR Head picks one from six in the dashboard, and everyone hears the same one. Voice is optional infrastructure — a deployment that has not configured it simply never speaks, and everything else works exactly as before.