How do you authorize a voice AI agent when there's no screen?

There is no consent dialog in a voice conversation. Voice agents run tools while they keep talking, and every consent pattern we rely on assumes a screen and a click. What authorization looks like without either.