Second Chair

Every voice agent assumes one human and one assistant. That assumption breaks the moment a third person is in the room, and the conversations where speech carries legal consequence are almost never two-party: an adviser sits with a couple, a clinician with a patient and their son. Second Chair is built for those rooms. AssemblyAI's Universal-Streaming API transcribes and diarizes the conversation live, attributing every word to a speaker as it is spoken. Each finalized turn is redacted, then checked against a FINRA and SEC rule pack in under a millisecond. When the advisor says something prohibited — "guaranteed to return eight percent" — a warning reaches their earpiece alone, carrying the regulation and the exact sentence to say instead. The clients never know. The agent has two output channels. The whisper channel is private. The room channel speaks aloud, and only when addressed by name — across a twenty-turn call it speaks twice and stays silent eighteen times. A second tier runs a language model over the same turn, detached, catching omitted disclosures a pattern cannot see without ever delaying the urgent warning. Every call produces an append-only audit record: speaker-attributed, free of personal identifiers, with the regulation cited beside each finding.