Voice Desk

Lint a voice-agent prompt against four safety blocks as hard gates — disclosure, no premature confirmation, barge-in, escalation — so a prompt that reads well but has no guardrails can't ship.

Prove the prompt you approved is the one actually running. Register approved versions, hash what's live, and catch the console-edited shadow prompt — IN SYNC / DRIFTED / UNKNOWN VERSION per agent.

Score each release against your own baseline on the voice behavior a text diff can't see — barge-in, latency, dead-air, escalation — and fail CI on a regression.

Turn a week of call transcripts into per-call verdicts and a GHL field export. One deterministic pipeline — whole week, no LLM required.

Generate the warm-transfer briefing for a voice-agent escalation and grade its completeness. Reason and ask are essential — miss either and it's a context-less dump.

Reconcile booked appointments against what really happened and flag the phantoms — the confirmations a voice agent logged that never became real visits.

Find the suppressed do-not-call numbers still sitting in your dialing lists, normalizing phone formats so a DNC contact can't hide behind a different spelling of the same number.

Resolve every queued call into the called party's local time and flag the ones that would fire outside the calling window — before the batch dials.

Catch PII spoken aloud in call transcripts — a card number read digit by digit — then re-scan the output for anything that survived. What document scanners miss.

Turn your retention policy into a dated purge schedule and flag what you're over-retaining — while never scheduling a class under legal hold. It schedules only; it deletes nothing.