Upload a folder of recordings, sign off on the QA rubric in chat, and the Agent transcribes with diarization, scores every call on your dimensions, and flags compliance violations down to the exact second. Schedule it to run at 2 a.m. and the .xlsx is waiting in your Drive at 8.
The Agent runs the previous day's calls at 02:00 every morning. The .xlsx is ready in Drive by 08:00. You only open it when violations cross a threshold — 100% coverage replaces 5% sampling.
Daily 02:00 · last 24h of calls · writes audit-{date}.xlsx
Timestamps are second-precise; the transcript shows the surrounding 60 seconds. Managers coach on the real moment — not on generalities.
A zip of yesterday's calls, or point the Agent at a Drive folder. MP3 / WAV / M4A / AAC all work. Recordings up to ~1 hour per file process comfortably; longer ones get chunked under the hood.
The Agent proposes dimensions and weights, you edit them in chat ("raise empathy weight, lower closing weight", "add a dimension for hold-time discipline"). No scoring happens until you say go. Save the rubric and tomorrow's run reuses it automatically.
One chat command — "run this every day at 2 a.m. against the previous 24 hours of calls" — and the Agent schedules the job. Output lands in /audits/{date}.xlsx in Drive each morning before standup. Violations sheet is separate so the compliance team can review without scrolling.
Cheap call-audit tools transcribe to one big text blob and your QA team has to guess "was that the agent or the customer". The Agent uses speaker diarization via `vb asr` and labels every line. Compliance violations (e.g. "did the agent disclose the recording?") become trivially answerable, not a manual relistening exercise.
A daily schedule runs at 2 a.m., the .xlsx is ready by morning, and the compliance lead opens it only if violations exceed a threshold. Most days, no one opens the file — that's the point. Manual call sampling ("we listen to 5% of calls") gets replaced by automated 100% coverage with human exception handling.
When the .xlsx flags a call at 12:47, you click the timestamp and the player jumps to 12:42 (5 seconds of lead-in). No "I think it was somewhere in the middle". This makes coaching real — the manager can replay the exact moment instead of arguing about generalities.
The QA rubric is co-designed in chat, weighted by you, signed off by you. The Agent applies it consistently across every call. When the dimensions need to change (new product launch, new regulation), edit the rubric once and tomorrow's run uses the update. No "the AI started scoring differently this week".
Phone-band audio (8 kHz, narrowband codecs) is harder than broadcast audio for every provider. The Agent picks from whichever providers are enabled in your workspace via `vb asr providers` — typical cue-level error rates we see on call audio are 5–8% on clean lines and 10–14% with heavy accents or background noise. Low-confidence segments are flagged in the .xlsx so QA reviewers focus on the ambiguous cues. For tonal languages (Mandarin, Cantonese, Vietnamese), you can pin a provider known to handle tonal language well in chat.
Sign in, hand it over to the Agent — the finished file lands in your Drive.