Cutting Call-Review Time to Zero

The problem

Front-desk call handling is one of the biggest levers on a dental clinic's revenue — missed bookings, weak follow-up, and inconsistent phone manner all show up directly in the schedule. But auditing that manually doesn't scale: once you're past a handful of locations, someone is listening to (or skimming transcripts of) thousands of calls a month, and that review is slow, expensive, and inconsistent between reviewers.

The client needed a way to score every inbound call across more than 20 clinic locations, automatically, without a human in the loop for the first pass.

What we built

The shape of the system is a multi-tenant ingestion pipeline sitting between GoHighLevel (where the calls actually happen) and a scoring layer powered by Google Gemini:

1. Every inbound call event is ingested from GoHighLevel in real time, scoped by clinic (locationId) so the 20+ tenants never cross-contaminate data.
2. Each call is transcribed and scored automatically against an 8-step evaluation framework — things like whether the booking was offered, whether the caller's question was actually answered, and call disposition.
3. Scored results land in a real-time dashboard backed by 16+ data tables and 22+ API modules, so a clinic manager can see today's calls without waiting for an end-of-month report.

A simplified version of the scoring request shape:

{
  "locationId": "loc_abc123",
  "callId": "call_9f21",
  "transcript": "Hi, I'd like to book a cleaning for next week...",
  "framework": "8-step-front-desk-v1"
}

The response feeds straight into the dashboard and the daily report — there's no manual queue a human has to clear before that data is useful.

The result

• 100% call visibility — every inbound call gets scored, not a sampled subset.
• 0 manual reviews required for the first pass — humans only get involved on the calls the system flags as worth a closer look.
• Daily reports go out automatically across three channels — SMS, email, and the internal dashboard — so the people who need the numbers don't have to go looking for them.

What I'd change next time

The 8-step framework was tuned iteratively against real transcripts rather than designed upfront — that worked, but it meant the first few weeks of scores needed a "this is still calibrating" caveat attached when shown to clinic managers. Building an explicit calibration period into the rollout communication from day one — rather than retrofitting it after managers started asking why scores looked different week to week — would have saved a round of "is the AI broken?" conversations that were really just "the model is still learning what good looks like for this clinic."