Every call the clinic takes, read and scored within two minutes
A three-site London clinic group could not see what happened on its phones. iGrowix built an autonomous system that transcribes every call, classifies the outcome, grades how reception handled it, and turns wasted calls into advertising decisions.
5,446
Calls transcribed and analysed in six weeks
171 hrs
Manual conversation review avoided entirely
1.3 min
Median time from call ending to full analysis
100%
Coverage across all recorded calls, not a sample
Thousands of calls a month, and no idea what happened on any of them
The group takes around 130 calls a day across three clinics, answered by eight reception extensions and arriving on seven different numbers, each published on a different website or advertising campaign. Every call was recorded by the practice's 3CX phone system. None of it was ever used.
Nobody knew which calls became appointments
The practice management system records bookings but not where they came from. A call that ended in an appointment and a call that ended in a hang-up looked identical in every report the group had. Conversion from enquiry to booking was a matter of opinion.
Advertising was being judged on call volume alone
Seven numbers across seven websites and campaigns generated a call count and nothing more. There was no way to tell a campaign that delivered patients from one that delivered wrong numbers, people wanting NHS treatment, or callers who were simply too far away.
Enquiries were being lost quietly
Callers with a genuine need regularly ended the call without an appointment and without any follow-up arranged. Nobody was counting them, nobody held a list, and nothing triggered a call back. The revenue leak was invisible because the evidence was deleted.
Reception performance was managed on impressions
Eight people answered the phones across three sites. With no way to review calls at scale, feedback came from the occasional overheard conversation or a patient complaint, so good handling went unrecognised and weak handling went uncorrected.
Recordings were deleted after twenty days
The phone system holds recordings for twenty days against a fixed storage quota and then removes them. Any dispute, any training example and any pattern older than three weeks was simply gone.
Manual review was never realistic
Six weeks of calls is 171 hours of audio. Listening to it properly would occupy one person for more than four working weeks a month, and would still produce impressions rather than numbers that could be counted, compared and acted on.
An automated pipeline from the phone system to the management report
We built and deployed a system that reads every call the way an attentive manager would, and records the answer as structured data rather than a subjective note. It runs unattended, end to end.
Collect
Connects to the 3CX system and pulls each recording within a minute of the call ending, with the number dialled, the extension that answered, the direction and the duration.
Transcribe
Speech recognition produces a timestamped transcript with the two speakers separated, accurate on compressed phone audio and a wide range of accents.
Understand
A language model reads the transcript against the clinic's own services, branches and prices, and returns structured fields rather than a summary paragraph.
Report
Calls are attributed to the website or campaign that produced them, then rolled into dashboards for reception performance, treatment demand and advertising waste.
Every call classified into one of eight outcomes
Appointment booked, rescheduled or cancelled; information only; wrong branch; wrong number; no conversation; other. The classification describes the result of the call rather than the request, so a caller who asks to book but is told to ring back is never counted as a booking. Managers can correct any classification, and the correction overrides the model everywhere.
Call-to-campaign attribution across seven numbers
Each tracked number is tied to the website or campaign that publishes it, so every outcome rolls up by source. Conversion, wasted calls and missed opportunities can be compared site by site, and the worst-performing number is visible at a glance instead of being averaged away.
Search phrases recovered from the conversation
The system reconstructs the search-style phrases behind each enquiry and separates those that lead to bookings from those that lead to wrong calls. Before a phrase is recommended as a negative keyword it is checked against calls that did convert, so a term that also brings patients is never excluded.
Missed opportunities tracked per patient, not per call
A caller with real intent who leaves without a booking and without a follow-up is flagged. Repeat calls from the same number count once, and the flag clears automatically if that person books later, so the follow-up list is a list of people to ring rather than a pile of duplicate entries.
A reception score out of ten, with evidence
Each conversation is graded against what the caller wanted rather than the outcome alone, so a booking that happened only because the caller was insistent scores lower than an unbooked call handled well. Every score carries a written reason, the transcript lines that justify it, and a single coaching point, which makes it reviewable instead of something to be taken on trust.
Recordings available beyond the PBX retention window
Audio is held locally for seven days and then streamed on demand from the phone system, so calls stay playable while storage stays flat. Transcripts, scores and reports are kept permanently, which means the analysis survives long after the audio is gone.
Runs unattended, with no staff listening time
The pipeline operates as a managed background service that starts itself, restarts after failure and resumes from the last synchronised call after any interruption, backfilling whatever it missed. No one schedules it, checks it or feeds it.
Access control and an audit trail for clinical data
Named staff accounts carry separate add, edit and delete permissions, with settings and user management restricted to a super administrator. Every sign-in, session and change is recorded in an audit log that cannot be edited from the interface.
What the practice manager sees every day
* Interface screenshots use anonymized sample data. Phone numbers are from Ofcom's reserved test ranges.
One conversation, converted into fields a manager can act on
Illustrative example. Real transcripts contain patient medical records and remain strictly inside the clinic perimeter.
- Outcome
- Appointment booked
- Treatment
- Emergency dental
- Patient Intent
- Same-day appointment for a chipped tooth and pain
- Urgency
- High
- Appointment
- Today, 11:30
- Price Discussed
- £30 examination
- Reception Score
- 9 / 10
- Source
- Emergency dentistry campaign number
Offered two specific slots, answered the price question directly, confirmed arrangements clearly.
Six weeks of calls, classified across 5,446 conversations
Of 3,006 genuine patient enquiries, 608 became appointments—a conversion rate of 20.2%. Before this system existed, none of these figures could be produced at all.
Three things the clinic group could not previously see
16.6%
of calls were wasted
905 calls were wrong numbers, people wanting a different branch, or requests for services the group does not provide. Where those calls came from paid search, every one was a click already paid for.
399
patients worth chasing
Callers with a real need who left without booking and without a follow-up arranged. Counted once per person rather than per call, and removed automatically from the list if they book later.
19.8%
of calls had no conversation
Voicemail, hang-ups and silence. A measurable answer-rate problem that nobody had quantified, concentrated in identifiable hours of the week.
Advertising Waste, Named & Actioned Precisely
Because every call carries the number it arrived on, wasted calls can be traced back to the website and campaign that produced them. The system extracted 4,683 distinct search-style phrases.
| Service Requested (Not Provided) | Wasted Calls | Immediate Advertising Action |
|---|---|---|
| NHS dental treatment | 77 | Exclude NHS terms; the group is 100% private |
| GP appointment | 30 | Exclude general practice and GP terms |
| Sexual health treatment | 24 | Keep testing terms, exclude treatment terms |
| Paediatric blood test | 17 | Exclude children's testing terms |
| HPV testing | 13 | Exclude terms or introduce the service |
Every call graded against what the caller wanted
Graded out of ten against caller intent rather than raw outcome alone. Every score carries a written justification and exact transcript citations for constructive coaching.
| Grade Range | Share of Calls | Operational Meaning |
|---|---|---|
| 9–10 · Exemplary | 22.4% | Need identified quickly, specific slots offered, details confirmed |
| 7–8 · Good, minor misses | 55.4% | Usually no alternative offered, or details not read back |
| 5–6 · Passive | 18.4% | Outcome driven entirely by the caller, not by reception guidance |
| Below 5 · Needs attention | 3.8% | Clear intent not converted, or questions answered incorrectly |
“The average across six weeks was 7.5 out of 10. The value is not the average: it is the 131 specific calls below 5, each one playable, with a transcript and a one-line coaching point.”
Which treatments people ask for, and which convert
Enquiries normalized to standard clinical categories. The gap between inquiries and bookings pinpoints where pricing, availability, and reception training generate the greatest revenue gain.
| Treatment | Enquiries | Booked | Conversion |
|---|---|---|---|
| Hygiene | 475 | 137 | 28.8% |
| Blood tests | 446 | 70 | 15.7% |
| Emergency dental | 242 | 112 | 46.3% |
| Sexual health testing | 221 | 49 | 22.2% |
| Dental examination | 191 | 70 | 36.6% |
| Teeth whitening | 81 | 15 | 18.5% |
Built to run unattended, with zero maintenance overhead
The pipeline operates as a managed background service that starts itself, restarts after any phone system failure, and resumes from the last synchronized call, backfilling whatever it missed. Half of all calls are fully analyzed within 1.3 minutes of the caller hanging up.
| Operational Measure | Production Result |
|---|---|
| Calls processed in the period | 5,446 calls |
| Coverage of recorded calls | 100% (zero sampling) |
| Median call-end to analysed | 1.3 minutes |
| Staff time spent reviewing calls | Zero hours |
| Equivalent manual listening avoided | 171 hours |
| Cloud processing cost per call | under 3p |
From an invisible channel to the clinic's best-measured one
The group's busiest sales channel was also its least visible. It is now the best measured one. Every call is classified, scored and attributed within minutes, the reception team is coached on real conversations rather than impressions, and advertising decisions are made from what callers actually asked for.