Call-recording scorecard + weekly coaching summary

The problem

A call that's answered but not booked costs the same marketing money as a booked one. We don't attach a dollar figure.

• Only 45% of answered home-services calls asked for the appointment. Source: Invoca, Home Services Lead Conversion Benchmarks Report 2026 (Industry report by a call-tracking vendor)
• HVAC booked 38% of calls and plumbing 43% in ServiceTitan's data. Source: ServiceTitan booking-rate data, as summarized by Skygnosis (Secondary write-up of ServiceTitan data)

Background reading: How many calls is your HVAC shop missing, and what it costs

Who it's for

Shops with 1+ people answering phones (CSRs, the owner's spouse, office manager) that already record calls in Twilio, or can export recordings from CallRail or ServiceTitan.

Tools and monthly cost

Example month: Illustrative arithmetic. Assume 400 calls of 3 minutes: recording 1,200 min x $0.0025 = $3.00, storage about $0.60 a month (growing unless you delete old recordings), speech-to-text $3.60-$7.20, AI scoring about $0.10-$0.20. Total about $7-$11 a month plus n8n. (Twilio's own transcription would be $0.05/min = $60.)

Setup

1. Turn on recording for inbound calls (e.g. record on the <Dial>) and set recordingStatusCallback to the workflow's recording webhook, adding ?agent=NAME per line or queue so calls map to a person.
2. Play a recording notice at the start of calls. Some US states and Canadian provinces require all-party consent.
3. Set TWILIO_AUTH_TOKEN and PUBLIC_WEBHOOK_BASE so the workflow can verify Twilio's signature, plus your speech-to-text and model API keys.
4. Import workflow-score-call.json and workflow-weekly-summary.json. Replace the mock 'sheet' URL with a Google Sheets node (or any table) for the call-scores log.
5. Score a week of real calls, read 10 of them yourself, and only then share the weekly summary with the team.

Download

Download call-scorecard.zip

Contains the n8n workflows, AI prompts, a Claude skill version (documentation only, not executed in our tests), the mock data, the test runner with its last results, and the mock server we tested against. All keys in .env.example are fake.

Test result

Every workflow was imported into n8n 2.45.4 and run end to end on our test machine. Twilio, Jobber and Stripe were replaced by a local mock server that accepts the same request shapes as their documented APIs and records every call. The AI steps used a small local model (Qwen2.5-3B-Instruct, 4-bit) and the speech-to-text steps used Whisper base.en running locally. Mock data (customers, quotes, invoices, recordings) is invented. Nothing was sent to a real phone, customer or account.

Actual output from the test run

Score row written for call C2. Names and numbers are invented.

{
  "date": "2026-10-21",
  "agent": "Tyler",
  "call": "CAc2c2c2c2",
  "seconds": 21,
  "problem": "\"water heater leaking\"",
  "greeting": true,
  "greeting_evidence": "ridgeline this is tyler hi how much do",
  "got_name": false,
  "got_name_evidence": "",
  "got_address": false,
  "got_address_evidence": "",
  "got_problem": true,
  "got_problem_evidence": "\"water heater leaking\"",
  "offered_time": false,
  "offered_time_evidence": "",
  "asked_for_booking": false,
  "asked_for_booking_evidence": "",
  "outcome": "not_booked",
  "score": "2/6",
  "unverified": "",
  "tip": "It's better to ask about the service call fee after identifying the problem and the customer's interest in the service."
}