Measure recoverable work, not ringing events. Before building an AI answering service for plumbers, an HVAC receptionist or a call overflow agent, determine what callers were trying to accomplish and what happened afterwards.
Collect a baseline
As a starting research design, request four representative weeks of aggregate call data and review a stratified sample of approximately 200–500 interactions where lawful and operationally feasible. These are suggested starting quantities, not statistically sufficient samples for every business.
Include:
- Calls offered, answered, abandoned and sent to voicemail.
- Opening-hours versus after-hours demand.
- New enquiries versus existing-customer requests.
- Repeat calls concerning the same task.
- Service-area, appointment-type or customer eligibility.
- Completed bookings, fulfilled jobs, cancellations and unsuccessful attempts.
- Existing callback recovery.
- Human handling, supervision and correction time.
- Complaints, escalations and failed transfers.
Use a task identifier where possible. Do not assume a telephone number uniquely identifies a person or that two calls from it concern the same task.
A useful baseline export schema is:
call_id, task_id, timestamp, location, routing_group,
call_reason, eligibility, answered_status, handling_minutes,
transfer_result, operational_record_id, final_task_status,
human_followup_minutes, repeat_contact_flag
Prefer aggregate or redacted material for early scoping. Obtain approval before copying recordings or identifiable records into a new tool.
Build a missed-call recovery waterfall
Use this sequence:
Missed calls → distinct enquiries → eligible enquiries → demand not already recovered → accepted bookings → fulfilled jobs.
Apply available capacity before assigning value. If the business cannot fulfil additional work, do not count all eligible enquiries as recoverable contribution.
Compare four operational options on the same workload:
- Better routing and callback ownership.
- Additional staff coverage.
- A human or hybrid answering service.
- The proposed AI workflow, including human rescue.
Deliverable: a baseline workbook with each input labelled extracted, provided, assumed or unknown. Keep revenue attribution assumptions separate from observed call outcomes.