Methodology
What your AI receptionist actually booked
Answering every call is easy to measure. Whether those bookings turn into collected money is a different question, and the one worth asking.
- Cash collected per job booked by a person
- Cash collected per job booked by the agent
- Jobs the agent books per month
Booked is not collected
An AI receptionist reports the metric it can see: calls answered, appointments made, hours covered. Those are real, and they are not the question.
The question is what happens to those appointments afterwards. A booking that cancels, or completes at half the ticket, or never gets invoiced, cost you the call it displaced. The only way to see that is to follow the booked jobs all the way to cash.
Two funnels, one comparison
Jobs are split by who booked them and each side is followed to collected revenue. Then the two are compared as cash per booked job — not booking counts, not completion rates.
Per-job is what makes the comparison fair. An agent that books twice as many jobs at half the realised value has moved nothing, and any metric that counts bookings will call that a success.
Each side needs at least thirty jobs. Below that you are comparing two small samples and calling the difference a finding.
Which jobs are the agent's is your call, not a guess
Nothing in an export says which bookings an AI made. The column holds whatever your shop types into it, and Ava is a voice agent at one company and a scheduler named Ava at the next.
So the product never infers it. You designate which labels in your own data mean the agent, and every job is classified against that list. A blank or unrecognised label stays unknown and is excluded from both sides — an unattributed job is a job we cannot attribute, not a human one.
The designation is applied when the analysis runs rather than when the file is uploaded, so adding a label later reclassifies records already imported. Nobody re-uploads six months of jobs to correct a checkbox.
The result that is worth as much as a finding
If the agent's jobs collect as well as your team's, there is no finding and the detector says so. That is a useful answer — it is the evidence you need to keep paying for the thing, and it came from your own books rather than the vendor's dashboard.
A gap in the other direction is not automatically an argument for switching the agent off, either. It is an argument for looking at which calls it takes, and what it does with the ones that turn into money.