What a Missed Call Actually Costs You

Every appointment business loses money to calls nobody answers. The arithmetic, what an AI receptionist does and doesn't fix, and how to check it against your own numbers.

The number nobody measures

Almost every business that takes bookings by phone can tell you how many appointments it made last month. Very few can tell you how many calls it failed to answer, because an abandoned call leaves no trace in the system that matters. It is not a lost sale in the CRM. It is not a complaint. It is simply someone who rang, waited, hung up, and rang somebody else.

Your phone system knows. Almost nobody looks. That is the entire opportunity, and it is why the first thing worth doing costs nothing: pull the call log.

Working out what it costs you

Four numbers, three of which you already have. Take a clinic, a garage, a law firm, a salon, anything where a booking is the transaction.

Input Where to get it Worked example
Inbound calls per month Phone system report 2,000
Abandonment rate Same report. This is the one that surprises people. 20%, so 400 calls
Share of callers who would have booked Your answered-call conversion rate is a fair proxy 30%, so 120 bookings
Value of a booking Average transaction, or lifetime value if you want the honest number $150

Two things people get wrong here. They use average transaction value when the honest figure is lifetime value, because a patient or client you lose on Tuesday does not come back on Thursday. And they forget after-hours entirely. A business open forty hours a week is unreachable for the other 128, and people book when they remember to, which is usually in the evening.

What automation actually fixes

A voice agent is worth building for exactly one reason: it removes the queue. Not by being clever, but by being parallel. Ten people calling at once are ten conversations, not a queue nine deep, and it does not matter whether it is 9am on Monday or 11pm on Sunday.

The second-order effect is usually bigger than the bookings. Your front desk stops spending its day on "what time is my appointment" and starts spending it on the calls that actually need a person: the upset customer, the complicated reschedule, the sale worth closing properly. You did not reduce headcount. You changed what the headcount does.

What it doesn't

It will not fix a business whose problem is that nobody is calling. If your abandonment rate is 3%, an agent will save you a rounding error and you should spend the money on marketing instead. We have told people this and it is usually the most useful thing we say to them.

It will not sell. It qualifies, books, answers factual questions, and hands over. Anything involving negotiation, judgement, or a distressed customer should reach a person quickly, and a well-built agent is aggressive about escalating rather than proud of handling everything.

And it will not survive a bad handover. If the transfer drops the caller into a fresh queue and makes them repeat themselves, you have added a step to a process that was already annoying. The transcript and the record have to move with the call.

What building one involves

The model is the easy part. Three things are not.

  1. Latency. Past about a second of silence, a caller stops believing they are in a conversation. That one constraint drives most of the engineering: transcription that starts mid-sentence, lookups that run speculatively, speech that begins on the first clause instead of the last.
  2. Your booking rules. Which staff can do which job, how long it takes, which location, what has to be blocked out and why. This is where the real complexity lives, and it must not be left to a language model to infer. The model understands the caller; a deterministic rule engine decides what is bookable.
  3. Where it runs. If you are in healthcare, finance, or anywhere with a data-residency obligation, the default answer of streaming customer audio to a third-party API is a problem. Models can run on your own hardware, inside your own network. That is an architecture decision made at the start, not a setting toggled at the end.

We wrote up how those three played out on a real deployment, inside a 26-provider dental group's own EMR, in the Ormond Dental case study.

How to decide in an afternoon

Pull last month's call log. Find the abandonment rate and the after-hours volume. Multiply through with your own conversion rate and your own booking value. If the annual number is not comfortably larger than a year of build and running cost, do not do it, and be suspicious of anyone who tells you otherwise.

If it is larger, the next question is not which vendor. It is whether your booking rules can be expressed precisely enough for software to follow them, because that, rather than the AI, is what determines whether the thing works.