What an AI receptionist actually costs — and what it actually saves

Every page selling one of these leads with a percentage. Conversion up forty percent. Missed calls down ninety. No denominator, no methodology, no name attached.

I'd rather show you the arithmetic and let you run it on your own numbers, including the version where the answer comes out negative and you shouldn't buy anything.

The cost side

There are three real line items and one that's usually hidden.

Build or setup. This is the work of connecting to your scheduling system, writing the scripts, defining escalation rules, and testing it against how your patients actually talk. Priced anywhere from nothing (self-serve products) to five figures (someone doing the integration properly). We charge from $9,800; a self-serve product charges zero and hands you the configuration work.

Monthly usage. Usually per-minute or per-call. The range is wide and depends on volume. Budget somewhere between one and four hundred dollars a month for a typical single-location practice on most platforms, more if your phones are genuinely busy.

Support and tuning. Ours is $1,200 to $3,800 a month depending on volume and how many systems we're watching. On a self-serve product this line is $0 and the work lands on your office manager instead — which is fine if they have the time and the appetite, and expensive if they don't.

The hidden one: your attention during setup. Somewhere between four and fifteen hours of somebody's time, whoever builds it. Scripts have to be approved by someone who knows the practice. This is real cost and it's the line people forget, then resent.

The savings side, honestly

Two categories, and they're not equally reliable.

Category one: revenue you weren't capturing. New-patient calls that went unanswered and became appointments instead. This is the big one and it's real, but it only exists if you were actually missing calls. Run the 4:55 test before you believe any number here.

Category two: labor you no longer need. This is softer than vendors admit. Practices rarely fire anybody. What happens is that the existing front desk stops drowning, and the next hire never becomes necessary. That is real value. It also shows up as a cost you avoided rather than a smaller payroll number this quarter, so be clear with yourself about which one you are buying.

Run it on your numbers

Five inputs. Use conservative figures — if the case only works with optimistic ones, it doesn't work.

• N = missed or abandoned calls per month (from your phone system's report)
• p = share of those that are new patients (if you don't know, use 0.2)
• c = share of those you'd realistically convert to a booking (use 0.4, which is pessimistic)
• V = value of a new patient over two years (from your own software)
• C = total monthly cost of the AI, all lines above, amortizing setup over 24 months

Monthly return = (N × p × c × V) − C

A worked example, deliberately unflattering. A single-location practice missing 60 calls a month; 20% are new patients; convert 40% of those; a new patient is worth $1,200 over two years. That's 60 × 0.2 × 0.4 × $1,200 = $5,760 a month in recovered value against a fully loaded cost of maybe $1,800. Payback lands inside the first two months.

Now the version where it doesn't work. A three-person specialty practice with 6 missed calls a month, mostly existing patients rescheduling, where a new patient is worth $400. That's 6 × 0.1 × 0.4 × $400 = $96 a month against $1,500 of cost. Don't buy it. Buy a better voicemail-to-text setup and spend the money on something else.

Both of those practices exist. The difference between them sits entirely in the inputs. Anyone quoting you a percentage without first asking for your numbers has skipped the arithmetic.

What actually determines whether it works

Cost is rarely the deciding factor. Four other things are.

Can it get into your schedule? Without the ability to write an appointment into your practice management system, you have bought a very expensive answering service. This is the biggest variable in the whole decision, and your software determines it rather than the AI. Some systems have a proper API. Some have a partner program. Some have nothing, and the workaround adds both cost and fragility. Find out which you have before you sign anything.

Is the escalation path right? The moment that decides whether patients trust it is the first genuinely urgent call. It needs to recognize urgency early, stop trying to be helpful, and hand off to a human fast. Test this specifically. Call in and describe an emergency in vague language. See what happens.

Did your staff get told, or asked? Every deployment that fails, fails here. The front desk finds out on go-live day that a robot is doing part of their job, concludes it's a prelude to layoffs, and quietly stops feeding it the information it needs. Bring them in during the build. Start with the tasks they already complain about.

Does anyone read the transcripts? Somebody has to read what it said, every day, for the first month. It will get things wrong. Those mistakes cost ten minutes to fix in week one and are expensive to discover in month six. Whoever sells you this should be doing the reading. If they are not, they have handed you a job without mentioning it.

The version where the honest answer is don't

• Your call volume is genuinely low and your front desk isn't stretched.
• Your practice management system is closed, and the workaround costs more than the problem.
• Nobody internally has any appetite for it. A system your staff resent will be sabotaged, gently and deniably, and it will fail.
• You're about to change practice management software. Wait. Build it against the system you're going to have.

Any one of those and you should keep your money. We say this in audits more often than you'd expect, and it's the cheapest possible outcome for the person paying us.

---

The audit does this arithmetic with your actual numbers and your actual software, and puts the answer in writing — including when the answer is no. $2,500, refunded if it can't find three times that.