AI Receptionist ROI: The Worked Math for a Medical Practice

Every vendor selling an AI receptionist also sells a calculator that proves you should buy one. One page-1 result puts the return at $100,000 to $200,000 or more a year for a mid-size practice. Another claims its voice AI books 40% of the callers who would otherwise have hung up. A third promises most clinics save $9,000 or more a month. All three numbers come from vendor pages, and none of them show the arithmetic.

This post shows the arithmetic. It walks one transparent model from six inputs, call volume, missed-call rate, bookable share, recovery rate, show rate, and visit value, through recovered revenue, subscription cost, and payback period. Every figure is labeled Modeled or tied to a named published source. Cart Gaze has no published client data, so nothing here is dressed up as a client result.

By the end you can rebuild the whole thing in a spreadsheet in about 15 minutes. Just as important, you will know the three situations where the honest answer is no, because a model that cannot say no is a brochure.

Why vendor ROI calculators flatter the answer

Vendor ROI calculators overstate the return because the assumptions that do the work are locked where you cannot edit them. The common moves: value each recovered call at patient lifetime value, count every answered call as a new booking, and apply no show-rate discount. A visible model beats a flattering one.

The page-1 results for this search are calculators from MedReception, PatientXpress, AppointFlow, Hookneural, OmniMD, and half a dozen others. Each is a lead form shaped like a spreadsheet: you type your call volume, it outputs a large number and a demo button. The inputs you can edit are the harmless ones. The inputs that decide the answer, the capture rate and the value per patient, are fixed by the vendor.

Three moves inflate the output. First, lifetime value substitution: valuing each recovered call at $2,000 to $6,500 of future lifetime revenue instead of the single visit it actually books. Second, no incrementality: assuming every call the AI answers is a booking you would have lost, when a share of missed callers simply call back later. Third, no show-rate discount: a booking that never walks in gets counted at full value.

Even the baseline statistics disagree with each other. Talkdesk's 2025 report puts unanswered calls at 23% of volume, while a study of 7,000 calls cited by AnswerNet puts the miss rate at 42%. Both get quoted as fact, which tells you neither is your number. If you are still working out what these tools do before pricing one, start with what an AI receptionist actually does and where it stops, then come back to the math.

The six inputs that decide AI receptionist ROI

Six numbers decide the return: monthly call volume, missed-call rate, the share of missed calls that wanted to book, an honest recovery rate, your show rate, and average visit value. Five of the six come from your phone log and practice management system. Only the recovery rate is a true assumption.

Published ranges, each sourced. Anchors for your own measurement, never your result.

Call volume and missed-call rate come straight out of your phone or VoIP system: total inbound calls last month, and missed plus abandoned divided by that total. Most cloud phone systems export both in a few clicks. This is the same export used in the missed-call cost worksheet, and your own export beats every industry average, because the published miss rates disagree by almost 2x.

The bookable share takes 20 minutes of tagging. Pull 30 to 50 missed calls and sort them into two buckets: wanted to book or was a new patient, versus billing, refills, and results. Vendor calculators quietly count all missed calls as lost patients; your tagged sample will usually say otherwise.

Show rate and average visit value live in your practice management system. Published no-show rates run 5% to 30% depending on specialty, which is exactly why you should use yours instead of an average. The other published anchors are sanity checks, not inputs: $125 to $200 of revenue lost per missed patient call is a widely cited range attributed to HFMA, and the claim that 85% of callers who reach voicemail never call back traces to PATLive. Verify both before you repeat them.

The worked model, line by line

The model multiplies six lines: 800 calls a month, a 20% miss rate, a 50% bookable share, a 25% recovery rate, an 80% show rate, and $150 per visit. The result is 16 kept visits and about $2,400 a month in recovered revenue. Every figure is modeled, so swap in your own.

Walk the discounts in order. Of 800 calls, a 20% miss rate leaves 160 missed calls, deliberately below the published 23% to 42% disagreement so the model stays conservative. Half of those, 80 calls, wanted an appointment. The rest were billing, refills, and results: calls worth answering, but not lost revenue.

The 25% recovery rate deserves the most scrutiny, because this is where vendor calculators claim 40% and up. If 85% of voicemail callers never call back, roughly 15% do come back on their own, and those bookings are not incremental. The model prices recovery net of them, and below a typical daytime desk conversion. If 25% still feels optimistic for your patient base, cut it and rerun; the spreadsheet is yours.

Each step shrinks the number. A model that skips a step is flattering you.

Then the show rate: 20 bookings at an 80% show rate is 16 kept visits, and 16 visits at $150 each is $2,400 a month of modeled recovered revenue. Notice what the model refuses to do. It does not touch lifetime value, and it does not count saved staff time. Both are real, and both are exactly where flattering math likes to hide.

Payback period and break-even, the two honest numbers

Judge the purchase on two numbers: break-even and payback. At a modeled $400 a month all-in, break-even is about 3 kept visits, since $400 divided by $150 per visit is 2.7. Payback on a $1,500 setup fee arrives in the first month if the model holds.

Start with the real cost side. Flat small-practice plans run $149 to $299 a month, but the all-in figure for a healthcare line typically lands around $300 to $700 once you count the tier where the vendor signs a BAA, which can add 10% to 30%, plus integrations. Setup fees run $500 to $3,500. The full breakdown is in what an AI receptionist really costs; this model uses $400 a month plus $1,500 setup.

Year 1, modeled: $400 x 12 plus $1,500 is $6,300 of cost, against $28,800 of recovered revenue, for a net of roughly $2,000 a month. At that pace the setup fee pays back in under a month. Those are the model's numbers, not a forecast of yours: change one input and the answer moves, which is the whole point of owning the spreadsheet.

Break-even is the number worth memorizing, because it is small enough to check by hand. At $400 a month and $150 per kept visit, the system has to produce about 3 kept visits a month to cover itself. You can audit that in your own schedule after 60 days without trusting anyone's dashboard.

Three cases where the ROI math fails

The model collapses in three situations: call volume too low for recovered revenue to clear the subscription, show rates weak enough to shrink every booking, and a schedule too full to absorb new patients. Each one is checkable before you sign, and each is cheaper to find now than later.

Modeled, other inputs fixed. The subscription is flat at $400, so volume decides the return.

Low call volume is the clearest failure. Run the same model at 150 calls a month and recovered revenue is about $450 against a $400 subscription: roughly $50 of margin, which one soft input erases. At 400 calls the model recovers about $1,200 and the case is real. Below roughly 200 to 300 calls a month, a per-minute plan or a simple missed-call text-back usually beats a flat subscription.

Weak show rates cut the same way. Drop the show rate from 80% to 60% and the same 20 bookings are worth $1,800, not $2,400. A specialty with a no-show rate near 30% should model the pessimistic case before the hopeful one, and remember that a caller recovered at 9 p.m. is not automatically as committed as one who called your desk twice.

No capacity is the failure nobody models. If your schedule is booked out for six weeks, a recovered booking mostly displaces another one, and the revenue line quietly goes to zero. The honest gain becomes staff time and after-hours coverage, which is real but smaller and harder to measure. Buying for that reason can still make sense, but price it as that reason, not as new revenue.

Run the model on your own numbers

Pull last month's call log, compute your six inputs, and run the multiplication twice: once conservative, once best-case. Pilot for 60 to 90 days only if both runs clear the all-in cost, then re-measure recovered bookings in your own schedule. If neither run clears it, keep the money.

Answer the questions in order. The cheap exits come before the subscription.

The flow asks the questions in the order that saves you money. If modeled recovery cannot at least double the all-in cost, hold off; the margin is too thin to survive your first surprise. If the money clears but your schedule cannot absorb new patients within 2 weeks, spend on capacity before automation. If patient health details will flow through the calls, shortlist only vendors that sign a BAA, and price that tier rather than the demo tier.

Whatever the flow says, treat the first 60 to 90 days as the real test: pilot on one call type, read the transcripts, and count recovered bookings in your own schedule rather than a vendor dashboard. If you would rather have the model run for you first, the free Growth Leak Audit does this exact math per practice, from your own numbers, before anyone talks tools.