Health system CIOs at Epic’s UGM reached for the same new word: tokenomics. Nobody can forecast the bill because the per-feature cost dashboard is still being built. But only one side of that multiplication is secret. The other side — volume — has been public federal data all along.
Ochsner Health CIO Amy Trainor, BSN, RN, told Becker’s that Epic’s own AI features are billed inconsistently — some flat, some consumption-based — and that the dashboard which would show what each feature actually costs has only recently started being built.
Four orders of magnitude is not a forecast. It is the absence of one. And the reason the range is that wide is that the bill is a product of two numbers: a price nobody has published, and a volume everybody already knows.
Here is the volume. Every ambulatory office visit billed to Medicare fee-for-service in 2024, by specialty — the exact encounters an ambient scribe, a chart-prep agent or a coding suggestion would fire on.
Two dials decide everything. Price per AI call is what a vendor charges for one model invocation — a note draft, a summary, an agent turn. Calls per encounter is the one nobody budgets for: a single visit does not trigger one call, it triggers a scribe pass, a re-draft, a chart summary, a coding suggestion, and whatever the clinician re-runs because the first output was wrong.
Each circle below is one specialty. Horizontal position is office visits per clinician per year. Circle area is how many clinicians are in that specialty. Move the dials and watch which specialties cross the $10,000 line.
The dial that moves the bill most is not the price. It is calls per encounter — and unlike price, it is not set by your vendor. It is set by how much your users like the thing. A tool nobody uses costs $10. The same tool, loved, costs $10,000. Every consumption contract is a bet that your best user stays average.
Today’s newsletter suggests bringing one slide with three columns — expected volume, 3×, 10× — and your cost at each. Here it is, drawn from real volumes, for the twelve busiest specialties. The dot is expected usage; the line runs out to 10×. The vertical rules are Trainor’s two numbers.
A benchmark you can move with a slider deserves to be interrogated, so here are the three places it breaks.
CMS suppresses any clinician/code/place-of-service cell with fewer than 11 beneficiaries. Low-volume specialties therefore lose their low-volume clinicians entirely, and the survivors inflate the average. Push Min. clinicians above 1,000 and several apparent outliers vanish: Nuclear Medicine (470 visits/clinician, n=50), Undersea and Hyperbaric Medicine (323, n=31), Peripheral Vascular Disease (341, n=44), Pathology (363, n=79). None of those are dense-visit specialties. They are small-n artifacts wearing the costume of one.
Nothing in this chart tells you what Epic charges. Nobody knows — that is the entire premise of the Becker’s piece. The price slider is a your-number-here field, not a measurement. Anyone publishing a confident total AI spend forecast right now is multiplying a real number by an invented one and reporting the product to two decimal places.
Switch the y-axis to % of the visit fee and the scatter flattens into a rank order, because the meter charges the same whether the encounter pays $88 or $169. The specialties eating the largest share are the lowest-paid ones: Nurse Practitioner ($93.27/visit), Physician Assistant ($90.91), Podiatry ($95.71). A flat per-call price is a regressive tax on the cheapest visit — which is to say, on primary care.
Medicare fee-for-service, performance year 2024, office place-of-service only, HCPCS 99202–99215. 73 specialties with at least 25 billing clinicians. Dollar figures are Medicare allowed amounts, weighted by service count — not submitted charges, not commercial rates. Specialties are CMS’s own Rndrng_Prvdr_Type field, which is derived from the specialty code on the claim and is not always what the clinician would call themselves. The cost model is straight arithmetic — visits × calls × price — with no volume discount, no cached-response tier, and no flat-fee floor, all three of which real contracts have.