Federal health officials are weighing a Medicare benefit category for AI that supports diagnosis and care. Per the Times, the live internal number is 60% to 80% of what a human is paid for the same service. Nothing has been proposed. But the fee schedule that number would land on already exists — so here it is. Set the rate. Pick which services an AI may bill. Watch the whole schedule sag.
A payment rate is not an opinion. It is an instruction that every staffing model, every build-vs-buy memo and every residency match eventually obeys. So the useful question about “60% to 80%” is not whether it is fair. It is: 60% of what, exactly?
Medicare Part B paid out $103.12 billion across 5,955 non-drug service codes in 2024. The 300 largest carry $88.87 billion of it — 86.2% — and every one of them is a dot below.
Forty-four of those 300 codes are evaluation and management. They are $41.50 billion of the money. That is what “the same service” mostly means.
Each dot is one HCPCS code. Horizontal position is how many times it was delivered in 2024; vertical position is the average Medicare allowed amount for one of them; area is total dollars. Codes you mark as AI-billable get a red drop-line down to what the same service would pay at the rate you set. Everything else stays where it is.
Every code is attributed here to the specialty that billed the largest share of its dollars. That is a crude attribution — a 99214 is billed by nearly half a million providers across every specialty in medicine — but it answers the question people actually ask, which is whose income the dial is attached to.
The phrase is carrying the entire policy. A 99214 is not a deliverable; it is a bundle — a history, an exam, medical decision-making of moderate complexity, and a signature attesting that a licensed human did all three and will answer for it. There is no code in the fee schedule for “the thinking happened, nobody is liable.” Pricing an AI at 70% of a 99214 quietly asserts that the bundle decomposes and that the missing 30% is exactly the part a machine doesn’t supply. Nobody has shown that.
Before deciding whether AI belongs in this table, look at what is already in it. Four of these 300 codes are skin-substitute Q-codes — Q4275, Q4331, Q4277, Q4250 — billed by 46, 9, 24 and 29 providers respectively for a combined $881.6 million. Per-square-centimeter products with no professional work component, carrying more money between them than all screening and diagnostic mammography in this table combined ($634.0M). The argument that a rate must reflect cognitive work is one this schedule has already lost. The companion piece takes that apart →
Find the codes you or your product touch, note the current allowed amount, and write down what your system would have to attest to in order to bill one of them. Who signs. What artifact records that a human reviewed it. What happens when neither of them caught it. That attestation layer is the part nobody at CMS will define for you, and it is the part that will decide whether the rate is 60% or 5%.
Also today: Two Targets, One Policy → — Medicare’s new AI prior-auth pilot points the same tool at skin substitutes and epidural injections. One chart shows they are not the same problem.