CMS launched WISeR — AI-assisted prior authorization — and per STAT it went live on a rushed timeline and delayed care. The pilot points the same tool at two service families: skin substitutes and epidural injections. Here is every Medicare provider who billed either one in 2024, on one axis. They are not the same problem, and one review threshold cannot serve both.
Prior authorization is a threshold policy. Somebody sets a line; requests above it get looked at. That works when the thing you are policing has a shape — a cluster of ordinary billers and a tail of unusual ones. Point the same line at two families with different shapes and you get two different programs wearing one name.
In 2024 Medicare Part B paid $4.72 billion for skin and tissue substitutes across 2,902 billing providers, and $918 million for epidural and facet injections across 11,041. Five times the money, from a quarter as many people.
The median epidural biller clears $35,136 a year. The median skin-substitute biller clears $1,681. The largest clears $146.6 million.
Every provider in both families, binned by what they billed. The axis is logarithmic because it has to be: the top of the skin-substitute row is five orders of magnitude from the bottom. Both rows are scaled to the share of their own family so the shapes are comparable. Drag the review threshold and everything above it turns red — that is who a prior-auth program at that line would actually be reading.
Look at the top row before you touch anything. It is bimodal: a hump of ordinary practices around $1,000 and a second, entirely separate population out past $1 million. The bottom row has one hump and no second population at all.
The chart opens with the line at $1 million. In skin substitutes that line covers 710 providers — a quarter of them — and 96.3% of the money. In epidurals the same line covers 60 providers and 7.9% of the money. One of those is a program. The other is not running.
Now drag it down to where it would bite in epidurals. To reach half the dollars there you have to review the top 10.4% of 11,041 providers — 1,147 practices, overwhelmingly ambulatory surgery centers doing entirely routine volumes. In skin substitutes, half the dollars sit with 83 providers. That ratio — 1,147 reviews against 83 — is the whole argument.
These are plotted exactly, not binned. The last column is the sum of distinct beneficiaries CMS reports for each of that provider’s codes — an upper bound on how many people they treated.
The fourth row is the one to sit with: $91.8 million in Medicare allowed charges for services delivered to at most 39 beneficiaries. No algorithm is needed to find that. A SUM and an ORDER BY found it in one query against a public file.
The case for AI in prior authorization is that human review does not scale. But in the family where the money is, the signal is so concentrated that the hard part was never detection — 29 NPIs hold 29% of $4.72 billion, and 83 hold half of it. The hard part is the decision: who is willing to deny a wound-care claim, on what documentation, with what appeal exposure. AI moves the cost of the reading, which was not the binding constraint. Meanwhile the epidural half of the same pilot needs to touch thousands of ordinary practices to find anything at all, and every one of those touches is a delay. Same tool, two problems, and it is the right tool for neither.
Q4xxx and A2xxx HCPCS code; “epidural & facet” is 62321–62327, 64479–64495 and 64633–64636. WISeR’s own service list is narrower and is set by CMS, not by me. This is the neighbourhood the policy is pointed at, not the exact addresses.If you are building anything that sits between a clinician and a payer, plot your own target distribution before you scope the model. If it looks like the top row, you have a concentration problem and you need an escalation and appeals workflow, not a classifier. If it looks like the bottom row, you have a volume problem and every false positive is a delayed injection. Which row you are on determines the product, and you can tell from one chart.
Also today: Sixty Cents on the Dollar → — the explorer. Every big Medicare service, re-priced at the 60–80% AI rate under discussion inside HHS.