Oracle Health’s clinical AI agent now reads the visit conversation and suggests professional-fee charge codes, with the clinician confirming. Before asking what that will do to the code distribution, it is worth looking at what the distribution has already been doing for twelve years without any help.
Documentation, coding, and chart prep were three separate purchase orders eighteen months ago. Oracle just made them three capabilities of one agent, generally available in the US, and put the coding suggestion in the orders workflow where the clinician signs off before submission.
The obvious worry is that a suggestion engine pointed at charge capture will push levels up. It probably will. But here is the part that makes the worry hard to prove: the levels have been going up the whole time anyway.
Every 99212–99215 billed to Medicare fee-for-service in an office setting, as a share of that year’s total. 2.1 billion visits across twelve years. Nothing in this chart is AI.
In 2024, Medicare allowed $85.43 for a 99213 and $119.77 for a 99214 — a gap of $34.34 per visit. There were 63,712,067 level-3 office visits. Move the slider to shift some of them up one level and watch the national number.
This is arithmetic, not a prediction. Nobody has published a confirm-rate for Oracle’s suggestions, and the suggestion is not the billing — the clinician confirming it is. But the arithmetic is the reason the feature exists, and it is the number a health system CFO is running right now.
Sixty-five specialties, each drawn as one hundred percent of its own established office visits, sorted by how much of that sits at level 4 or above. This is the surface a coding agent gets pointed at — and the headroom is very unevenly distributed.
Claims data records what was billed. It cannot tell you whether the visit deserved it. A 15-point shift toward level 4 is equally consistent with a sicker, older, more comorbid Medicare population, with better documentation of care that was always being delivered, and with drift. Anyone who shows you this chart and says “upcoding” has added a claim the data does not carry.
Push Min. clinicians up and the top of the ranking reshuffles. Nuclear Medicine sits at 86.4% level-4-and-above on 43 clinicians; Neuropsychiatry at 82.4% on 108; Geriatric Psychiatry on 94; “Undefined Physician type” on 97. CMS also suppresses any clinician/code/place-of-service cell under 11 beneficiaries, so these specialties are missing their low-volume members entirely and the survivors skew high. Set the floor at 2,000 and what remains is a genuine clinical gradient — cardiology, nephrology, oncology, endocrinology at the top; podiatry, dermatology and optometry at the bottom.
Oracle’s announcement scopes pro-fee code suggestions to ambulatory visits inside its own EHR. This chart is Medicare fee-for-service only, place-of-service office only, and level 2–5 established visits only. It excludes new-patient visits, every commercial and Medicare Advantage encounter, and every health system that does not run Oracle. It is the right shape and the wrong size.
Today’s newsletter makes the point that steps get bundled and judgments don’t. Level selection was, for twelve years, a judgment made by a clinician under a documentation rule. Oracle is proposing it as a step. The chart above is what the judgment looked like. Whether the next twelve years look different is the only real test of the distinction.
CMS Medicare Physician & Other Practitioners PUF, performance years 2013–2024, HCPCS 99212–99215, place of service = office. Shares are computed on summed Tot_Srvcs. Dollar figures are Avg_Mdcr_Alowd_Amt weighted by service count for PY2024 — Medicare allowed amounts, not submitted charges and not commercial rates. The specialty panel covers the 65 specialties with at least 25 billing clinicians and at least 20,000 established office visits, which is 172,923,678 of the year’s 173,015,417. Specialty is CMS’s Rndrng_Prvdr_Type, derived from the claim’s specialty code.