clinicians.build · interactive · september 5, 2026

The Ratchet Already Ran

Every ambient scribe pitch promises richer documentation and the revenue that follows. Before you price that, look at what happened to Medicare office visits between 2013 and 2024, with no ambient AI in the room at all. The level‑4 share went from 40.8% to 54.6%. Press play, then add the scribe on top.

Primary source: Becker’s Hospital Review — Simon Mathews, MD, “The productivity ratchet: what are we actually buying when we buy AI scribes?”, Sept 3, 2026
Data via MIMI Labs: CMS Medicare Physician & Other Practitioners, by Provider and Service
CPT 99212–99215, office place of service · 12 annual vintages · 2.24 billion visits

Each band below is a share of all established-patient office visits billed to Medicare fee‑for‑service that year. Real CMS data through 2024. Then drag the scribe slider and the chart extends into a hypothetical 2025–2028 in which an ambient tool moves that share of 99213s up to 99214, and the same share of 99214s up to 99215.

scrub the year — then add a scribe
2024
off
2013–2024 is CMS data. 2025–2028 is a model, not a measurement.
level 4 share
99214, of all office E/M
level 3 share
99213 — the shrinking stock
blended $/visit
at 2024 allowed amounts
vs 2013 mix
per clinician/yr, 372 visits
99215 — level 5 99214 — level 4 99213 — level 3 99212 — level 2

Eleven years, one level.

In 2013, Medicare paid for 92.8 million level‑3 office visits and 79.3 million level‑4s. In 2024 those numbers are 63.7 million and 94.4 million. The level‑2 visit has nearly disappeared — 7.6% of the mix down to 3.1%. Total visit volume is lower than 2013. The mix is what moved.

Some of that is real: sicker panels, an older population, the 2021 E/M documentation overhaul that re-based the codes on medical decision-making. Some of it is that documentation improved. And some of it is that every incentive in a fee‑for‑service system points the same direction, and coding drifts that way on its own, slowly, without anybody deciding to do it.

The claim and the measurement, side by side

Three numbers get quoted in scribe business cases. Here is each one next to what has actually been measured.

Only the third pair goes the buyer’s way. Burnout is the one thing the evidence overdelivers on — 51.9% to 38.8% within a month of adoption in a six-system study — and it is the one thing that does not appear on a P&L. The two that finance underwrites, revenue and time, both come in at roughly a quarter of the pitch.

Where the difference gets found

A system that modelled $13,000 per clinician and is collecting $3,044 has a hole. The hole is not big. Across the 42 highest-volume Medicare specialties, the shortfall works out to one to two additional visits a week — a 20-minute slot that becomes 15, two add-ons on a Thursday. It is small enough that it never has to be announced, which is exactly why it is the likeliest way it gets closed.

We have run this experiment before. Human scribes spent twenty years inside fee‑for‑service and a meta-analysis of 39 studies found throughput up about 0.3 patients per hour and RVUs up 0.55 per hour. The saved minutes did not stay with the clinician; they became volume, because volume was what got paid. That is Jevons: today’s saved thirty minutes is the baseline tomorrow’s schedule is built on.

The number nobody has scheduled is burnout the quarter after the productivity targets get re-based to the post-scribe schedule.

What this chart is not telling you

Push the scribe slider past about 20% and stop trusting it. The model shifts a fixed share of level 3s up to level 4 every year, but the stock of level 3s is finite and already shrinking — 36.8% of the mix and falling. By the modelled 2028 at a 30% rate the level‑3 band nearly vanishes, which is not a forecast, it is what happens when you apply a percentage repeatedly to a depleting pool. The real ceiling is lower and arrives sooner. That is the single most important thing this graphic gets wrong on purpose: up-coding is a stock, not a flow, and a scribe spends it once.

The 2021 discontinuity is doing work you can see. Watch the level‑3 band between 2020 and 2021 — it drops 2.5 points in a single year. That is the CMS office E/M overhaul taking effect, not clinician behaviour. Any trend line drawn straight through 2013–2024 is averaging across a rule change, and any vendor drawing one is telling you a story about behaviour that is partly a story about regulation. The 2020 dip in total volume is COVID, for the same reason.

And the whole thing is Medicare fee‑for‑service only — no Medicare Advantage, which is now about half of all beneficiaries, no commercial, no Medicaid, office place of service only, established patients only. Codes get re-valued: the same 99214 was $102.91 in 2013 and $119.77 in 2024 in allowed dollars, so every dollar figure here holds prices at 2024 to isolate the mix effect. If you want the specialty-by-specialty version with the arithmetic exposed, that is the companion piece.