Every federally funded health center in America, sorted by the share of patients best served in a language other than English — against the nine settings where the ambient-scribe evidence was actually generated.
Line up the settings of every major ambient-scribe study and you get UCLA, Mass General Brigham, Emory, UCSF, Yale, UC Davis, Kaiser Northern California, Penn, Stanford. Nine names. Line up the clinics where documentation burden actually peaks — long visits, three languages, heavy social complexity — and you get 1,352 organizations that appear in none of them.
The graphic below plays itself. It has one job: put both counts on the same page, at the same scale, once.
8,977,620 patients — 27.8% of the 32.3 million people these centers see — are recorded by HRSA as best served in a language other than English. In 190 of these grantees, covering 4.9 million patients, that share is above half. Every dot in the tall left-hand stack is a center where nearly everyone is served in English; every dot out on the right is a center where nearly nobody is.
“A gap in evidence is not a verdict. It is a to-do list.” — Dr. Gigi Magan
The demographic gradient in the underlying technology is not speculative. The 2020 PNAS audit of five commercial speech recognition systems — Amazon, Apple, Google, IBM, Microsoft — measured word error rates against matched interview audio:
Nearly double, traced to the acoustic model and thin training data. Vendors have invested heavily since. Nobody has published the check in any of the 1,352 settings above.
If you hold a procurement pen, add one line to your next ambient-documentation RFP: stratified accuracy data by preferred language and demographic group, from a deployment setting comparable to ours, before contract. You will probably not get it. The answer you get back — and how fast — tells you more about the vendor than the demo did.
One dot is one grantee, not one clinic and not one patient. A single grantee can run dozens of sites; the largest here serves 295,386 people and the smallest 134. The tall left stack is visually dominated by small rural centers, so the field over-represents organizations and under-represents patients. The 499 grantees to the right of the 25% line carry 16.4 million patients — half the national panel — from a third of the dots.
“Best served in another language” is a reporting field, not an audio measurement. It is a self-reported UDS count. It says nothing about accent, dialect, or code-switching in English — which is exactly where a scribe pipeline degrades. A center sitting at 5% can still run most of its visits in accented English.
The nine is a hand-assembled absence. It is the set of study settings Magan located after checking 41 references against primary sources. The honest claim is “none found in a federally qualified or community health center,” not “none exist.” If you know of one, that is the most useful reply this piece could get.
Vintage caveat. HRSA's UDS tables carry no performance-year column; 2024-12-31 is the file publication date used as the year proxy. Seven grantees report no language field and are excluded — hence 1,352 rather than 1,359.
Cardiology has already built the machinery this field is missing. The AHA's AI Assessment Lab ran Ultromics' EchoGo Heart Failure against roughly 90,000 real-world echocardiograms and published subgroup findings by race and age alongside the accuracy numbers. Ambient documentation is deployed far more widely than that algorithm and has no equivalent body doing that work.
So the assessment layer for the most-deployed clinical AI in America is vacant, and the entry requirement is a QI dashboard, not an R01. Decline rates by language. Note quality by population. Edit burden on interpreter-mediated visits. None of that needs a grant.