GAO published GAO-26-109116, "Science & Tech Spotlight: AI for Medical Notes and Coding," this morning. Two figures from it will be in a vendor's deck by Friday: 95% coding accuracy and $1 million saved. Both are really in the document. Neither is a GAO finding.
Here is the sentence they come from, whole: "According to one AI medical coding software developer, when its software was deployed at a health system with five hospitals, it generated medical codes with more than 95 percent accuracy, and the system's emergency departments reduced annual coding costs by more than $1 million." A developer's claim, about the developer's own software, at one unnamed system. GAO reported it accurately and attributed it plainly. The attribution is what gets dropped in the retelling.
Read the attribution line before you quote the number. It is doing all the work.
There are few independent studies evaluating the accuracy of these tools. Some tools may not store patient recordings and transcripts, which may limit the extent to which health care providers can conduct independent assessments. Inaccuracies may result in patient harm or over- or under-reimbursement from insurers to providers.
GAO's own key takeaway, verbatim: "The accuracy of these tools may be difficult to verify, and the overall effects on health care spending are uncertain." And the deflation worth keeping: the headline benefit in the same report is that "clinicians reduced their documentation time by 20 percent, or two minutes per appointment." Twenty percent and two minutes are the same fact. Only one of them sounds like a reason to sign.
Two honest limits on all of the above. First, GAO says so itself, in the document's own footer: "This document is not an audit product and is subject to revision based on continued advances in science and technology." A Spotlight is a briefing for Congress, not a verdict. Second: this report is about clinical documentation and medical coding generally — it does not single out radiology or imaging, and nothing in it is a finding about your modality or your code mix. It matters to you because you are the one being pitched these tools, not because GAO said anything about you.
Sources (primary, verified today): GAO-26-109116, "Science & Tech Spotlight: AI for Medical Notes and Coding" (July 2026) — every quoted passage taken verbatim from the report PDF, including the developer-attribution sentence, the Challenges section, and the not-an-audit-product notice. The 21%-to-28% clinician adoption figure cited in the report is attributed by GAO to the American Medical Association (2026); the 20%/two-minute documentation figure is attributed by GAO to "one study." GAO contact of record: Sarah Harvey. Caveat: a Science & Tech Spotlight is a technical briefing product, not an audit or a recommendation, and this edition takes no position on any named vendor.
GAO published a Spotlight on AI for medical notes and coding this morning. Two numbers from it will be in a vendor deck by Friday: 95% coding accuracy, $1 million saved. Both are really in the document. Neither is a GAO finding. The sentence, whole: "According to one AI medical coding software developer, when its software was deployed at a health system with five hospitals, it generated medical codes with more than 95 percent accuracy, and the system's emergency departments reduced annual coding costs by more than $1 million." A developer's claim. About its own software. At one unnamed system. GAO attributed it plainly — the attribution is just what falls off in the retelling. Because two pages later, in GAO's own voice: "There are few independent studies evaluating the accuracy of these tools. Some tools may not store patient recordings and transcripts, which may limit the extent to which health care providers can conduct independent assessments." And the benefit everyone will round up? "Clinicians reduced their documentation time by 20 percent, or two minutes per appointment." Twenty percent and two minutes are the same fact. Only one sounds like a reason to sign. Three questions for the demo: Which independent study — one that isn't yours? Do you retain recordings and transcripts, so we can audit a code? What happens if it captures more diagnoses than we would have — because GAO says that raises spending, and it's also what an over-reimbursement finding looks like. GAO-26-109116. It's four pages. Read the attribution lines. Read the source, not the coverage.
Use the MRP Personal Design System (Signature). Create a 1080x1350 animated data piece titled "Read the Attribution Line." Bone field (#F4F1EC), Ink type (#141414), Coral Bright (#FF6A4D) reserved for the single big display number, Coral Deep (#D9401F) for the kicker and small tags. Cormorant for the number and headline, Manrope Light for captions and body. Tall, airy, left-aligned, generous whitespace — a document being read closely, not a chart. Sequence (about 12 seconds, calm and forensic — no alarm): 1. Kicker top-left, letter-spaced uppercase in Coral Deep: "PUBLISHED THIS MORNING · GAO SPOTLIGHT" — types on quickly and settles. 2. An enormous Cormorant "95%" settles in from slight opacity in Coral Bright, left-aligned. Beneath it, a Manrope Light caption fades in: "coding accuracy, per today's GAO report." Hold — let the reader believe it. 3. Beat. Beneath the caption, small Manrope Light body type types on one line at a time, like someone reading aloud: "According to one AI medical coding software developer," — and as that clause lands, the words "According to one AI medical coding software developer" glow briefly in Coral Deep while the big 95% desaturates from Coral Bright to plain Ink. The number does not move. It just stops being GAO's. 4. A thin ink hairline draws across the frame. Below it, in Cormorant italic: "There are few independent studies evaluating the accuracy of these tools." Small Manrope tag beneath in Coral Deep: "— GAO, same report, two pages later." 5. End card: everything clears but one Cormorant line — "Read the attribution line." Footer in Manrope Light: "GAO-26-109116 · July 16, 2026." Close with the SignatureMark (Italiana "Meschelle Peterson" + code63labs handle), bottom-left. Motion language: slow eases, opacity and color only, no bounces, no particles. The one move that matters is a color drain, not a motion. Restraint is the aesthetic.
FIRST SOURCE · one verified original-source finding, composed for one reader · this edition: radiology imaging centers