A fourteen-hospital network in the American Midwest published results this month from eighteen months of running AI in radiology, and the most useful number is not the 41% reduction in report turnaround. It is the 6% override rate the system settled at, after starting at 18%.
What they actually deployed
Not autonomous reads. The network kept every diagnosis with a radiologist and split the work in two. A triage layer, using a cleared device, ordered the worklist by urgency and flagged studies with suspected intracranial haemorrhage, large-vessel occlusion and pneumothorax for immediate read. A drafting layer, which is not a cleared diagnostic device, produced a structured draft: technique, comparison, findings by organ system, and a suggested impression with the sentences it drew from.
The radiologist edits the draft. Nothing reaches a referring physician without a signature.
Why drafting beats reading
The economics of radiology AI have been stuck for years because the value proposition was framed as replacing reads, which forces a regulatory conversation about diagnosis and a labour conversation about headcount. Drafting sidesteps both. It attacks the part of the job that is transcription and structure — 40% to 60% of the time on a routine chest CT — and leaves interpretation alone.
The network measured this directly. Median time from study completion to signed report fell from 4.7 hours to 2.8. The after-hours backlog, the queue that builds between 10pm and 6am, fell 62%, because drafts were ready when the morning radiologist logged in.
The override rate is the real metric
An 18% override rate at go-live sounds alarming and is normal: it is the rate at which radiologists changed a finding or an impression materially. It fell to 6% over eight weeks, and the mechanism was not a better model. It was a local fine-tune on 40,000 de-identified reports from the network's own template conventions — how they phrase a negative finding, which measurements they include, what goes in the comparison line.
That is the transferable lesson. Generic drafting quality is a solved problem at the level of grammar and structure. Institutional voice is not, and it is cheap to fix with local data if you plan for it.
Where it still fails
Three categories, consistently. Rare findings, where the draft omits something a radiologist would have caught on a second look. Hardware-adjacent artefacts — implants, coils, motion — where the model describes the artefact as pathology often enough to matter. And paediatric studies, where the training distribution is thin and normal variants are common.
The network's rule is blunt: any study with an implant flag, any patient under 18, and any examination the triage layer scores as low-confidence goes to a radiologist with no draft. That removes about 14% of volume from the drafting path and, by their account, most of the serious near-misses.
What to copy, what to skip
Copy the split. Triage and drafting are separate products with separate regulatory postures, and buying them separately keeps the cleared part cleared. Copy the local fine-tune, and budget for it — it took six weeks and one part-time data engineer.
Skip the temptation to measure accuracy alone. The numbers that changed behaviour inside the network were turnaround, backlog and override rate, because those are the ones a department chair can act on.