# Transcript: 4.5 years of AI in everyday radiology (3.5 min)

Narrated by Sergey Morozov. Explainer of Morozov et al., J Am Coll Radiol 2026, doi:10.1016/j.jacr.2026.09.026. Spoken numbers are rounded; the exact values are in the article.

Hello. I am Sergey Morozov. The 3R Swiss Imaging Network has used AI in everyday radiology. Here is what we learned along the way. When a computer starts thinking slower than we do, we simply replace it. So what should we do with an AI whose answer arrives after the report is already signed? At the 3R Swiss Imaging Network: Four and a half years, 20 centers, 10 AI tools from 7 vendors, and nearly 400,000 AI-assisted exams. Everyone asks whether AI is accurate. Far fewer ask whether its answer arrives in time to make a difference. To find out, we triangulated three data sources. The AI platform dashboard with almost 100,000 exams, the RIS with about 21,000 exams read with and without AI, and 58 radiologists whom we surveyed twice. We looked at three clocks. Clock one is AI latency. About two minutes from the end of the scan to the result. The algorithm itself needs only a fraction of that. Almost three-quarters of the time is data routing. Picture a sprinter stuck in traffic. Clock two asks whether the result arrives in time. More than eighty percent of results are ready before the radiologist starts the report. About 7% arrive only after the report is signed. We call this the Too Late rate. And to our knowledge, it is the first time it has been measured. Clock three is the radiologist's report time. When the AI result is there in time, it pays off. Reports were 26% faster in trauma radiography and 18% faster in knee MRI, adjusted for the individual radiologist. That is roughly a three-to-one return on the cost of the AI solution, and still about two-to-one after adjustment. On top of the clocks, we asked the radiologists themselves. More than nine in ten use AI, and they rate their AI tools above seven out of ten. Yet their perceived productivity stays just below three out of five, under the neutral midpoint. Economists know this as the Jevons paradox. Time saved may simply turn into more cases. From this triangulation come our monitoring metrics. The three clocks tracked per tool every quarter. We also propose a standardized questionnaire for radiologists and two practical steps. Push studies to AI as soon as the scan is complete, and put time-to-display targets into contracts. And then comes the real paradox. Today, AI sometimes cannot keep up with the radiologist. Fix that, and with rising volumes, the radiologist may no longer keep up with AI. Will we still be able to monitor what it does? So we have to do two things. Monitor AI and prepare radiologists to work with it. AI should run like a Swiss train. On time, comfortable, and taking us exactly where we need to go. Right result, right moment, from the 3R Swiss Imaging Network, published in JACR. If you have questions, write to the 3R research team.