What the study actually measured
Most published evidence for AI-assisted radiology reporting comes from a single vendor demonstrating its own product's impact at one site. The study behind this story is different. Published in the Journal of the American College of Radiology (JACR, September 2026) and led by Dr. Sergey Morozov and colleagues, it tracked 10 different AI tools from seven separate vendors, running on a shared orchestration platform across the 3R Swiss Imaging Network — 20 outpatient imaging centers in French-speaking Switzerland staffed by 58 radiologists.
The dataset is large and the observation window is long: about 389,000 AI-assisted imaging exams processed between January 2021 and June 2025, nearly five years of real clinical use rather than a short pilot. As Radiology Business reported, the researchers "found clear benefit, with statistically significant efficiency gains and widespread radiologist adoption."
That combination — multiple vendors, one network, nearly five years — is what makes this study notable. It is not evidence that one company's algorithm works. It is evidence that AI-assisted reporting workflows, implemented at scale across a private-practice network, produce measurable radiology efficiency gains independent of which vendor supplied the tool.
The numbers, modality by modality
After adjusting for factors like radiologist experience, the study found AI availability was associated with lower median report turnaround times in the network's highest-volume modalities:
| Modality | Turnaround time change | Note |
|---|---|---|
| Trauma radiography | -26% | Highest-volume modality in the network |
| Knee MRI | -18% | Second highest-volume modality studied |
Adoption was broad, not confined to early adopters. Across the network, 91% of radiologists actively used at least one AI tool, and 66% described themselves as regular users. As reported in the RSNA Daily Bulletin, musculoskeletal imaging accounted for the bulk of AI activity, and adoption among breast, chest, and brain imaging specialists reached roughly 76%.
The researchers also quantified capacity, not just speed: for trauma radiography alone, AI performed the equivalent of about 0.46 of a full-time radiologist's workload — a concrete way of expressing what "efficiency gain" means on a schedule rather than only in a turnaround-time percentage.
The real bottleneck wasn't the AI
The study's most useful finding for buyers isn't the turnaround-time percentage — it's the diagnosis of what limits it. Median total latency, the delay between image acquisition and AI results reaching a radiologist, was about 2.06 minutes per exam. Of that delay, roughly 72% was attributable to data routing — PACS retrieval and network infrastructure — not to the AI model's processing time.
As the authors put it, reported by Radiology Business: "infrastructure latency, not algorithm speed, was the primary barrier to clinical utility." The study also measured a "too-late rate" — how often AI results arrived after a radiologist had already started or finished a report — which averaged 7.2% overall but ran as high as 13% for chest CT, where delays most erode the value of an AI draft.
The lesson generalizes beyond this one network: an AI tool's accuracy is only part of its value. If results don't reach the radiologist's worklist before the report is started, the efficiency gain never materializes — regardless of how good the underlying model is.
Why multi-vendor, independent evidence matters more than another case study
Radiology groups evaluating AI are used to seeing single-vendor white papers — useful, but hard to generalize from, since the vendor controls the site, the comparison, and often the publication. A study spanning seven vendors and 10 tools, run by a practice network rather than a vendor, and published in a peer-reviewed journal, carries a different kind of weight. It is evidence for the category of AI-assisted reporting workflows, not a referendum on one product.
That doesn't mean every tool performed equally — the researchers also tracked radiologist satisfaction by tool and found it varied widely, with some algorithms earning strongly positive feedback and others landing negative. The practice-level efficiency gain was real; the experience of getting there was uneven across the 10 tools, which is itself a reminder that "AI adoption" at the network level is really many smaller adoption decisions stacked together.
An ROI framework for US outpatient imaging centers
Translated for a US outpatient imaging center or radiology group sizing up an AI investment, three takeaways follow:
Radiology efficiency gains are reproducible, not anecdotal
Statistically significant turnaround-time reductions held across multiple vendors and nearly five years of real use — not a single pilot quarter. That is a stronger basis for an ROI case than one company's before/after chart.
Evaluate the delivery pipeline, not just the model
With infrastructure latency responsible for most of the delay in this study, buyers should ask vendors how fast a report draft reaches the reading radiologist end-to-end — PACS/RIS integration and routing included — not just how accurate the model is in isolation.
Expect uneven results across modalities and tools
Gains concentrated in the highest-volume modalities (trauma radiography, knee MRI) and satisfaction varied by tool. A credible vendor evaluation should look at results by study type, not a single blended average.
Where xAID fits
The Swiss network's bottleneck was the handoff between acquisition and radiologist — exactly the seam AI CT reporting is built to close. A foundation-model approach produces one comprehensive, structured report draft per study instead of a stack of narrow detection outputs to reconcile, and xAID's in-house radiologist reviews every preliminary before it reaches a client's worklist ready-to-sign. The study's core finding — that efficiency gains depend as much on how fast and cleanly a draft reaches the radiologist as on the model itself — is the same design question xAID is built around.
Frequently asked questions
What did the JACR study find about AI and radiology efficiency?
A September 2026 study in the Journal of the American College of Radiology examined 10 AI tools from seven vendors used across 20 outpatient imaging centers and 58 radiologists in the 3R Swiss Imaging Network, covering about 389,000 AI-assisted exams over roughly 4.5 years (January 2021 to June 2025). It found statistically significant radiology efficiency gains in high-volume modalities, including trauma radiography and knee MRI, along with widespread radiologist adoption.
How much did AI reduce radiology report turnaround time?
After adjusting for radiologist experience and other factors, the study found report turnaround time fell 26% for trauma radiography and 18% for knee MRI, the two highest-volume modalities studied. For trauma radiography alone, AI performed work equivalent to roughly 0.46 of a full-time-equivalent radiologist.
What was the biggest obstacle to AI's impact, according to the researchers?
Not the AI algorithms themselves. Median total latency, the delay before AI results reached radiologists, was about 2.06 minutes per exam, and roughly 72% of that delay came from data routing and retrieval infrastructure rather than AI processing time. The researchers concluded that infrastructure latency, not algorithm speed, was the primary barrier to clinical utility.
Is this independent evidence or one vendor's marketing claim?
It is independent, peer-reviewed research spanning 10 AI tools from seven different vendors running on a shared orchestration platform, not a single company's case study. That multi-vendor design is what makes the turnaround-time findings meaningful evidence for AI-assisted reporting as a category, rather than one product's marketing claim.
What does this mean for US outpatient imaging centers evaluating AI?
The results suggest radiology efficiency gains from AI-assisted reporting are reproducible at the practice level, not confined to one algorithm or specialty. But the gains tracked closely with integration quality, not just model accuracy, so buyers should evaluate a vendor's full delivery pipeline, including data routing, PACS/RIS integration, and how fast a report draft reaches the radiologist's worklist, alongside a radiologist-reviewed, ready-to-sign workflow.
Source: S. Morozov, N. Heracleous, D. Korka, C. Thouly, B. Dufour, O. Novarina, B. Rizk, "AI Latency, Report Turnaround Time, and Adoption in a Multi-Vendor AI Ecosystem: A Multi-Site Observational Study," Journal of the American College of Radiology (September 2026), doi.org/10.1016/j.jacr.2026.09.026, as reported by Radiology Business, The Imaging Wire, and the RSNA Daily Bulletin. Figures are rounded as reported.