What Would Happen If AI Doubled Radiologist Reading Productivity?
RADIOLOGY IN THE AGE OF AI & VLMS | ARTICLE 4 OF 14
A Swedish randomized trial showed that AI safely cut radiologist workload in mammography screening by 44%, without missing more cancers. One radiology AI startup now claims its system automates 25% of routine tasks and has set 90% as its target.
If even a fraction of those numbers hold at scale, the question is not whether AI will change how we read. It is, ‘how can this safely be implemented?’
What the data say
The MASAI trial, a prospective randomized study from Sweden, remains the strongest controlled evidence on AI-assisted screening. AI triage halved the number of mammograms requiring radiologist double-reading while cancer detection held steady. That is not a vendor claim. It is a peer-reviewed RCT.¹
Langlotz and colleagues at Stanford modeled what similar efficiencies could mean for the U.S. radiologist workforce over five years, estimating a 14-49% reduction in radiologist hours worked, not headcount.² The distinction matters.
The most grounded real-world evidence comes from Northwestern Medicine, published in JAMA Network Open in 2025:
Generative AI deployed across 11 hospitals, 24,000 reports, five months, all X-ray types. Average reading efficiency gain was 15.5%.
The individual ceiling, for radiologists who used the tool most fluently, reached 40%, without accuracy loss.³ Unpublished follow-on data from the same group suggests CT gains could reach 80%.
One more detail worth noting: Northwestern built this system entirely from institutional clinical data, no commercial LLM, at a fraction of big-lab cost. It is the DeepSeek moment for healthcare AI, and we will return to it in Article 12 on build versus buy.
The paradox
AI reading efficiency is growing, but the radiologist shortage is growing alongside it, and demand is outpacing both.
The AAMC projects a shortage of 17,000-42,000 radiologists, pathologists, and psychiatrists combined by 2033.⁴ Fellowship programs are oversubscribed. Open positions sit vacant for months and salaries are at record highs. Near-term job displacement is structurally improbable in that environment.
AI is getting more capable and the case for hiring more radiologists is getting stronger. For now, those forces run in parallel.
“For now” is the key point.
The question the literature is not asking
Even granting the near-term paradox, there is a distribution question the radiology literature has mostly avoided: who captures the productivity gains?
History offers an uncomfortable answer. Across virtually every industry where automation increased output per worker, efficiency gains flowed primarily to employers, platform owners, and investors. Workers saw some wage growth and job preservation, but not proportional gains.
Economist James Bessen documented this pattern in his work on innovation and wages. His inverted U-curve is instructive: automation initially creates jobs as lower costs expand the market, then labor demand peaks and falls as machines substitute for human tasks. Acemoglu and Restrepo framed it plainly: automation increases the size of the pie, but labor gets a smaller slice.
A 2025 perspective by Ruthven and Agten in the European Journal of Radiology Artificial Intelligence applied this framework directly to our field.⁵ Their core argument: once AI surpasses human reading accuracy, practices face pressure to reduce their largest expense, which is salaries. Staffing decisions are increasingly made not by radiologists but by department heads, practice owners, and private equity firms. Between 2012 and 2024, radiologist ownership of U.S. practices fell from 63% to 46%, while PE ownership rose from 1% to 13%. If AI enables the same reading contract to be serviced with fewer radiologists, that is margin expansion, and the gain flows up, not across.
This is not a political argument. It is a structural one. Employed radiologists and group partners need to understand it now, because it will shape contract negotiations and practice decisions for the next decade.
Where the gains come first
Not all of radiology is equally exposed. High-volume, relatively homogeneous screening tasks are where AI provides the most reading lift soonest: mammography, lung nodule follow-up, normal chest X-ray triage, bone age. These tasks have well-defined outputs, large training datasets, and measurable ground truth.
Complex cross-sectional work is a different problem. The abdominal MRI with an indeterminate adrenal lesion, the chest CT with three competing differentials: that is where clinical judgment and subspecialty expertise still carry most of the cognitive load. AI is a collaborator there, not a reading multiplier. Not yet anyway.
On the industry side, Cognita CXR, the generative radiology AI that recently received FDA Breakthrough Device Designation, reports an 18% reading efficiency gain in deployed environments.⁶ We will examine what the Breakthrough Designation means for the regulatory pathway in Article 9. The number is consistent with Northwestern: real and meaningful, with the ceiling still being found.
What practice leaders should be asking now
The question is not whether AI will change reading workflow. It is already changing it in practices that are deploying it. The more useful question is: what does 1.5x or 2x reading throughput mean for your specific contracts, staffing model, and market position, and who benefits from that efficiency?
Some groups will use AI-driven productivity to take on more volume. Others will use it to offer faster turnaround and differentiate on service. Some employed radiologists will see those gains returned as compensation or reduced call. Others will find the math used to justify holding headcount flat.
The groups best positioned are those asking these questions now, before contracts are renegotiated around AI assumptions someone else set.
Up Next in Article 5:
In the next article we will look at the broader radiology labor market: the shortage, the pipeline, and what happens if imaging volume plateaus while AI reading throughput keeps climbing.
The bottom line
Reading productivity gains from AI are real, peer-reviewed, and accelerating. The best current evidence puts average gains at 15.5% with individual ceilings at 40%, without accuracy loss. Screening-heavy practices will feel this first.
The near-term workforce crisis provides insulation from displacement, for now. But it does not address the distribution question: efficiency gains in medicine, as in other industries, have a historical tendency to flow toward capital rather than labor. The Bessen U-curve has not reversed itself for any other profession, and there is no obvious reason radiology will be the exception.
Understanding this clearly is part of navigating the next decade well.
AI can increase output per radiologist if it behaves like a well-trained fellow.
If it behaves like a first-year, supervision friction is too high.
If you want to deploy AI in a way that expands effective capacity, protects revenue, and surfaces risk early, I can help.
I am identifying a small number of forward-leaning partner sites to build and pilot independent AI performance evaluation software in real clinical workflows.
More information:
https://orainformatics.com/aiconsulting/
References
1. Lång K, et al. AI-Supported Screening for Breast Cancer: A Prospective, Population-Based Paired-Reader Study (MASAI trial). Lancet Oncology, 2023. https://doi.org/10.1016/S1470-2045(23)00334-5
2. Langlotz CP et al. What Effect Will AI Have on the Radiologist Workforce? AuntMinnie / MedRxiv, 2024. https://www.auntminnie.com/imaging-informatics/artificial-intelligence/article/15814368/what-effect-will-ai-have-on-the-radiologist-workforce
3. Huang Y, Etemadi M, et al. Generative AI Boosts Radiology Productivity Up to 40% in Large Multi-Site Clinical Deployment. JAMA Network Open, 2025. https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2834943
4. Christensen EW et al. Short-Term Strategies for Augmenting the National Radiologist Workforce. AJR, 2024. https://ajronline.org/doi/10.2214/AJR.24.30920
5. Ruthven H, Agten C. Perspective: AI Productivity Will Not Benefit Employed Radiologists. European Journal of Radiology Artificial Intelligence, 2025. https://doi.org/10.1016/j.ejrai.2025.100033
6. Mosaic Clinical Technologies / Radiology Partners. FDA Breakthrough Device Designation for Cognita CXR Generative AI Model for Radiology. Business Wire, March 4, 2026. https://www.businesswire.com/news/home/20260304633206/en/Mosaic-Clinical-Technologies-Announces-FDA-Breakthrough-Device-Designation-for-Cognitas-Generative-AI-Model-for-Radiology