"Radiology AI. Two paths," described AI that assists radiologists (Smart Impression drafting impressions from the radiologist's findings) and AI that aims to interpret on its own (RP's Mosaic). RadNet's October 27 investor call on its ONRAD acquisition adds a useful wrinkle: the most profitable teleradiology company in the conversation got there without either one.
What ONRAD actually automated
ONRAD reports adjusted EBITDA margins above 30%. CEO Dave Engert credited the plumbing, not the reading: image ingestion, pulling priors, worklist assignment and report delivery. When ONRAD absorbed Direct Radiology, a low-margin operation, it brought it to roughly 30% margins with that same workflow technology. No AI reads required. ONRAD handles about 8,000 reads a day, including roughly 250 strokes with a seven-minute average turnaround.
Prices are going up, not down
Engert said ONRAD has raised customer prices to cover rising radiologist pay, and expects that pressure to continue. Radiologist scarcity still buys pricing power.
Radiologists may enjoy the rising income, but it comes with a catch. Every raise makes the radiologist a bigger line item, and the biggest line item is always what management tries to shrink. Scarcity doesn't protect radiologists. It supports the effort to replace them.
Then comes the AI
RadNet CEO Howard Berger wants every exam routed through RadNet's AI workflow, with draft reports for radiologist review, and projects 50% or more productivity gains, with 50 to 60% of imaging passing through AI by early 2027. These are projections, not results. Note the CFO's point: ONRAD's expected $60 to 65 million in EBITDA doesn't depend on any of it. RadNet is buying a business that already works and betting AI makes it work with fewer radiologists.
The number to watch
Not the percentage of exams "touched by AI." Watch paid radiologist-hours per completed read. That ratio is the clock in this site's tagline.

Radiology Partners, High-Yield Debt, and the Economics of AI

October 22, 2026 | 0rad.com
Radiology Partners (RP) is taking on substantial additional debt to acquire Everlight Radiology, an international teleradiology company with more than 800 radiologists and approximately 2.5 million annual examinations. The acquisition raises an interesting question: How important will artificial intelligence be to RP's ability to manage its debt and improve profitability?
To help finance the acquisition, RP recently arranged $485 million in high-yield bonds paying 9.75% annual interest. These bonds, commonly called junk bonds, carry higher interest rates because investors perceive greater financial risk. The new borrowing alone will cost approximately $47 million annually in interest.
RP already carries substantial debt. Credit-rating agencies estimate its total debt at roughly 8 to 10 times annual adjusted operating earnings, depending on the methodology used. That is a high level of financial leverage, particularly for a business facing continued pressure from radiologist compensation and reimbursement.

Where Mosaic AI Fits In

RP is simultaneously developing Mosaic Drafting, an AI system designed to analyze imaging examinations and generate preliminary radiology reports for physician review and approval.
The potential financial implications are considerable.
Suppose AI allows radiologists to interpret 30% more examinations per hour. RP could use that additional productivity to expand capacity, increase physician compensation, improve operating margins, or some combination of these.
But for a highly leveraged company, reducing physician labor expense per examination would be particularly valuable.
Consider a hypothetical savings of just $10 per examination across 10 million annual interpretations. That would represent $100 million in annual savings, before AI development and operating costs, more than twice the annual interest on the new Everlight bonds.
The actual savings remain unknown, and RP has not demonstrated that Mosaic can achieve reductions of this magnitude.

Could AI Eventually Replace Some Radiologist Labor?

Today, Mosaic Drafting requires radiologist review and approval. It is not an autonomous interpretation service.
Nevertheless, as AI becomes more capable, it could progressively reduce the physician time required for each examination. If productivity increases faster than imaging volume, fewer radiologist-hours may ultimately be necessary.
The financial benefit to RP would depend on how physician compensation adjusts to those productivity gains.
A more distant possibility is autonomous AI interpretation of selected examinations. That would raise significant regulatory, liability and reimbursement questions. Current U.S. rules do not provide a general mechanism for AI software to replace radiologists and independently bill for professional interpretations.

The Financial Question

RP's debt does not mean the company must replace radiologists to survive. It has other options, including revenue growth, refinancing and improvements in its existing operations.
However, its financial structure creates a strong incentive to convert AI-driven productivity improvements into operating cash flow.
The acquisition of Everlight increases the number of examinations across which RP could potentially deploy Mosaic, making even relatively small reductions in interpretation costs financially meaningful.
For radiologists, the important questions are whether AI will meaningfully reduce interpretation time, how physician compensation will reflect those efficiencies, and whether productivity gains will ultimately reduce staffing requirements.
For RP, the critical measure of Mosaic's success may be less about the quality of its draft reports than the reduction in physician labor cost per completed interpretation.

AI helping versus replacing radiologists- a Tale of Two pathways

There are two ways AI can enter the reading room. It can sit beside you, or it can sit in your chair. Both are already happening, and two recent examples show the difference.

Pathway one: AI as assistant

Intermountain Health reports that about 75% of its radiologists now use Smart Impression, a generative AI feature in Microsoft’s PowerScribe. Most have had it for roughly six months.

The workflow will feel familiar. You read the study and dictate your findings. The AI drafts the impression from those findings. You edit and sign. The radiologist still does the interpretation; the software handles the summary.

Intermountain cites lower cognitive load and good acceptance, but no hard numbers on productivity, turnaround time or staffing. Becker’s Hospital Review

The takeaway: radiologists will adopt AI quickly when it removes drudgery and leaves them in charge. Three out of four in half a year is fast for any new tool in medicine.

Pathway two: AI as interpreter

Radiology Partners’ Mosaic Drafting starts at the other end. Instead of working from your findings, it looks at the images and generates the draft report itself. The radiologist’s role shifts from reading the study to reviewing the machine’s read.

That sounds like a small change in sequence. It isn’t. One tool makes you faster at your job. The other is on the way to taking your job, disguised as quality control.

What’s the difference?

Both tools produce a draft. Both keep a radiologist’s signature on the report. From a distance they look alike, and adoption headlines will lump them together. But they lead to very different places:

• Assistive AI increases what each radiologist can do, increasing productivity.
• Interpretive AI reduces how much radiologist is needed per exam, so some of the interpretation fee goes to the software company, with high margins.

The Intermountain numbers show that AI in the reporting workflow is highly adopted. What is still to be determined is which pathway wins. That will be driven by radiologist decisions, legislative actions, societal pressure and cost analysis.  

Welcome to 0Rad, aka ZeroRad.  Consider it a way to make sense of the changes surrounding one small, impacted profession brought about by AI.  

This site is the simmering result of a post by Paul Kedrosky, the first part of which (below) is essentially required reading if you are a radiologist. 

Paul’s post struck a chord that I've been unable to shake.  Mark Twain really said "History never repeats itself, but the kaleidoscopic combinations of the pictured present often seem to be constructed out of the broken fragments of antique legends."  I hope you gain some sense or amusement from the posts that follow
.  

The Wisdom of Teamsters, plus Product Traps and Housing BombsWhat happens when people decide an occupation is toast? This is an important question given what we see happening in many occupations that once looked appealing, and now, post large language models (LLMs), look less so.paulkedrosky.com