The blank screen used to be where radiology work began. Now, at a growing handful of practices scattered from California imaging centers to global teleradiology networks, radiologists open their workstations to find draft reports already populated, AI-parsed findings waiting for review, the cognitive grunt work already done.
RadNet's digital health division posted $96.9 million in annual recurring revenue as of March 31, 2026, according to company figures released this year—nearly double the $49.8 million recorded twelve months earlier. The jump reflects something broader than incremental automation. A clutch of operators, ranging from 300-radiologist group practices to international teleradiology services, have begun replacing legacy picture archiving and communication systems wholesale, swapping out reporting stacks and PACS infrastructure for what they call AI-native operating environments. These platforms don't bolt algorithms onto existing workflows. They rewrite the workflow itself.
The approach breaks from a decade of tinkering. For years, radiology IT meant layering point solutions atop creaky infrastructure: a detection algorithm here, a triage tool there, all feeding into the same dictation software radiologists had tolerated since the early 2000s. That model is giving way, unevenly but unmistakably, to unified platforms where foundation models draft preliminary reports in seconds and orchestration layers route images through governance checkpoints rather than ad hoc vendor integrations.
Everlight Radiology, a teleradiology network operating across six international markets, signed a five-year contract with Sirona Medical in May to jettison its PACS and reporting stack entirely. The replacement, a cloud-native platform Sirona calls RadOS, will support what the company describes as "agentic AI automations" for both clinical reads and operational scheduling. Jeff Oakman, Everlight's global chief operating officer, said the partnership "will fundamentally transform how our radiologists practice."
Radiology Partners, which employed more than 300 radiologists as of July 2025, unveiled its own proprietary system that month. Dubbed MosaicOS and marketed as "fully cloud- and AI-native," the platform integrates viewer, worklist, and AI orchestration into a single environment. One component, Mosaic Drafting, uses multimodal foundation models to generate pre-drafted reports. The tool is under institutional review board evaluation and pursuing FDA clearance. The practice plans to license the platform to outside clients once regulatory review wraps.
Natoe AI, a startup that launched AI-native teleradiology services in July, inserts draft pre-reads into the workflow before any radiologist logs in. Rakesh Deshmukh, co-founder and CEO, frames the pitch bluntly: "A blank report is the most expensive thing in radiology—it means an expert is spending their time on assembly instead of judgment."
Three Hundred Million Studies and Counting
Radiology has long been the proving ground for medical AI. The specialty accounts for roughly three-quarters of FDA-cleared AI devices, a proportion that has held even as the agency's live registry climbed past 1,500 entries by mid-year, according to tracking by Radiological.ai and data cited in Stanford's AI Index Report covering FDA filings through December of last year.
But regulatory clearance is not the same as clinical integration. A survey conducted by KLAS Research in September 2025 found that about half of healthcare organizations deploy some form of imaging AI. Platform governance adoption, however, sits closer to 10 or 20 percent. KLAS characterized broad orchestration as a "future move" for most respondents. Europe leads other non-U.S. regions in deployment, shaped in part by the EU AI Act, which imposed transparency obligations in August and phases in high-risk medical AI requirements through 2028.
The teleradiology sector, where several of these AI-native models have emerged, is projected to expand from $23.9 billion in 2026 to $124.8 billion by 2033, a compound annual growth rate north of 26 percent, according to a June 2026 market analysis by Grand View Research. Workforce pressures add urgency. A December projection from the Health Resources and Services Administration, cited by the American College of Radiology in its February 2026 workforce analysis, anticipates a shortage of 141,160 physicians by 2028. Subspecialty radiology training seats declined through 2023, underscoring the staffing constraints that AI-native practices believe they can alleviate.
What Changed
Regulatory infrastructure matured faster than many anticipated. The FDA finalized guidance on Predetermined Change Control Plans for AI devices in August of last year, a framework that lets vendors update algorithms post-market without filing new submissions if modifications fall within a pre-approved envelope. The agency also created special controls in June for certain radiology machine-learning quantitative software, streamlining pathways for iterative development. Practices aligning with vendors that have robust PCCP strategies can now evolve their tools in production. Those that don't risk obsolescence as competitors iterate faster.
Workforce economics accelerated the shift. RadNet said in its first-quarter update that it expects more than 70 percent of its imaging studies to flow through clinical AI by year-end, with every radiologist report processed through DeepHealth's Reporting Pro, a tool launched on June 10. The company ties AI penetration directly to margin improvement in a capital-intensive specialty wrestling with technologist shortages. BCG Executive Perspectives highlighted RadNet's approach in April, framing the DeepHealth operating system as a case study in "AI-human system redesign" amid persistent vacancies.
Platform economics also tipped. GE HealthCare's Edison AI Orchestrator, Siemens' teamplay marketplace, and Philips' AI Manager all offer vendor-neutral routing and governance layers, a shift from the walled-garden model that dominated the prior decade. Aidoc, which closed a $150 million Series E led by Goldman Sachs Alternatives on April 29, 2026, said it has deployed across nearly 2,000 hospitals and analyzes upward of 60 million patient cases annually. CEO Elad Walach offered a stark timeline: "By 2030, every complex diagnostic decision should be supported by AI that enables earlier detection and reduces preventable error." CARPL.ai, a radiology AI marketplace, raised a $10 million Series A in July led by the International Finance Corporation.
The convergence of these forces—regulatory clarity, workforce constraints, and platform consolidation—has pushed a subset of practices to abandon incremental adoption in favor of wholesale infrastructure replacement.
Inside the Operating Systems
RadNet and DeepHealth launched Reporting Pro in June. The tool complements existing detection and triage algorithms, but integrates directly into the reporting workflow. Digital health annual recurring revenue climbed from $49.8 million on March 31 of last year to $96.9 million twelve months later. BCG cited RadNet's "significant investment" in AI-human workflow redesign as a potential blueprint for other providers navigating staffing shortages, though the consultancy stopped short of declaring the model generalizable across varying practice scales and payer mixes.
Radiology Partners rolled out Mosaic Reporting in 2025 and added Mosaic Drafting as a second phase. The drafting module, which leverages multimodal foundation models to generate preliminary report text, remains under institutional review and is pursuing FDA approval. Nina Kottler, associate chief medical officer for clinical AI at the practice, said in July of last year that "MosaicOS rewrites the rules, connecting AI, clinical expertise and scalable technology into a single solution that empowers radiologists." The practice intends to commercialize the platform to external clients, though no launch date has been disclosed.
Everlight Radiology's five-year Sirona Medical agreement, signed in May, replaces PACS and reporting infrastructure across its six-market footprint. Andy Donaldson, Sirona's chief technology officer, called the deal evidence of a "cloud-native era" in radiology operations. Whether that framing proves prescient or premature will depend on how many peer practices follow Everlight's lead in the next 18 to 24 months.
HOPPR launched Presto Agent on June 24, a tool designed to inject AI draft reporting directly into existing PowerScribe 360 and PowerScribe One workflows. The approach sidesteps the rip-and-replace model, appealing to practices reluctant to abandon legacy dictation systems. Radiologist Daniel Riherd said the agent "runs right inside PowerScribe… saved me a ton of time." Joshua Adam Tarrence, a doctor of osteopathic medicine, noted it "fits right into the system I already report in… takes a lot of the manual busywork out." CMO William Boonn and CEO Khan Siddiqui positioned the tool as solving deployment bottlenecks for risk-averse IT departments.
Yale New Haven Health reported a 12 percent boost in reporting efficiency during its first week using Rad AI Reporting, according to a June 10 press release. Rad AI hired Leonard Law, a former product leader at Google Cloud and Coinbase, as chief product officer the following month. Co-founder and CEO Doktor Gurson described the moment as generational: "Radiology is in the middle of a once-in-a-generation transformation. Leaders will build technology that radiologists actually want to use."
Lunit, a breast-imaging AI vendor, said in April it supports more than 330 screening sites across the Americas and roughly 1 million annual screening mammograms, with FDA clearance for next-generation 3D mammography algorithms. Viz.ai presented data at the International Stroke Conference in March showing a 44 percent reduction in interfacility stroke transfer door-in-door-out time, a metric with direct implications for patient outcomes and hospital throughput.
Governance Lags Behind Deployment
The American College of Radiology and the Society for Imaging Informatics in Medicine approved the first joint practice parameter for imaging AI in May, outlining protocols for selection, deployment, monitoring, and continuous quality improvement. Tessa Cook, who chaired the effort, said the parameter represents a "first-of-its-kind" framework covering the algorithm lifecycle from procurement through ongoing performance audits.
The ACR's ARCH-AI quality-assurance program, launched in June 2024, recognizes practices that meet AI governance benchmarks, with an emphasis on local acceptance testing and longitudinal monitoring. The Assess-AI registry, which opened in November 2024, collects real-world algorithm performance data, capturing AI outputs, report text, and DICOM metadata for post-market surveillance.
"AI is different from previous technologies," Christoph Wald, vice chair of the ACR Board of Chancellors, said in June 2024. "Even a U.S. Food and Drug Administration-cleared AI product must be tested locally to ensure it works safely and as intended." A systematic review published in Radiology: Artificial Intelligence in January called for standardized outcome measures and return-on-investment frameworks across diverse care settings. An October benchmark study titled "Radiology's Last Exam" tested frontier multimodal AI against human experts and cautioned against unsupervised clinical deployment.
Reimbursement, however, remains bundled for nearly all imaging AI applications. The American Medical Association's CPT Appendix S taxonomy, updated on June 8, clarifies categories for assistive, augmentative, and autonomous software outputs, laying groundwork for future coding proposals. ACR coding updates for this year include new and revised Category III codes for AI-derived outputs such as malignancy risk scores.
The fiscal year 2025 New Technology Add-on Payment approval for Annalise.ai's CT brain triage tool for obstructive hydrocephalus, cited by the ACR, suggests a pathway exists for add-on reimbursement in narrow use cases. But such approvals remain isolated. Revenue impact at AI-native practices flows primarily from throughput gains and operational efficiency rather than direct per-study fee adjustments.
That disconnect creates strategic ambiguity. Practices investing in AI-native platforms are making capital bets on margin improvement through volume and speed, not on near-term shifts in the fee schedule. The calculus works if studies per radiologist per hour climb and technologist vacancies ease. It falters if deployment costs outpace productivity gains or if downstream referrers balk at AI-drafted reports.
What Happens Next
KLAS signaled in its 2025 research that this year would pivot toward "stabilizing with AI," a shift from experimentation to consolidation around proven use cases with measurable returns and governance frameworks. Imaging and radiology workflows were named as focal areas. BCG wrote in May that AI "won't fix your health system; redesigning it will," a caution against treating algorithms as standalone solutions rather than components of broader operating-model transformation.
The next 18 months will likely surface clearer winners. Consolidation around unified reporting platforms seems inevitable, as does expansion of draft-reporting and agentic automations embedded directly in radiology operating systems. Increased adoption of Predetermined Change Control Plans will separate vendors capable of iterative improvement from those locked into static algorithms. More outcomes papers will appear, moving from time-savings claims to harder endpoints: cost per study, length of stay, diagnostic accuracy, and quality-adjusted life years.
Practices launching today confront a binary choice, or something close to it. They can retrofit AI onto legacy infrastructure, preserving existing vendor relationships and dictation habits while layering on point solutions as budgets and tolerance allow. Or they can rebuild from the platform up, a path that demands capital, governance maturity, and willingness to navigate regulatory iteration.
The early cohort—RadNet, Radiology Partners, Everlight, Natoe AI, and a handful of others—is attempting to prove that the unit economics of the latter approach hold. Walach's claim that every complex diagnostic decision will be AI-supported by 2030 may turn out to be conservative. The question radiologists face is not whether they will practice in that world, but whether they will arrive there through incremental patches or by rewriting the operating system now. The answer, for an increasing number, appears to be the latter. Whether that translates into a durable competitive advantage or simply raises the table stakes for everyone remains to be seen.
