We previously used Chat GPT to assess a set of 5 "AI Reimbursement" articles from Katie Palmer at Stat Plus. Here, we expand the approach to a larger set of 8 of her business strategy articles.
##
The attachment contains eight substantive articles. Read together, they support a comprehensive review of radiology’s AI frontier—and of the commercial, clinical, and policy choices likely to confront digital pathology next. 2026 STAT 1013 Bigger Katie pal…
Radiology’s AI transformation is becoming a transformation of the diagnostic business itself. Across eight STAT articles published between December 2025 and October 2026, Katie Palmer follows developments that initially appear separate: report-generating foundation models, practice acquisitions, opportunistic screening, hospital prediction systems, Medicare payment changes, and consumer purchases of cancer-risk assessments. Together, they reveal an industry reorganizing around the ability to extract more information from existing clinical data—and to control the infrastructure through which that information reaches patients.
The familiar question, “Can AI interpret an image?” remains important. But Palmer’s reporting increasingly directs attention to another question: Who can turn that interpretation into a useful, sustainable clinical service?
For pathology, radiology provides an unusually valuable preview. The observation that radiology runs roughly five years ahead is best understood as a strategic rule of thumb, rather than a fixed timetable. Nevertheless, its experience offers an opportunity to study problems before they become fully developed in pathology: integration costs, ownership of diagnostic platforms, evidence requirements, professional responsibility, and the struggle to capture the financial value of better information.
A chronological reading shows how these issues connect.
December 2025: Foundation models begin to change the product being sold.
“Off an $80 million acquisition, Cognita’s CEO on the power of scale in radiology AI”Palmer’s interview with Cognita Imaging CEO Louis Blankemeier describes a fundamental expansion in ambition. Earlier radiology algorithms generally searched for particular findings. Cognita’s vision-language models aim to generate preliminary reports encompassing a much broader range of abnormalities.
Radiology Partners’ acquisition of Cognita matters because it combines software development with a large clinical organization. The practice supplies access to clinical workflows, radiologist feedback, and opportunities for evaluation and deployment. Distribution becomes part of the development strategy.
The company reports approximately 20% average time savings in a retrospective evaluation and increased detection of selected findings. These are company-reported results described in an interview, not independently established benefits across all practice settings. Blankemeier also acknowledges that validating free-text reports is difficult: a conventional radiology report is an imperfect reference standard, and automated text comparisons may fail to distinguish clinically important errors from harmless differences in wording.
The strategic implication is substantial. The emerging product is assistance with a large portion of the examination and reporting process. Its value may therefore be captured through the economics of the existing radiology service, without a separate payment for each algorithmic function.
April 2026: Opportunistic screening exposes the downstream consequences of inexpensive computation.
“AI could check millions of CT scans for heart risk. Who will pay for it?”Algorithms that identify coronary artery calcium on existing chest CT scans appear to offer an especially attractive form of innovation. They extract additional information without requiring another scan, and the article reports a Medicare payment pathway of approximately $15 per analysis.
Yet the principal economic consequences occur after the analysis. Patients may receive statins, consultations, additional imaging, or invasive investigations. Hospitals may view these referrals as revenue opportunities; payers must consider whether the additional utilization produces sufficient benefit.
One study described in the article increased statin prescribing substantially, but did not establish a reduction in cardiovascular events. Another deployment illustrates the patient-selection problem: the vendor estimated that most patients with elevated calcium scores might not warrant notification because they were already appropriately managed or had competing clinical priorities.
The article establishes a distinction that recurs throughout the collection: detecting additional abnormalities is not equivalent to identifying additional opportunities to improve care. Effective screening requires clinical context, selective follow-up, and responsibility for what happens next.
May 2026: Sepsis prediction reveals the competitive power of installed infrastructure.
“In the battle of sepsis algorithms, performance alone doesn’t predict victory”Although focused on hospital prediction rather than imaging, this article illuminates a commercial problem running through the entire collection. An independent developer can produce a strong algorithm and still struggle against a product already embedded in the electronic health record.
Epic’s advantage combines familiarity, integration, purchasing convenience, and relatively low incremental cost. A competing system must justify the organizational effort required to adopt it.
Palmer also distinguishes early warning from recognition of what clinicians already suspect. A model that responds to orders for blood cultures or lactate testing may identify sepsis accurately while providing limited additional lead time.
Bayesian Health’s response combines clinical validation, FDA clearance, prospective supplemental payment, implementation support, and a broader software offering. That combination suggests that successful clinical AI companies increasingly need capabilities extending well beyond model development.
The article’s most unsettling observation concerns developers moving toward billing applications. The market may offer a clearer return for documenting illness than for preventing deterioration. Payment arrangements influence which problems attract investment.
July 2026: CMS attempts to define a coherent category for clinical software.
“CMS signals intent to revamp how it pays for clinical software and AI”This article describes proposals involving “software as a medical service,” New Technology APCs, a new status indicator, and changes affecting selected algorithmic laboratory services. The developments are reported as proposals at the time of publication.
The underlying difficulty is that software does not fit comfortably into payment methods built around physician time, equipment, supplies, and historical costs. Development and validation may be expensive, while the incremental cost of another analysis may be small. Subscription and per-use contracts further complicate comparisons.
For pathology and molecular diagnostics, the most consequential question is where a diagnostic service begins and ends. Is computational interpretation an integral part of producing the laboratory result, or a separately identifiable service applied afterward?
Consistency across specialties is an understandable policy objective. But similar computational methods do not necessarily establish equivalent clinical services. A molecular assay’s interpretive algorithm, an image-based risk score, and a report-drafting assistant may perform very different functions.
CMS’s own broader policy discussion confirms interest in the incremental value of clinical software and in payment approaches that support technology-enabled care while controlling spending. CMS
August 2026: Temporary Medicare payments reveal the importance of pricing denominators.
“What Medicare incentives for AI-based devices mean for tech companies—and hospitals”Palmer examines new technology add-on payments as a practical tool for overcoming hospitals’ reluctance to purchase unfamiliar technology. Temporary reimbursement can make an uncertain investment easier to approve.
But software subscriptions must be translated into patient-level payment amounts. The article’s striking example is InVision’s cardiac amyloid software: its reported maximum add-on changed from $162.50 to $2,275 as the proposed billing approach shifted toward positive findings.
That difference illustrates the influence of the denominator. Spreading a subscription across all analyzed patients produces one apparent per-case cost; allocating it across a much smaller set of positive cases produces another.
Maximum NTAP amounts also should not be confused with guaranteed receipts. Payment depends on case-level eligibility and costs, and hospitals need the operational capacity to identify and code qualifying cases.
The article shows that obtaining a payment pathway and realizing its financial value are separate accomplishments. It also raises the longer-term question: what sustains adoption when the temporary supplement ends?
September 2026: Developers and clinical practices begin to converge.
“In radiology, AI is blurring the line between technology development and clinical practice”This is arguably the collection’s most consequential article for understanding industry structure.
Radiology groups are acquiring and developing technology. Technology companies are supporting or acquiring radiology practices. New entrants propose to redesign the entire operation around AI, including scheduling, quality checks, reporting, communication, and follow-up.
Owning or closely partnering with a practice offers several advantages: a place to deploy the technology, access to user feedback, control over workflow, and the ability to retain operational gains. It can reduce dependence on selling individual algorithms to one institution after another.
Palmer also distinguishes efficiency claims from improved clinical outcomes. A practice may become more productive without proving that its AI improves population health. Conversely, a useful improvement in scheduling or report turnaround need not produce a mortality benefit to have value.
The article leaves an essential business question open: Will these organizations remain attractive businesses if AI produces modest, incremental gains rather than dramatic productivity increases? That is a more demanding test than announcing an “AI-native” practice.
September 2026: Breakthrough designation becomes less useful as a substitute for comparative evidence.
“The last ‘breakthrough’ devices to get special treatment from Medicare”This report describes the withdrawal of preferential payment qualification pathways for future breakthrough-designated devices, with transitional provisions for qualifying products.
The central distinction is between a technology’s promise and its demonstrated clinical improvement. FDA breakthrough designation identifies potential; it does not itself establish that the marketed product improves outcomes over available alternatives.
For developers, the change increases the importance of planning comparative evidence early. For policymakers, it exposes a competitive trade-off. Stronger evidence requirements can improve purchasing discipline while also favoring organizations with enough capital, patients, and infrastructure to generate that evidence.
This connects directly to the articles on Epic and practice ownership. A policy intended to improve evidence can also influence market concentration. That does not make the policy misguided, but it means its competitive effects deserve attention alongside its evidentiary rationale.
October 2026: Clairity takes risk prediction directly to consumers.
“Why this startup is taking its AI to predict breast cancer risk directly to patients”Clairity’s $249 offering through Everlywell provides another answer to slow institutional adoption: let patients purchase the analysis themselves.
Its regulatory and clinical design supports that strategy. The output does not require a radiologist’s review, allowing telehealth providers to order the analysis of an existing mammogram and return a five-year risk estimate.
The distribution problem becomes easier, but the clinical handoff remains difficult. Patients must bring the result to their own clinicians, who must integrate it with existing risk information and determine appropriate management.
The article reports encouraging recognition of imaging-based risk assessment in NCCN guidance, alongside established risk-assessment approaches. That recognition does not by itself establish the incremental benefit of buying this particular test.
Consumer access therefore changes who finances adoption without resolving all the questions that slowed institutional purchasing. The result can reach the patient before the care pathway is fully developed. 2026 STAT 1013 Bigger Katie pal…
Taken together, the articles describe three simultaneous expansions: what AI can analyze, how it is distributed, and how its value is captured. Narrow finding detection is expanding toward comprehensive report generation and future-risk prediction. Distribution is expanding beyond conventional software sales toward practice ownership and consumer access. Revenue is being sought through separate reimbursement, existing professional payments, operational savings, downstream services, and patient cash.
These are related developments, but they should not be treated as one homogeneous market.
| Type of AI service | Principal source of value | Central question |
|---|---|---|
| Reporting and workflow assistance | Time, capacity, consistency, professional experience | Does it improve the complete workflow after checking and correction? |
| Diagnostic detection and triage | More accurate or earlier recognition | Does the additional finding change care appropriately? |
| Future-risk prediction | Better selection of screening or prevention | Does management based on the score improve outcomes relative to existing assessment? |
| Broad hospital surveillance | Earlier intervention across many patients | Does useful lead time outweigh alert burden and implementation costs? |
This distinction matters for evidence and payment. A tool that reduces transcription effort has a different evidentiary burden from a test that sends previously healthy people into additional screening. A system that improves an existing examination has a different business model from a new diagnostic service sold separately.
The deepest common theme is control over the clinical workflow. Epic’s installed infrastructure, Radiology Partners’ clinical scale, RadNet’s imaging network, and Everlywell’s consumer channel are different forms of access to the point where a product is used.
Algorithm developers can create value and still fail commercially if another organization controls deployment, interpretation, purchasing, and the customer relationship. Practice ownership is one response: bring the technology and its operational benefits into the same organization.
This is an important qualification to the conventional reimbursement discussion. Some AI businesses may thrive without ever securing a dedicated code. They can earn a return by making an already reimbursed service more efficient or by improving the practice’s ability to deliver it. Other products genuinely create a new service and need a separate financing mechanism.
The unit of innovation is becoming the clinical process. The coronary-calcium article shows why a score needs a prevention pathway. The sepsis article shows why a prediction needs timely clinician action. The Clairity article shows why a risk estimate needs a management plan. The practice-ownership article shows why an accurate model may remain commercially weak until integrated with scheduling, reporting, communication, and follow-up.
That suggests a change in how innovation should be evaluated. The relevant comparison is often between care delivered with the entire AI-supported process and care delivered without it. Model accuracy remains necessary, but cannot describe the total effect.
A second major theme is that computational abundance creates demands on scarce human attention. Running another analysis may be inexpensive. Reviewing its output, explaining uncertainty, investigating findings, and coordinating care may be costly.
This tension becomes more pronounced as foundation models identify a wider range of abnormalities. Greater breadth can improve care, but it also creates more opportunities for incidental findings and unresolved recommendations. The institution that generates the information may not be the institution that bears the resulting workload.
The same issue could become central in pathology. A tissue image may support multiple additional predictions, but each clinically reported prediction creates obligations concerning validation, interpretation, communication, and appropriate use. The availability of an inference does not establish that routinely reporting it is beneficial.
A third theme concerns who receives the gains from productivity. Faster reporting can support shorter turnaround times, more capacity, reduced overtime, greater professional satisfaction, or higher organizational margins. Those outcomes are not interchangeable.
An employer may convert time savings into additional reading volume. A physician may prefer to spend the time consulting with referring clinicians. A payer may eventually seek lower prices. A software vendor may capture part of the gain through licensing.
Consequently, an algorithm’s productivity benefit cannot be fully understood without examining contracts and organizational incentives. Even a technically successful product leaves a distributional question: who benefits from the time it saves?
The collection is strongest as a synthesis of industry change, rather than proof that the technologies collectively improve outcomes. Several articles follow overlapping companies and trends. Their recurrence should not be mistaken for independent replication. Company interviews, reported research results, policy proposals, and finalized decisions also carry different evidentiary weight.
Palmer’s reporting is especially effective at asking what happens after detection and who benefits financially. A useful counterweight is to avoid assuming that worthwhile innovation must reduce total spending. Improved health can justify additional expenditure. Similarly, administrative efficiency can be valuable without establishing a new clinical benefit.
The appropriate questions are whether the service improves something that matters, whether the improvement justifies its costs and harms, and whether a workable arrangement exists to finance it.
Foundation models also change the validation problem. A classifier with a defined target can be evaluated against that target. A generated report may contain many assertions, omit important findings, misstate comparisons, or describe the right abnormality in a misleading way.
The Cognita interview recognizes this difficulty. Its approach relies heavily on radiologist review rather than superficial similarity between generated and reference text. But evaluation should also consider the combined human–AI system: whether a plausible draft influences what the clinician notices, whether errors survive review, and whether time savings persist after adequate verification.
Clinical corrections are potentially valuable development data, but an edited report is not automatically a definitive ground truth. Systems need independent evaluation as well as feedback from routine use.
The most plausible direction is toward fewer integration points supporting more functions. This is an inference from the collection, not a settled forecast. Hospitals and diagnostic practices have limited capacity to adopt and govern a growing catalog of separate products. Broad platforms and comprehensive services can reduce that burden.
The result may favor large practice networks, established information-system vendors, and companies able to combine multiple useful functions. Independent developers may increasingly reach customers through partnerships, embedded products, or acquisitions.
There is a corresponding danger: distribution power can become more decisive than comparative quality. Easier integration is a real benefit, but it should not make external evaluation or switching suppliers impractical.
The most useful reforms would make evidence generation and adoption more predictable. Several priorities emerge.
Define the intended use and comparator early. Developers, clinicians, and payers should agree on the decision being improved and the relevant alternative. A new risk score must be compared with existing risk assessment; a report assistant must be evaluated against the actual reporting workflow.
Match evidence requirements to the claim. Time savings require credible workflow measurement. Earlier intervention requires meaningful lead time and evidence of appropriate action. New screening strategies require assessment of downstream benefits and harms.
Support adoption that generates interpretable evidence. Prospective evaluations, pragmatic comparisons, and time-limited payment arrangements can reduce uncertainty. Broad deployment without a credible comparison may produce large volumes of data without answering the central question.
Pay attention to implementation and responsibility. A clinical service should specify who reviews the result, communicates it, and arranges follow-up. These functions should be included in the economic assessment.
Preserve the ability to compare and replace models. Interoperability, portable data, and independent evaluation can help prevent an institution’s first integration decision from determining its options indefinitely.
Clarify regulatory responsibilities as business models converge. Internally developed tools, research deployments, commercially distributed software, and clinical services require careful distinctions. An exemption from establishment registration should not be casually treated as a blanket exemption from all applicable oversight. FDA materials describe the practitioner exemption specifically in the context of registration and listing. fda.gov
For pathology, the opportunity is to prepare for these organizational choices while building digital capability. Radiology’s experience suggests that scanning and algorithms alone will not determine success. Clinical validation, workflow integration, data rights, distribution, and payment classification will become strategic assets.
Several radiology developments have plausible pathology counterparts:
| Radiology development | Question for digital pathology |
|---|---|
| Practices acquire AI developers | Will laboratories or pathology networks internalize algorithm development? |
| Models draft broad diagnostic reports | How will pathologists verify image-based findings and integrated interpretations? |
| Existing scans yield new risk information | Which additional tissue-derived predictions warrant clinical reporting? |
| Platforms distribute multiple applications | Who will control access to the digital slide workflow and customer relationship? |
| AI improves productivity within existing payments | Which functions are operating tools, and which constitute distinct diagnostic services? |
The comparison should preserve the differences between the specialties. Pathology’s specimen preparation, staining, scanning, laboratory quality systems, and molecular workflows introduce their own constraints. Nor does a payment approach developed for radiology automatically fit a laboratory service.
The most consequential lesson is that the organizations best positioned to benefit from AI may be those that can connect validated interpretation with clinical delivery and capture the resulting value. For pathology leaders, the immediate strategic work is therefore broader than choosing an algorithm: it includes deciding which capabilities to own, which to purchase, how to validate them, and who will remain responsible for the diagnostic service delivered to the patient.
