Docket ID No. OSTP-TECH-2025-0067
Download this comment (PDF)
Introduction
The White House Office of Science and Technology Policy (OSTP) issued a Request for Information (RFI) to identify laws and regulations that may hinder the development or deployment of Artificial Intelligence (AI) technologies in the United States. This comment responds to OSTP’s RFI, specifically explaining how existing Centers for Medicare and Medicaid Services (CMS) regulations on medical billing may disincentivize the adoption of AI tools that may augment clinical decisionmaking.
The growing adoption of clinical AI tools reflects the progress made on the nascent technology. Indeed, the number of AI-enabled devices approved by the FDA has risen from two in 1995 to 235 in 2024, a significant shift driven by the increasing availability of computing power, data, and investments (FDA, 2025; Beam and Kohane, 2018). These devices are being developed and implemented for everything from administrative purposes, such as clinical note-taking and post-operative follow-ups, to assisting with diagnostic tasks such as skin cancer detection and CT/MRI interpretation (Greene, 2025; Bian et al., 2020; Al-Karawi et al., 2024; Katal et al., 2024; Krakowski et al., 2024).
The extent to which these devices have been adopted, however, has varied significantly. A single site study based on one health system found that the most commonly adopted technology was clinical documentation assistance (100% adoption rate); however, technologies that assist with diagnostic procedures are not being fully deployed (Poon et al., 2025). Another study from the American Medical Association found that while 68% of physicians recognized some value of AI in the clinical workplace and 66% used it in some capacity, six out of the ten top use cases were for administrative tasks, not diagnostic ones (AMA, 2025).
The reasons for the lack of adoption are fairly clear. One study found that 47% of survey respondents cited financial concerns and 40% cited regulatory uncertainty as reasons for not adopting AI tools (Poon et al., 2025). Other studies have found a lack of resources for the implementation of the technology and poor billing structure leading to perverse incentives to adopt (Hassan et al., 2024; Parikh, & Helmchen, 2022). These trends underscore an important point regarding the adoption of AI in healthcare. While clinicians are readily adopting technologies that assist in administrative tasks that may reduce the amount of time that doctors spend on paperwork and data entry rather than patient care, there is a lack of adoption of technologies that augment clinical tasks that may assist with patient outcomes or treatment decisions.
AI Technologies in Healthcare
Healthcare AI technologies are being developed for a wide range of applications; however, their diffusion across the sector has been uneven. To understand how policy and reimbursement structures affect this variation, it is useful to distinguish among the major types of tools currently in use. AI applications in healthcare generally fall into three categories: administrative tools that automate documentation and coordination; generalist clinical tools that synthesize diverse medical data to support clinical decisions; and specialized clinical tools that perform discrete diagnostic functions.
Administrative Tools
AI tools for administrative support, particularly clinical documentation, have shown the clearest and most consistent benefits in healthcare settings. These systems automate paperwork and note-taking, allowing clinicians to spend more time on patients. The benefits of this have been confirmed across studies, including increased attention paid to patients, lower reported burnout, and a reduction in after-work-hours note-taking (Olson et al., 2025; Duggan et al., 2025). These tools have also demonstrated small improvements in accuracy, with one study finding that accuracy improved by 5 percentage points when using AI transcribers (Chomutare et al., 2025). These tools have also been used at an administrator level. Evidence has shown measurable efficiency improvements when these tools are used in hospital-level decisionmaking, such as improving data integration, patient management, and cross-department coordination, including reductions in patient wait times and reductions in elective surgery queues (Russo et al., 2025; Alves et al., 2024).
Generalized Clinical Tools
AI tools that are intended to use and understand multiple types of medical data such as text, images, and laboratory data are playing a broader role in healthcare. These generalist models are designed to handle multiple clinical tasks through one interface, such that they may be able to interpret a chart, review a scan, and suggest possible diagnoses. For example, Google Research’s Med-PaLM 2, a large language model trained on medical literature and the PaLM-E vision-language model, can interpret clinical questions and produce responses through a chain of retrieval process that verifies claims against medical literature. In practice, this may include assistance with diagnoses, including radiology reports and imaging summaries of X-rays or CT scans, reviewing laboratory data, or summarizing patient histories and medication interactions (Singhal et al., 2023).
Specialized Clinical Tools
AI tools designed for single, well-defined tasks represent the most established clinical applications of the technology. These tools are typically trained and validated for one diagnostic function. Applications, however, vary significantly across systems. For example, the recently launched IDx-DR system is the first autonomous diagnostic tool that can be used to screen for diabetic retinopathy in primary care settings (Abràmoff et al., 2018). In cardiology, HeartFlow FFR-CT algorithm improved the identification of patients with coronary artery disease and reduced unnecessary invasive testing (Nørgaard et al., 2014). In dermatology, image-based AI systems have been used for distinguishing between benign and malignant skin lesions and have performed as accurately as doctors (Esteva et al., 2017). This has allowed for screening without on-site dermatologists.
Billing for Healthcare Services
Healthcare reimbursement in the United States is primarily focused on the delivery of services. Under the fee-for-service (FFS) model, providers are reimbursed for each instance in which a service is provided, with payments determined by Current Procedural Terminology (CPT) codes for procedures and International Classification of Diseases (ICD-10) codes for diagnoses. These payments are calculated based on the relative value unit (RVU), which is a standardized measure used by CMS to quantify resources associated with medical services (Zuckerman et al., 2016). Each CPT code corresponds to an assigned RVU, which is composed of three components, including work RVUs, practice expense RVUs, and malpractice RVUs. Work RVUs represent the physician’s time, skill and effort required to perform a service, and they form the largest portion of total payment. Practice expense RVUs cover the costs of maintaining a practice, including staff, equipment, rent, and supplies, and vary depending on whether services are performed in facility or non-facility settings. Malpractice RVUs account for professional liability insurance costs and generally contribute a small share of total reimbursement.
Each RVU component is adjusted with a geographic practice cost index to ensure it is in parity with local cost variations and then by a conversion factor, which is a dollar amount updated annually by CMS to determine final payment (Sloan, 2011). This formula ties a large portion of reimbursement to human labor and tangible resource inputs, which has important ramifications for the adoption of clinical AI technologies.
Because RVUs were designed to quantify physician work and operating costs, they overemphasize manual and time-intensive labor while undervaluing efficiency gains. Given that clinical AI tools typically operate in ways that reduce physician time rather than increase billable events, they fail to neatly fit into any RVU category. For example, if an AI device were categorized as a work RVU, its efficiency would justify lower reimbursement since less clinician time would be required per encounter (Katz & Melmed, 2016). Of course, this may mean that a physician would be able to see more patients to bolster their total reimbursement; however, this could inadvertently lead to burnout as a result of overwork or create incentives for overutilization, as physicians increase patient volume to maintain income despite lower per-encounter payments. With evidence indicating a positive association between physician trust and time spent with patients, such a dynamic could make patients less likely to seek care in the first place (Keating et al., 2004; Croker et al., 2013).
If it were treated as a practice expense, it would be absorbed into overhead because CMS treats such costs as fixed inputs to care delivery rather than as revenue-generating activities, meaning that the savings or quality improvements from AI adoption would not translate into higher reimbursement (Burgette et al., 2021). In other words, while AI systems may reduce costs and improve outcomes by shortening hospital stays, preventing readmissions, or improving the accuracy of diagnoses, the reimbursement structure inadvertently favors an increase in the volume of claims over integrating these tools into the medical process. This leads to a misalignment between innovation in AI systems and reimbursement.
Recommendations
Create New CPT Codes for AI-augmented care
CMS should establish a new category of AI-augmented CPT codes that explicitly accounts for productivity and outcome-improving AI technologies. This would require that CMS develop a new AI-modifier that is appended to existing CPT codes to identify when an FDA-approved AI tool contributed to a diagnosis or treatment. CMS should additionally incorporate an AI efficiency adjustment in its RVU calculations to reflect improvements in efficiency, rather than only physician time. This would likely require collaboration with AMA’s CPT Editorial panel to define the proper criteria for AI-eligible procedures. This approach would align reimbursement with gains in diagnostic and efficiency performance rather than human labor alone.
Develop Value-Based Payment Models that Incentivize AI Adoption
CMS should establish an AI demonstration program through the Center for Medicare and Medicaid Innovation to test value-based payment models that focus on outcome improvements associated with the use of AI. This could include 1) shared savings models that allow providers to retain a portion of the cost savings from AI-enabled improvements; 2) outcome-based reimbursement metrics that focus on diagnostic accuracy and patient safety; or 3) data-sharing partnerships between CMS, health systems, and AI developers to track quality and cost outcomes across pilot sites. By tying reimbursement to outcomes rather than services, this approach aligns provider incentives with the benefits of AI adoption. It would also provide AI developers, health systems, and CMS with evidence on the fiscal and clinical effects of AI adoption prior to the full-scale adoption of such products.