HHS Could Use a Safe Harbor to Move Validated Hospital AI Faster
Clinical AI tools are already capable of changing bedside care, but many hospitals still cannot afford to deploy them at scale. The barrier is financial rather than scientific: validated software often exists, yet the payment structure lags far behind the pace of FDA review and hospital purchasing cycles. A targeted HHS safe harbor could close that gap by letting life sciences and technology companies fund adoption under strict guardrails, as long as the tools remain product neutral, evidence based, and free from sponsor control over clinical recommendations.
The case for doing so is not theoretical. A stroke-detection AI that identifies hemorrhage or ischemia on CT and routes the alert to the right team is now embedded in routine care at many hospitals, in part because Medicare reimbursement arrived early enough to support adoption. In a study spanning 166 hospitals and more than 14,000 patients, the technology shortened time to treatment by nearly 40 minutes, a 44% reduction. In stroke care, minutes matter because neuronal loss accelerates quickly once perfusion drops or a large vessel is occluded. Software that compresses the diagnostic-to-intervention interval can change whether a patient walks out of the hospital or leaves with permanent disability.
A different clinical problem shows what happens when the infrastructure is missing.
Hypertrophic cardiomyopathy can sit behind a normal-looking ECG for years, then surface only after syncope, arrhythmia, or sudden cardiac arrest. AI-based tools can now flag suspect patients on routine electrocardiograms before symptoms appear, yet diagnosis still fails far too often; roughly 60% of cases are misdiagnosed, and the diagnostic odyssey frequently includes four incorrect labels before the right one is reached. The biological logic is straightforward: the signal is there in the data, but the pathway to act on it is incomplete. Without reimbursement, hospitals may recognize the utility of a tool and still leave it unused on the shelf.
CMS has already acknowledged the need for a better financing bridge. Its proposed Software as a Medical Service category would standardize reimbursement for AI tools that interpret scans, electrocardiograms, or clinical charts and return a diagnosis or risk score. The agency has also created a dedicated AI office and is testing outcome-based payment models for technology-enabled chronic care. The first wave includes 36 new codes. Even so, rulemaking moves slowly, and claims-data cycles move even more slowly. AI tools are being cleared faster than payment policy can keep up, leaving a growing gap between regulatory approval and real-world deployment.
Hospitals have lived through this kind of lag before. CMS created the New Technology Add-on Payment program to provide temporary reimbursement for selected new drugs and devices while a more durable pathway developed. That bridge was never designed for the current volume of AI tools, and CMS has proposed retiring it. If that happens without a replacement, one of the few mechanisms that helps validated technologies cross the adoption gap disappears just as the pipeline of cleared products expands.
Federal regulators have already solved analogous problems in other domains. Since 2006, HHS has protected donated electronic health record software through a dedicated safe harbor and Stark Law exception, recognizing that accelerating digital infrastructure could benefit patients. In 2020, a similar approach extended to donated cybersecurity technology. Hospitals now depend on both forms of donated infrastructure at scale, and neither arrangement is treated as an Anti-Kickback Statute violation when the guardrails are clear. AI clinical pathway software fits the same policy logic when the tool supports medical decision-making rather than commercial steering.
A workable safe harbor would need three limits. First, the tool could not favor one company’s drug, device, or service over another’s. Second, its recommendations would have to be grounded in peer-reviewed evidence or established clinical guidelines. Third, the sponsor could not shape the output the clinician sees. In practice, that means a company could help pay for the infrastructure, but it could not purchase influence over what a clinician is told to prescribe, order, or escalate. For clinicians, the distinction protects independent judgment. For patients, it protects the integrity of the clinical pathway.
The equity implications are hard to ignore. Rural and critical access hospitals report less than 50% adoption of even basic predictive tools, while top-tier system-affiliated hospitals reach 86%. These are precisely the institutions that can gain the most from AI triage, chart review, and risk stratification, yet they are also the least able to absorb the upfront cost. Every year spent waiting for payment policy to catch up widens that divide. The result is not neutral delay; it is uneven access to software that can help identify stroke, cardiomyopathy, sepsis, deterioration, and other time-sensitive conditions sooner.
HHS has the authority to act through the same rulemaking process used for earlier safe harbors: propose a rule, take comment, and finalize it. The clinical evidence for many of these tools is not the primary obstacle. Financing is. Until payment catches up with validation, hospitals will continue to face a strange mismatch: software that can improve care in real time, but only if someone can pay to bring it through the door.