Seventy-one percent of U.S. hospitals were running predictive AI integrated into their electronic health records by 2024, according to ONC’s Hospital Trends data brief. That number moves faster than any IND, any protocol, and any regulatory framework currently on the books. The deployment velocity is real. The evidentiary foundation underneath it is considerably less certain.
The signal arrived quietly in Nature Medicine’s recent analysis of AI decision support scaling. The core question it poses is deceptively simple: can evidence generation keep pace with AI adoption in clinical settings? Buried inside that question is a problem with direct consequences for every sponsor running a trial that touches an AI-assisted clinical workflow, every CRO validating data collected through an AI-augmented EHR, and every regulatory affairs team preparing a submission that includes real-world evidence generated in those environments.
The Evidence Gap No One Is Budgeting For
Consider what a large real-world trial actually found when it tested this assumption at scale. A randomized study of over 9,600 patients across 16 primary care clinics in Kenya evaluated “AI Consult,” a generative AI-powered clinical support tool integrated into an electronic medical record system. The result: clinicians using the AI tool demonstrated improved quality of decision-making. Short-term patient outcomes did not significantly change. Read that twice, because it is the contradiction at the center of the entire AI decision support debate.
Better process. Same outcomes. That gap between what AI changes at the point of care and what it changes in measured patient results is precisely where evidence generation breaks down for sponsors.
If an AI decision support tool modifies clinician behavior during a trial without producing a detectable signal in the primary endpoint, you have a confounder that your protocol almost certainly did not anticipate. If that same tool is embedded in the site’s EHR system and feeding data into your eClinical stack, your source data verification process has a new variable that your monitoring plan was not designed to interrogate. The tool is live, the trial is running, and the evidentiary framework for what the AI actually changed is still being written.
The FDA’s response to this environment has been structural rather than prescriptive. In December 2024, the agency published final guidance titled “Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions,” subsequently updated in August 2025. The guidance addresses lifecycle management of AI/ML-enabled device software functions. What it does not address with operational specificity is how a sponsor or site should document, flag, or account for AI decision support modifications occurring in the clinical environment during an active trial. That interpretive gap is the operational problem sponsors are now inheriting.
Who Gets Caught in the Crossfire
Oncology and cardiology sponsors running multi-site trials at academic medical centers face the most immediate exposure. These institutions have the highest AI adoption rates, the most complex EHR integrations, and the greatest likelihood that an FDA-cleared AI decision support tool is operating in the background of your site visit. Aidoc, for instance, received FDA 510(k) clearance for its CARE AI foundation model, a comprehensive triage solution that operates across radiology and clinical decision workflows. When that system is running at a site, it is influencing the clinical gestalt that generates your source data. Your monitoring plan does not know it is there.
Decentralized and hybrid trial designs are exposed from a different angle. When a patient is interacting with an AI-assisted telehealth platform, an AI-augmented ePRO tool, or an AI-powered symptom checker between site visits, the behavioral and clinical data flowing back to your EDC is already shaped by an upstream AI layer. The data flow integrity challenges in AI-assisted clinical trials documented by regulatory compliance specialists at ELIQUENT Life Sciences include AI tools being used for protocol deviation assessment, data cleaning, and risk signal identification in active trials. Each of those applications touches the audit trail your inspector will walk through.
The counterintuitive reality here is this: the sponsors who believe AI decision support is only a commercialization problem, something to manage after approval, are the ones who will find it embedded in their Phase 3 data packages. Adoption at the site level does not wait for the sponsor’s validation timeline.
The Operational Directive
If you are a clinical operations leader overseeing a multi-site trial at institutions using AI-integrated EHR systems, your site feasibility questionnaire needs a new section. Before randomization begins, you need documented answers to three questions at every site: which AI decision support tools are active in the clinical workflow, whether any of those tools interact with the data fields feeding your EDC, and what change control processes the site has in place if those tools are updated mid-trial. The FDA’s December 2024 Predetermined Change Control Plan guidance establishes that AI/ML-enabled software functions are expected to have managed modification pathways. Your sites should be held to an equivalent standard of disclosure. If they cannot answer these questions, treat it as a protocol deviation risk category, not a vendor relationship issue.
For sponsors relying on real-world evidence generated in AI-augmented clinical environments, the pre-registration requirement the FDA has articulated for RWE credibility becomes even more load-bearing. The study design must account for AI tool use at data collection sites as a potential effect modifier, or the resulting evidence will not survive a serious methodological challenge at the advisory committee level.
Watch the comment period for any follow-on FDA guidance addressing AI in clinical trial operations specifically. The August 2025 update to the Predetermined Change Control Plan guidance touched AI/ML device functions but left the clinical trial intersection largely unaddressed. The agency’s AI/ML-based Software as a Medical Device Action Plan, first published in January 2021, identified post-market surveillance and adverse event reporting for AI tools as an open action item. Five years later, that action item has not produced trial-specific operational guidance. The next draft that lands in the Federal Register on this topic will carry immediate protocol implications for any sponsor running trials at AI-heavy institutions, and the window to shape it through public comment will be short.
References
- Nature Medicine — “AI decision support is scaling-up fast — can the evidence keep up?”
- ONC Health IT — “Hospital Trends in Use, Evaluation, and Governance of Predictive AI, 2023–2024”
- EurekAlert — “AI Consult trial: 9,600 patients across 16 primary care clinics in Kenya”
- Berkley Life Sciences — “FDA AI/ML SaMD Framework: Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions (December 2024, updated August 2025)”
- Aidoc — “FDA 510(k) Clearance for CARE AI Foundation Model”
- Clinical Leader / ELIQUENT Life Sciences — “Understanding and Preserving Data Flow Integrity in AI-Assisted Clinical Trials”
- Intuition Labs — “FDA AI/ML-Based SaMD Action Plan (January 2021) and Regulatory Framework Development”
Moe Alsumidaie, MBA, MSF, is founder and Chief Editor of Vanguard Publications, which publishes Clinical Trial Vanguard, Pharma Vanguard and BullScope, and Head of Research at CliniBiz. He has two decades in clinical trial operations and data science, with earlier roles at Genentech, Abbott Vascular and Stanford University Medical Center, and is a guest lecturer in clinical trial sciences at Rutgers University.
