A study published August 19, 2026, in PLOS Digital Health dropped a finding that should stop every pharmacist working in a technology-enabled care environment. Of 1,357 AI and machine learning-enabled medical devices that have received FDA clearance or approval, only 34 were linked to registered prospective clinical trials, and only three had been evaluated for patient-centered outcomes such as mortality, morbidity, or hospital readmissions.
Three. Out of 1,357. That is a 0.2% clinical validation rate for a category of tools that 71% of U.S. hospitals are now actively deploying in clinical workflows.
What the Study Actually Found
The research, led by investigators at the Stanford Center for AI in Medicine and Society and published in PLOS Digital Health, reviewed the entire FDA-cleared AI and machine learning medical device database against prospective clinical trial registrations.
The findings: 1,357 AI/ML-enabled medical devices have received FDA clearance or approval. The majority cleared through the 510(k) pathway, which requires demonstration of substantial equivalence to a predicate device but does not require evidence of clinical effectiveness. Of those 1,357 devices, only 34, representing 2.5%, were linked to registered prospective clinical trials. Only three devices had been evaluated for patient-centered outcomes: mortality, morbidity, or hospital readmissions.
The authors noted that clinical decisions increasingly depend on algorithmic outputs. Yet despite rapid adoption, one question remains largely unanswered: do these tools improve patient outcomes? They called specifically for revised authorization requirements, stating that FDA clearance should require prospective evidence of clinical effectiveness and equity before deployment, not as an afterthought after years of clinical use.
The clinical domains where AI/ML devices are most heavily concentrated: radiology and pathology for diagnostic imaging, cardiology for ECG and monitoring applications, and clinical decision support for medication management, sepsis prediction, and readmission risk. The pharmacy-adjacent tools, including drug interaction checkers, dosing recommendation systems, medication reconciliation automation, and adverse event prediction algorithms, fall directly within the device categories most deployed without outcomes validation.
Why This Is the Most Important Clinical Signal for Pharmacists Right Now
Every hospital deploying AI-assisted clinical decision support is using tools that have been cleared by the FDA but not necessarily validated against the question that determines patient value: does the patient do better because of this tool?
The sepsis prediction algorithm flagging a patient for early intervention may reduce sepsis mortality, or it may generate a false positive rate that causes clinician alert fatigue and leads to missed true positives. Without a prospective outcomes study comparing the AI-assisted workflow against the prior workflow, the health system deploying the algorithm cannot answer that question with evidence.
The AI-generated drug dosing recommendation adjusting a medication for renal function may produce better therapeutic outcomes than manual dose adjustment, or it may produce equivalent outcomes with more clinician time spent reviewing and overriding alerts. Without outcomes data, the institution cannot know.
The AI medication reconciliation tool identifying likely discrepancies at discharge may reduce readmissions, or it may identify the same discrepancies that a pharmacist would have caught through manual review, producing equivalent outcomes at higher technology cost and lower pharmacist skill utilization.
In all three cases, the gap between FDA clearance and outcomes validation is the information void that clinical judgment must fill. And the clinician most qualified to fill it, in any medication-related AI decision support domain, is the pharmacist.
The Human in the Loop Role That AI Structurally Requires
The Autonomous Pharmacy Framework, covered in this newsletter earlier this year, describes pharmacy’s five-level automation trajectory through the lens of self-driving car analogy. At Level 3 and above, AI handles increasingly complex tasks and the human’s role shifts from execution to oversight, from doing to judging.
The 1,357-device, 3-outcomes-validated reality is a description of where clinical AI sits today: tools deployed at Level 3 confidence that have Level 1 outcomes evidence. The gap between deployment confidence and outcomes evidence is exactly the space where pharmacist clinical judgment creates irreplaceable value.
The pharmacist who reviews an AI drug interaction alert is not simply accepting or overriding a suggestion. They are performing a clinical judgment that the tool itself cannot perform: is this alert clinically significant for this specific patient, at this specific dose, with this specific comorbidity profile, at this specific point in their care trajectory? That judgment requires the complete patient picture that no algorithm currently integrates as fluidly as a trained pharmacist does at the point of care.
When that pharmacist overrides the alert and documents why, they create two simultaneous outputs: a better clinical decision for the patient in front of them, and a data point in the institution’s AI performance tracking system that informs whether the tool’s alert logic should be refined. The documented clinical override is the feedback loop that makes AI tools better over time. Without it, the institution is deploying a tool and hoping it improves outcomes rather than measuring whether it does.
This is the pharmacist-as-AI-governance-layer argument. It is not theoretical. It is the operational description of what happens every time a pharmacist reviews a CDS alert, weighs it against their clinical judgment, and acts. The difference between a pharmacy team that treats that review as a routine workflow step and one that treats it as a governance function is documentation, measurement, and the institutional acknowledgment that pharmacist clinical judgment is the validation layer that makes AI-assisted clinical decisions trustworthy.
The Equity Dimension the PLOS Study Specifically Raised
The researchers called for prospective evidence of both clinical effectiveness and equity before AI deployment. The equity framing matters for pharmacists in a specific and often overlooked way.
AI/ML tools trained on historical clinical data inherit the patterns of that data, including patterns of under-documentation, diagnostic delay, and therapeutic inequity that have historically affected Black, Latino, Indigenous, and low-income patient populations. An AI sepsis prediction model trained on data from a predominantly white suburban hospital may perform differently in an urban safety-net hospital serving a different demographic population. An AI drug dosing recommendation tool trained on clinical trial data that under-enrolled older women or patients with complex multimorbidity may generate systematically different recommendations for that population.
The pharmacist who understands algorithmic bias, who questions whether an AI tool’s performance has been validated in the specific patient population their institution serves, and who flags systematic over-alerting or under-alerting patterns in specific demographic groups is performing a health equity function that almost no current health system AI governance structure formally assigns to anyone.
That function, formalized, is an AI equity auditor role. It requires clinical training in the drug domain being monitored, familiarity with the demographic composition of the patient population, and systematic documentation of performance differences across patient subgroups. Pharmacists with medication expertise, patient population knowledge, and outcome documentation skills are the natural candidates for that role.
The FDA Regulatory Trajectory That Will Accelerate This
The PLOS study’s call for revised authorization requirements coincides with ongoing FDA activity on AI/ML regulatory frameworks that pharmacists in health system and clinical informatics roles should track.
The FDA has been developing an action plan for AI/ML-based software as a medical device since 2021, with ongoing updates reflecting the agency’s recognition that the current 510(k) pathway was not designed for continuously learning AI systems that may perform differently after deployment than at the point of clearance. The agency’s predetermined change control plan framework allows developers to describe in advance how their AI tool will learn and update post-clearance but does not yet require post-market outcomes validation as a condition of maintained clearance.
The PLOS study’s finding that 97.5% of cleared AI devices have no linked prospective trials provides the specific evidence base that regulatory reform advocates will use in FDA engagement in 2027 and beyond. The pharmacist who is fluent in this regulatory trajectory, who understands the difference between FDA clearance and clinical outcomes validation, and who can communicate that distinction to health system leadership, is providing a level of institutional intelligence that very few clinical staff currently deliver.
The AI Governance Role Being Built Right Now
Formal AI governance is becoming an organizational function in health systems, driven by the same evidence accumulation that this PLOS study represents. Hospital systems including UCSF, Mayo Clinic, and Stanford have established AI governance committees and, in some cases, dedicated AI safety or ethics roles specifically focused on clinical AI deployment.
The pharmacist is uniquely positioned to serve as an ethical steward of AI use in healthcare, ensuring ethical, transparent, and patient-centered implementation of AI tools that influence both clinical care and operational decision-making.
That positioning has a specific practical basis. Pharmacists already interact with more AI-assisted clinical decision support tools, across more patient touchpoints, than almost any other clinical role in a health system. Every medication order that passes through clinical decision support is a pharmacist touchpoint. Every drug interaction alert, every dose adjustment recommendation, every formulary substitution suggestion, every contraindication flag is a data point in an AI system’s performance record. The pharmacist who systematically reviews, overrides, documents, and tracks those interactions across a patient population is generating the kind of post-market performance data that the PLOS study found almost no one is currently collecting.
Building that data collection function into existing pharmacy workflows and presenting it to quality leadership as an AI outcomes evaluation protocol, is the specific proposal that gets pharmacists into AI governance roles.
Your Action This Week
Find out which AI-enabled clinical decision support tools are currently deployed in your institution or pharmacy workflow. The list will likely include some combination of: electronic medication reconciliation assistance, drug interaction checking, dosing recommendation support, sepsis or deterioration prediction alerts, readmission risk scoring, and formulary optimization suggestions.
For each tool on your list, ask one question: what is the institution’s current process for evaluating whether this tool is producing better patient outcomes than the approach it replaced?
If the answer is “we track alert acceptance and override rates,” that is performance monitoring, not outcomes validation. If the answer is “we have not compared 30-day readmissions, adverse drug events, or length of stay between the AI-assisted and pre-AI periods,” that is the gap the PLOS study documented at national scale reproduced in your specific institution.
Propose filling it. An AI outcomes evaluation protocol for medication-related clinical decision support tools, designed and led by pharmacy, requires: a pre-implementation baseline for the outcomes metric of interest, a defined post-implementation measurement window, a method for comparing outcomes between patients touched by the AI tool and a matched comparison group, and a reporting cycle that feeds results to the quality and technology governance committees.
That proposal does not require a data science team or an informatics budget. It requires a pharmacist who understands clinical outcomes measurement and who has access to the basic utilization and outcome data that every health system already collects.
The pharmacist who raises this is not raising a problem. They are demonstrating the clinical and technological governance leadership that health system administrators and CMOs are actively looking for, and that most of their pharmacy colleagues haven’t yet thought to offer.
The tools are deployed. The outcomes are unvalidated. The governance role is open.
The pharmacist who fills it first will define it.
Sources: PLOS Digital Health (Crigger E et al. Prospective Clinical Trials of AI and Machine Learning Medical Devices: A Cross-Sectional Study. PLOS Digit Health. 2026 Aug 19. doi:10.1371/journal.pdig.0000783), STAT News (Most AI Medical Devices Lack Clinical Proof They Work, Study Finds, August 19, 2026), Health IT Analytics (PLOS Study: Only 3 of 1,357 FDA-Cleared AI Devices Tested on Patient Outcomes, August 2026), The Verge (Doctors Are Using AI Tools That Haven’t Been Tested on Patient Outcomes, August 2026), FDA (Artificial Intelligence and Machine Learning in Software as a Medical Device, Action Plan 2021-2026), Pharmacy Times (Pharmacists as Ethical Stewards of AI in Healthcare, 2026), ASHP (Practice Advancement Initiative 2030: AI Standards Recommendation), Drug Topics (Total Pharmacy Solutions Summit Coverage: AI in Pharmacy Governance, June 2026), Nature Medicine (AI Clinical Decision Support: Post-Market Performance and Bias Considerations, 2025)