AI in Healthcare Is Moving Faster Than Regulators Can Track. Here's What's Actually Deployed.

Healthcare AI has passed the pilot phase. Across radiology, pathology, primary care, drug discovery, and clinical administration, AI systems are not being evaluated — they are operating in production workflows, influencing diagnoses, and reshaping how clinical organisations staff and manage their operations. The regulatory frameworks designed to govern medical devices are processing a backlog of AI applications that grows faster than the approval pipeline can clear it. The result is a significant gap between what is deployed and what has been rigorously validated.

Where AI Is Already Embedded in Clinical Practice

Radiology has the deepest AI penetration of any clinical specialty. FDA-cleared algorithms for detecting pneumothorax, pulmonary embolism, intracranial hemorrhage, and breast cancer on imaging studies are running in hospitals across the US and EU. Several health systems have implemented AI triage systems that flag critical findings and route them to the front of the reading queue regardless of when the scan was ordered — a change that has measurably reduced time-to-treatment for conditions where hours matter.

Clinical documentation is the second major deployment area. AI scribing tools — ambient listening systems that transcribe and structure clinical encounters directly into the EHR — have been adopted by large health systems and independent practices alike. The time savings are substantial: physicians report recovering 60–90 minutes per day previously spent on documentation. Epic, the dominant EHR vendor, has integrated AI ambient documentation natively, which means deployment no longer requires a separate implementation project.

Drug discovery has seen the most dramatic scientific results. AlphaFold's protein structure predictions have become standard infrastructure for early-stage drug development across major pharmaceutical companies. AI-designed molecules are entering human clinical trials. The pace of target identification and lead compound generation has accelerated to a degree that is restructuring how pharma companies staff their discovery operations.

Where the Regulatory Gap Creates Risk

The FDA's 510(k) clearance pathway — designed for medical devices with a predicate device — was not built for AI systems that continuously update, perform differently across patient populations, and can degrade when input data distribution shifts. An AI algorithm cleared on one hospital's imaging data may perform significantly worse on another hospital's scans if scanner models, patient demographics, or imaging protocols differ. Post-market performance monitoring requirements for AI medical devices exist but are widely acknowledged as inadequate for detecting real-world performance degradation.

The algorithmic bias problem in healthcare AI is documented and serious. Multiple studies have shown that commercial AI diagnostic tools perform significantly worse on patients from underrepresented populations in the training data. A dermatology AI trained predominantly on lighter skin tones has meaningfully lower accuracy for darker skin tones. A sepsis prediction algorithm validated on a US academic medical centre population may miscalibrate when deployed in a community hospital with different patient demographics. These are not theoretical risks — they are documented outcomes from deployed systems.

What This Means for Healthcare Organisations Evaluating AI

The right question for any healthcare AI adoption decision is not "has this been FDA-cleared" but "has this been validated on a patient population and clinical environment that matches ours?" FDA clearance tells you the algorithm met a regulatory threshold at a point in time. It does not tell you it will perform well in your specific context.

Governance structures for AI oversight — clinical AI committees, performance monitoring protocols, human-in-the-loop requirements for high-stakes decisions — are becoming non-negotiable for responsible deployment. Organisations that deploy AI without these structures are taking on institutional and liability risk that will become apparent when a performance failure occurs.

Practical takeaway: If you are evaluating AI tools for a healthcare application, require the vendor to provide disaggregated performance data by demographic subgroup and to describe their post-market surveillance methodology. Both requests will tell you a great deal about whether the vendor has thought seriously about real-world deployment.