What Clinical AI Actually Looks Like

In the radiology department of a London teaching hospital, a stroke patient's CT scan arrives at 11 PM. The radiologist is on call, handling four other cases. An AI system flags “large vessel occlusion—likely LVO candidate.” The radiologist glances at the summary. It takes 90 seconds to confirm.

That 90 seconds used to be 12 minutes. This is not a science fiction outcome. Real-world use of Viz.ai stroke detection software shows a 39.5-minute reduction in the time neurointerventional radiologists are notified of high-risk patients, translating to documented 40% reduction in disability at 90 days. This is clinical AI that works.

But it works in a very specific way. And understanding the difference between that narrowness and the broader hype is essential for anyone deploying, investing in, or regulating clinical AI.

‍

Where AI Actually Wins

The FDA has authorized 1,104 AI-enabled radiology devices, comprising 76% of all FDA-cleared AI/ML devices through December 2025. The concentration is not accidental. Radiology is a problem AI handles well: high-volume, standardized inputs, clearly defined outputs, and a clinical workflow where an alert can slot cleanly into existing triage.

The evidence is mixed but real. Studies show AI-assisted radiologist performance improves sensitivity without degrading specificity for chest radiograph abnormalities, with some deep learning models achieving AUC of 94% in pneumothorax and pulmonary edema detection, outperforming human readers in isolated tasks. AI is measurably good at spotting what humans miss on CT, chest X-ray, and mammography when the imaging is high-quality and the pathology is clear.

But here is what the press releases omit: AI catches things radiologists miss. AI also misses things radiologists catch. More importantly, the gap between a model that performs well in validation and a tool that changes clinical outcomes remains vast.

‍

The Last-Mile Problem

Between a validated algorithm and a deployed system that improves patient outcomes lies clinical workflow. This is where AI breaks.

Research on AI implementation in clinical workflows shows that clinician adoption ranges from 45% to 89% of eligible cases, with alert override rates documented at 12% to 38% depending on implementation context. Put another way: a third of AI alerts are ignored or actively overridden. This is not a sign that the algorithm is wrong. It is a sign that the workflow is broken.

Why? Start with the basics. Most clinical AI is deployed as either EHR-embedded systems or standalone dashboards. EHR embedding seems logical—the alert appears where the clinician already is. But it competes with dozens of other alerts for attention. A radiologist reading 40 cases in a shift does not stop for every flagged abnormality. She trusts the system only after repeated exposure to its signal-to-noise ratio. Standalone systems are worse: they require context-switching, separate login credentials, and the assumption that a clinician will consult an auxiliary tool in the middle of established workflow.

The Viz.ai stroke model works because it solves this differently. The alert is not a suggestion. It is a dispatch order that integrates with the hospital's intervention team activation protocol. The AI is not advising the radiologist to look again. It is triggering a cascade that involves the interventional suite, neurosurgery, and transport. In other words, Viz.ai succeeded not because its algorithm is brilliant, but because it treats workflow integration as product architecture, not afterthought.

‍

Trust, Explainability, and the 3 AM Problem

The deeper bottleneck is trust. A radiologist with 20 years of experience does not defer to a black-box algorithm, especially when liability rests with the clinician, not the vendor. Studies on AI-clinical workflow integration show that concerns about loss of autonomy, unreliable user interfaces, and lack of skilled personnel are major barriers to adoption.

Consider the understaffed emergency department at 3 AM. A junior physician, exhausted and running on coffee, sees an AI recommendation. Does she trust it? Only if: (1) she understands what the model is claiming—not “risk of pneumonia” but “65% confidence of basilar infiltrate on left side”; (2) she has seen the model's failure modes—what does it miss? what does it falsely flag?; and (3) she has a clear decision rule for what to do if the AI disagrees with her clinical impression.

None of this is built by the machine learning team. All of it is built by clinical implementation specialists, change management, and training.

‍

The Real Constraint

There is no shortage of clever algorithms. In 2025 alone, the FDA cleared 72 AI-enabled devices in Q4, with 76% in radiology. Most of these will never meaningfully change clinical practice. The ones that do will be the ones where someone invested in the hard work of embedding AI into clinical workflow, not as a research project but as operational infrastructure.

This is why the companies winning in clinical AI are not the ones with the best papers. They are the ones with clinical operations teams, health system relationships, and the patience to work through the unglamorous process of getting a tool adopted in a hospital actually organized to use it.

The algorithm is necessary. The algorithm is not sufficient. This is the chasm between the AI that impresses at conferences and the AI that saves lives at 3 AM in an understaffed unit.