Will AI replace radiologists?

AI isn’t replacing radiologists today. In real hospitals, AI tools still play it safe. They help prioritize scans, flag urgent cases, and spot narrow issues. A radiologist still reads the image, signs the report, and carries the risk if something’s missed.

But that story is starting to shift, and investors should pay attention.

On March 25, 2026, NYC Health + Hospitals CEO Mitchell Katz said his system could replace “a great deal” of radiologists with AI if regulators allowed it. His vision is simple. AI does the first read. Humans step in only when something looks abnormal.

That matters because of who said it. This isn’t a startup making a pitch. It’s the largest municipal health system in the U.S., serving over one million patients a year across more than 70 locations. When an operator at that scale starts talking about cutting labor instead of just adding software, you’re looking at a budget decision, not a tech experiment. That same logic fits into a broader trend where major platforms are starting to put healthcare AI tools directly in front of patients and providers.

Radiologists aren’t going away. But the role is starting to narrow in a meaningful way. Hospitals under cost pressure are asking a different question now. How much routine image volume can AI handle before a human ever gets involved?

Why normal-screen triage is becoming the margin control point

The commercial logic is easy to overlook.

Screening workflows contain huge pools of negative studies.

A hospital does not need software that can do all of radiology to change the labor equation. It needs software that can safely keep a meaningful share of likely normal studies from taking the same specialist time. Once that becomes possible, the scarce asset changes. The valuable layer is the gatekeeper deciding which scans reach a radiologist for full review.

In a 2026 paired noninferiority trial published in Nature Medicine, an AI-supported breast-screening workflow classified low-risk cases as normal and routed the rest to human readers using AI support. Radiologist workload fell 63.6%. Cancer detection rose from 6.3 to 7.3 per 1,000 women screened, though recall rates also increased. That does not prove universal autonomy. It does show that, inside a tightly bounded screening workflow, software can materially change how much human reading labor is needed.

A 2025 Nature Medicine paper on population-based mammography screening reported that prospective studies were already showing AI-enabled workflows could improve screening metrics and reduce reading workload. The buyer’s question has shifted. It used to be whether AI could help a radiologist work faster. Now it is which high-volume, negative-heavy workflows can safely absorb enough first-pass reading to change staffing economics. That bigger efficiency question is also showing up in debates over whether AI agents can meaningfully reduce friction across the U.S. healthcare system.

Katz described a world where hospitals buy less radiologist time per unit of screening volume. For a safety-net system, that is a brutally rational goal. The American College of Radiology has warned that radiology faces unsustainable workloads and delays in diagnostic information. RSNA has published that imaging demand is outpacing growth in the number of radiologists entering the field. In that environment, software that removes even a slice of low-yield reading work becomes a budget tool.

Value may start moving away from the radiologist minute and toward the software layer that filters the queue. Whoever controls that layer can shape labor intensity, throughput, turnaround times, service-line economics, and vendor leverage. The winner is the company that gets embedded in the operating rule hospitals use to decide which scans deserve scarce expert attention.

The old augmentation-versus-replacement debate misses the operational shift. Hospitals can put price pressure on the specialty through credible first-pass substitution in screening settings where abnormal, ambiguous, or high-risk cases still escalate to humans. That narrower claim is also the one buyers can actually use. Mammography is the clearest proving ground because the workflow is repetitive, the negative pool is large, and the economics of cutting reader time are easy to measure. Chest X-ray and other high-volume pattern-recognition tasks are close behind. The radiologist remains central to abnormal interpretation, final responsibility, multidisciplinary communication, and the edge cases that break neat workflows.

But once normal studies stop demanding the same level of human attention, the staffing math changes before the org chart does.

What to watch next

First, watch whether regulators open clearer paths for narrow autonomous or semi-autonomous use in defined screening settings.

The FDA still treats these tools as medical devices subject to premarket review and maintains a public list of AI-enabled medical devices authorized for marketing in the United States. The bottleneck is as much legal permission as model quality. Real change starts when software can sit at the front of the workflow without automatic human rereading of every case.

Second, watch procurement language, not pilot announcements.

Hospitals have been piloting radiology AI for years. The real signal will come from contracts and public statements that describe these systems as staffing reducers, coverage extenders, or cost-control tools. Katz already offered a clear preview of that language. The next phase begins when more operators put the same logic into purchasing documents.

Third, watch the guardrails around sign-off and liability.

Mandatory human review on nearly every study keeps the savings capped. Oversight concentrated on flagged, abnormal, or uncertain cases changes the economics fast. That is where this argument either hardens into a delivery model or dies at the conference-demo stage.

Fourth, watch who buys first.

The earliest aggressive adopters will likely be systems with thin margins, screening-heavy volume, and constant staffing pressure. That is the real test.

In the near term, hospitals are starting to treat negative scans as work that no longer needs a human review. Once that mindset takes hold, the initial screening step becomes the most valuable part of the business.

Over a longer horizon, that shift may connect with other computational approaches now being explored in medicine, including how quantum computing could eventually be applied in healthcare.