At Advocate Health facilities across Wisconsin and North Carolina, an AI imaging system now quietly reviews scans looking for the kind of findings that get missed in a busy radiology queue: a small pulmonary embolism tucked in a lower lobe, an incidental brain bleed on a scan ordered for something else entirely. What began as a pilot at 22 sites in October 2024 has, by 2025 and into 2026, become one of the more mature real-world deployments of diagnostic AI in American hospital care.
From Pilot to Standard Workflow
The initial rollout focused narrowly on pulmonary embolisms, incidental pulmonary embolisms, and intracranial hemorrhages—three conditions where a delayed diagnosis can be fatal but where subtle imaging signs are easy for a fatigued radiologist to overlook during a high-volume shift. After the pilot showed the tool reliably flagged cases for faster review, Advocate announced in July 2025 that it would expand AI use across its clinical imaging workflow to cover a wider range of emergencies, including rib fractures, cervical spine fractures, abdominal free air, pneumothorax, aortic dissection, and brain aneurysms. Advocate’s own internal modeling, based on early outcomes from the two-state pilot, projects that nearly 63,000 patients a year will benefit from faster prioritization and earlier diagnosis once the tools are fully scaled—a meaningful figure against the roughly 8 million imaging studies the health system performs annually. Independent research on the specific pulmonary-embolism algorithms Advocate uses, developed by imaging AI vendor Aidoc, has found AI-informed radiologists reaching a sensitivity of 99.2 percent for detecting the condition, with the software’s calls matching radiologist interpretations in 97.8 percent of cases reviewed.
Why This Matters Beyond One Health System
Advocate’s expansion is part of a broader shift industry observers describe as “the pilot era ending.” For years, AI vendors sold hospitals narrowly scoped proof-of-concept tools that ran quietly in the background without changing clinical workflow. Advocate’s approach is different: the AI results feed directly into radiologist worklists, reprioritizing which scans get read first based on urgency the algorithm detects, not just the order they arrived.
A Similar Push for Rare Disease in Children
California’s Valley Children’s Hospital, serving more than 1.3 million children across its catchment area, has taken a parallel but distinct approach, applying AI tools like Epic’s Cosmos platform to help identify rare diseases in pediatric patients faster than manual chart review would allow. Rare pediatric conditions are notoriously hard to diagnose because individual physicians may see only a handful of cases in a career; pooling de-identified data across large patient populations lets pattern-recognition algorithms surface associations a single doctor would likely miss.
The Economics Driving Adoption
Money is a major undercurrent. UnitedHealth has projected AI could save the insurer close to $1 billion in 2026, while HCA Healthcare expects roughly $400 million in AI-driven cost savings, partly through automating revenue-cycle management rather than clinical decision-making. For health systems like Advocate, faster and more accurate imaging triage also reduces costly downstream complications—a missed pulmonary embolism that turns into an ICU admission is far more expensive than the software license that might have caught it earlier.
The Integration Bottleneck
Not every hospital can simply plug in similar tools. A widely cited Carta Healthcare survey found that successful AI pilots often stall not because clinicians distrust the technology, but because integrating AI outputs into existing electronic health record workflows is logistically difficult—EHR integration has now overtaken trust as the leading barrier to adoption. Advocate’s multi-site success partly reflects years of behind-the-scenes IT work to make sure flagged results actually reach the right radiologist’s screen at the right moment, rather than sitting in an unused dashboard. Even with strong detection numbers, researchers involved in evaluating these systems have been careful to note that human oversight remains essential: authors of the pulmonary-embolism accuracy studies underlying Advocate’s rollout explicitly say their findings support continued radiologist review alongside the AI rather than replacing it, since the algorithm’s job is prioritization and flagging, not autonomous diagnosis.
What’s Next
Advocate’s model—starting narrow, proving safety and speed gains, then broadening the clinical scope—appears to be the template other large systems are watching closely. Analysts expect more multi-site health systems to follow with phased expansions covering additional conditions such as sepsis onset and ICU deterioration, particularly as the FDA’s running tally of authorized AI/ML medical devices, which independent trackers put at well over 1,300 entries and climbing through early 2026, continues to grow and give hospitals more validated options to choose from. Aidoc itself expanded its own regulatory footprint in January 2026, securing clearance for what it describes as the first multi-condition AI triage solution for body CT, a platform now also running at AdventHealth, WellSpan Health, Mount Sinai, and Yale New Haven Health alongside Advocate—suggesting the “start narrow, then broaden” playbook is becoming an industry norm rather than one hospital’s isolated strategy. The next test will be whether outcomes data—not just detection speed—shows patients are actually being treated faster and surviving conditions that used to slip through the cracks.
