Healthcare AI Leaves Pilots as FDA Device Tally Reaches 1,614

The pilot era gives way to production

For about two years, AI in healthcare mostly meant pilots: small trials in one department, built around a single tool and rarely tied to a budget line. By late 2026, executives, consultants and researchers describe a different picture. AI is being written into the daily workflows of hospitals, and the standard for success has moved from whether a model works to whether it can be governed, audited and shown to pay off. Leading into the year, executives were already saying the pilot era was ending and that the focus was moving to systems that can be trusted at scale23. Health systems told MobiHealthNews that 2026 would test whether AI could deliver measurable outcomes inside clinical and operational workflows, under clearer guardrails24.

The evidence now suggests the move to production is real. It is also lopsided. Most of the scaling has happened in documentation and administration, while regulated diagnostic AI keeps growing in a narrow band dominated by radiology. Meanwhile, federal regulators have pushed more of the job of vetting AI onto the hospitals that buy it. Together, these three trends say more about where clinical AI is going than any single product launch.

Ambient scribes are the first true enterprise deployment

The clearest case of AI moving from pilot to operations has nothing to do with diagnosis. It is the ambient scribe, which listens to a patient visit and drafts the clinical note. A study in the American Journal of Managed Care looked at 6,561 US hospitals and found that 62.6% of hospitals running Epic had adopted ambient AI. Adoption was markedly higher at nonprofit hospitals (70.2%) than at for-profit ones (28.8%), and higher at metropolitan hospitals than rural ones1. Physician surveys point the same way. Doximity found that use of voice documentation tools rose from 20% to 29% of surveyed physicians in less than a year3. The American Medical Association's 2026 survey found that 81% of physicians now use AI at work, up from 38% in 20237.

Deployments at single health systems show the scale. Cleveland Clinic says more than 4,800 clinicians used an ambient scribe across 3.5 million patient encounters in the first year of an enterprise rollout. Roughly 80% of its ambulatory workforce was onboarded in under 16 weeks8. Kaiser Permanente's rollout of Abridge across dozens of hospitals and hundreds of clinics is still cited as one of the largest generative AI deployments in healthcare21. In Spain, Quirónsalud's scribe went from 2.7% of outpatient visits to about 31%, with more than 2.3 million assisted encounters10.

The numbers on benefits are where the coverage splits most sharply. Vendor-adjacent material claims daily savings of 40 to 60 minutes per primary care physician2. One development firm says integrated AI systems return more than three times their cost, with payback in three to six months28. Peer-reviewed research is much more cautious. A JAMA study of 8,581 clinicians at five academic health systems linked scribe adoption to 16 fewer minutes of documentation per eight scheduled patient hours, and after-hours EHR time did not change significantly3. Black Book Research found that only 8% of adopters reached positive ROI in the first year9. The reasonable conclusion is that scribes are popular and do help clinicians, but the efficiency gains are modest and uneven. Their staying power may depend more on clinician retention and well-being than on hard cost savings.

Radiology dominates regulated diagnostic AI

Diagnostic AI follows a different path, one shaped by FDA authorization. The agency's AI-Enabled Medical Device List, updated September 4, 2026, contains 1,614 devices with decisions through June 29, 202618. The Imaging Wire reported that the FDA authorized 92 AI-enabled devices in the first quarter of 2026 and 89 in the second20. Reported totals vary across outlets, from about 1,357 to 1,451 to 1,524, but the gap mostly reflects which snapshot of a regularly updated list each outlet used261912. The FDA says the list is not comprehensive and runs behind actual authorizations18.

Radiology's dominance is the most consistent finding. Radiology is the lead panel for 1,230 of the 1,614 devices, or 76.2%. Cardiovascular is a distant second at 9.5% and neurology third at 4.5%18. Almost all of these devices, 96.2%, reached the market through the 510(k) pathway, which relies on similarity to existing products. Only 40 came through De Novo and 21 through premarket approval18. Large imaging vendors hold many of these authorizations because AI is now built into scanners themselves, not only sold as separate software. GE HealthCare reportedly leads radiology AI with 120 cumulative authorizations12.

The shift to operations is visible here too. Hospitals are no longer adding one triage algorithm at a time. They are moving to platforms. Sutter Health is scaling an "AI operating system" approach with Aidoc that connects imaging and EHR data, so critical findings set off care coordination right away25. A May 2026 partnership between Aidoc and Sol Radiology in Southern California links imaging AI with tools that start treatment pathways26. In January 2026, Aidoc received what was described as the first FDA clearance for a clinical device powered by a foundation model: a body CT triage tool covering 14 conditions19. One analysis describes the new workflow this way: AI flags urgent scans, takes measurements before the radiologist opens the study, and passes structured reports into referral and billing21. Executives quoted by HIT Consultant summed up the change as moving from asking whether AI can detect a finding to asking whether AI can make sure the finding is acted on23.

Regulators step back and hospitals take on the risk

On regulation, the coverage agrees more than it might seem, even if the conclusions drawn differ. On January 6, 2026, the FDA issued revised Clinical Decision Support guidance, which Commissioner Marty Makary framed as cutting unnecessary regulation33. It was re-issued on January 2939. The main change is that the FDA will now use enforcement discretion for tools that give a single clinically appropriate recommendation, as long as clinicians can independently review the basis for it32. The guidance also says software that summarizes a radiologist's findings and proposes a diagnostic recommendation may qualify for enforcement discretion if a provider stays in the loop33.

Law firms differ on what this means for generative AI. Some describe the guidance as easing the path to market for AI copilots that summarize findings or suggest diagnoses37. Covington points out that the guidance never actually addresses AI, despite being announced as AI-focused, which leaves developers uncertain about chatbots and generative tools33. The FDA did say at a March 2026 town hall that an LLM-enabled function could meet the independent-review criterion if clinicians can check the basis for its recommendations39. Software for time-critical decisions, such as stroke or sepsis alarms, generally still counts as a device38. The agency's August 18, 2026 discussion paper on generative AI devices, with comments due October 19, is exploratory and not guidance36.

The practical effect is easy to see. More decision-support tools will reach clinicians without FDA premarket review, which moves the vetting burden onto the organizations deploying them31. States are filling some of the gap. Texas requires disclosure when AI is used for diagnosis and review of AI-generated records. Rhode Island requires patient notice and accuracy review of AI visit documentation. Louisiana requires verbal disclosure before a visit is transcribed by AI36. Several states, including Washington, Iowa and Alabama, now bar AI from being the sole basis for coverage denials31.

What it adds up to

The reading supported here is that healthcare AI has truly entered operations, but on administrative ground first. Diagnostic AI keeps expanding through a regulatory channel that is fast, permissive and heavily concentrated in imaging. Deloitte found that more than 80% of health systems are prioritizing agentic AI for clinical operations and revenue cycle work. It also notes that larger organizations, with stronger data foundations, are better placed to move beyond pilots29. Combined with the adoption gaps between nonprofit and for-profit hospitals and between urban and rural ones1, this suggests the production era may widen existing gaps in capability between health systems.

The bigger risk is governance. Speakers at HIMSS26 named legacy systems, unclear governance models and gaps in workforce readiness as continuing barriers30. With federal oversight loosening and an evidence base that shows real but limited benefits, hospitals now hold most of the responsibility for checking that the tools they deploy are accurate, fair and worth the cost. The coming year will show which systems built that capacity alongside their AI rollouts.