Hospital AI Deployments Scale as FDA Eases Decision-Support Rules

From pilots to platforms

In 2026, hospital AI stopped being a set of small experiments and became something health systems buy for the whole enterprise. The reporting points the same way across the board. The open question now is less whether hospitals will deploy AI at scale than who checks it once it is running, and federal regulators have moved part of that job onto the hospitals.

The deal flow shows the change clearly. Health systems signed enterprise-scale AI partnerships throughout the year, moving past isolated pilots into systemwide rollouts covering documentation, decision support, virtual care and workforce management.23 Some examples:

• The University of Texas System launched a generative AI deployment across all eight of its health institutions.23
• The Department of Veterans Affairs awarded a contract to expand its ambient AI scribe from a 10-site pilot to more than 130 medical centers.23
• HonorHealth rolled out Abridge's ambient documentation tool to about 3,000 clinicians and skipped the pilot phase entirely.23
• Mount Sinai and Cedars-Sinai built OpenEvidence's clinical decision support platform into Epic, giving clinicians answers drawn from the medical literature inside their normal workflow.23

Survey data supports the anecdotes. Eliciting Insights reported a 67% year-over-year rise in health systems implementing or planning three or more AI solutions. Ambient note-taking had reached 68% adoption.27 Nearly two-thirds of Epic hospitals were reportedly using ambient AI tools as of January 2026.24 Ardent Health said its clinicians passed 1 million patient encounters supported by ambient AI, with the tool used in 87% of ambulatory encounters.25

The planning advice has kept pace. John Snow Labs, which sells healthcare AI tools, gave cost ranges in its guidance for 2026 deployments:21

• $50,000 to $500,000 for basic tools
• $1 million to $10 million for custom platforms
• $100,000 to $1 million a year for infrastructure

It also recommended that hospitals appoint chief AI officers and create AI ethics boards.21 Industry figures surveyed by Chief Healthcare Executive generally expected 2026 to be the year AI moved from pilot programs to full enterprise deployment. Several named ambient listening as the clearest example, as EHR vendors build those features in directly.22

The quiet center of gravity: documentation, not diagnosis

Read together, the coverage shows that the biggest deployments so far are mostly administrative and documentation tools, not diagnostic ones. Becker's list of enterprise deals is dominated by scribes, virtual nursing platforms, care-gap agents and literature-search assistants.23 Kaiser Permanente's AI scribe rollout across 40 hospitals is described as one of the largest generative AI deployments in healthcare.24 Banner Health gave a private chatbot built on Anthropic's Claude to more than 55,000 employees.24

This matters for regulation. Under the FDA's framework, an ambient scribe that drafts a note for a clinician to review has no medical-device purpose. Neither does a scheduling or intake agent.3 So the fastest-growing kind of hospital AI sits largely outside FDA premarket review. Diagnostic AI, the tools that read images and signals, remains inside it.

Radiology still owns the regulated market

On the diagnostic side, the FDA's AI-Enabled Medical Device List keeps growing, and it remains mostly an imaging list:

• Through Q2 2026: the September update counted 1,614 authorized AI-enabled devices, up 5.9% from the first quarter. Radiology accounted for 1,230 of them, or 76%.19
• 2026 so far: the agency authorized 92 AI devices in the first quarter and 89 in the second. Radiology made up about three-quarters of each quarter.19
• End of 2025: the cumulative count was 1,451, with radiology holding the same 76% share.12

The main sources broadly agree on these figures. Some explainers still cite older totals: one 2026 guide says only "more than 1,000" devices and marks the number as needing verification.13 Another tallied 92 devices with 2026 decision dates because it used a version of the list that ended in March.14 The differences come from the FDA's quarterly update schedule. The agency itself warns the list is not comprehensive.11

Large imaging manufacturers lead, partly through acquisitions. GE HealthCare had 134 AI authorizations, followed by Siemens Healthineers with 101 and Philips with 62. Aidoc, with 34, is the top pure-play AI company.19 Several reports describe Aidoc's January 2026 clearance as the first foundation-model-powered clinical AI device: a single CT triage tool covering 14 conditions, with reported mean sensitivity of 97% and specificity of 98% in its pivotal study.18

One detail shows how far regulated AI still is from the generative tools spreading through hospital workflows. As of March 2026, no authorized device on the list used generative AI or large language models.20 The headline-making generative tools and the FDA-reviewed devices are almost entirely separate groups.

The FDA moves the line

The main regulatory change this year came in January. At the Consumer Electronics Show, FDA Commissioner Marty Makary announced a looser approach to clinical decision support (CDS) software and non-invasive wearables.4 STAT described the changes as following through on the Trump administration's promise to deregulate AI. It also warned they could let unregulated generative AI tools into clinical workflows.6

The key change concerns CDS tools that give a single recommendation. The FDA previously treated software offering one recommendation as a regulated device. It now says it will use enforcement discretion, meaning it will generally not enforce device rules, when only one option is clinically appropriate and the clinician can independently review the basis for the advice.26 The agency's own examples show where the line falls:8

• A cardiovascular risk score built from weight, smoking status, blood pressure and lab results is exempt. The same score becomes a device if it predicts risk within 24 hours or uses genomic data.
• Software that summarizes a radiologist's report with diagnostic recommendations is exempt. Software that reads the image itself to write the report is still regulated.

Sources give different dates for the guidance. Trade and legal coverage ties the announcement to January 6.26 Several compliance trackers list January 29, 2026 as the date of the final document.579 The FDA added a CDS FAQ marked current as of June 29.39 The more important point for hospitals is the substance.

Coverage of that substance is more divided. Supporters, including a physician writing in KevinMD, call it a sensible shift that matches how clinicians actually reason and accepts low-risk tools that assist doctors.2 Compliance-focused analysts stress the limits. Software meant for time-critical decisions generally still counts as a device.9 Anything that analyzes pixels stays under FDA oversight, which the American College of Radiology confirmed after the update.10 A Live Compliance analysis puts the trade-off bluntly: more AI decision-support tools can reach clinicians without premarket review, and the burden of vetting them moves to the organizations that deploy them.1

That last reading is the most accurate. The guidance does not deregulate AI imaging. It does widen the range of AI that reaches a clinician's screen without an FDA reviewer looking at it first.

Generative AI and the regulatory gap

The January guidance offered almost nothing specific to generative AI. Critics note it does not address clinician-facing medical chatbots, unconstrained large language models, or patient-facing symptom checkers.10 The FDA began to address that gap on August 18, 2026. It published a discussion paper asking for feedback on possible approaches to generative AI devices, including risk assessment, competency-based premarket evaluation, post-market monitoring, foundation models and agentic AI. Comments are due October 19.5 The paper is exploratory and binds no one.510

For now, then, the generative tools hospitals are buying in bulk are largely governed by vendor contracts, internal governance and a growing set of state laws.

States fill the vacuum

That state-level patchwork is growing fast. Examples include:15

• Texas: HB 149 requires healthcare providers to disclose AI use. SB 1188 requires practitioners to review AI-generated records.
• Rhode Island: providers must notify patients when AI documents a visit and review the output for accuracy afterward.
• Louisiana: clinicians must tell patients aloud before an appointment is recorded for AI transcription.
• California: AB 3030 requires disclosure on certain AI-generated patient communications.
• Utilization review: several states, including California, Washington, Iowa, Maryland and Alabama, bar AI from being the sole basis for denying or delaying care.

Federal policy pulls the other way. A December 2025 executive order directs action to preempt "onerous" state AI laws, while a proposed ONC rule would remove AI "model card" transparency requirements.1 This sets up a real conflict between state lawmakers adding protections and a federal government removing them.

What hospitals actually have to build

The practical result is that hospitals now carry more of the responsibility for AI quality. Some are building systems for it. Texas Children's Hospital set up a formal governance framework that reviews tools for safety, effectiveness and ethics before deployment.24 Mayo Clinic, which runs about 150 AI models, tests tools much like a clinical trial: first with a small group of patients under doctor oversight, then with wider populations.28

Those are well-funded academic centers. The bigger risk is how much readiness varies. Smaller systems are adopting the same ambient and decision-support tools with fewer staff to validate them. Meanwhile, the January guidance itself names automation bias as a concern, meaning clinicians may trust machine output too readily.10

The preparation that matters most for large-scale deployment is governance, not computing power or budget. That means keeping a list of every AI function in use, deciding which ones count as devices, assigning human reviewers, and monitoring performance after launch. In 2026 the FDA has said, through both its guidance and its silence on generative AI, that much of this falls to hospitals. Systems that build these processes now will be in a better position if a generative AI framework becomes binding.