AI Is Already in Your Hospital: What It Means for Your Job

Adam Hart was working the emergency room at Dignity Health in Henderson, Nevada when the hospital's computer flagged a newly arrived patient for sepsis. Protocol: administer a large IV fluid bolus immediately. Hart examined the patient and recognized he was on dialysis. Fluid overload could kill him. Hart raised the concern. His supervising nurse told him to follow the protocol. A nearby physician eventually intervened, and the patient received a slow infusion instead.

This isn't a story about AI failing. It's a story about what happens when AI tools arrive without the governance to support the humans using them.

Seventy-five percent of US health systems are now using at least one AI application, up from 59% in 2025. Clinical note-taking alone has reached 68% adoption. This is not a pilot moment — it's a deployment moment. And the workers best positioned are those who understand these tools well enough to evaluate and challenge them, not just use them.

Here's what the evidence actually shows, organized by role. Go directly to your section.

If You're a Physician or APP

The ambient scribe story is mostly good news, and it has real data behind it.

The Permanente Medical Group ran a 63-week study of AI scribes across 7,260 physicians and more than 2.5 million patient encounters. The result: an estimated 15,791 hours of documentation time saved. Eighty-four percent of physicians reported a positive effect on patient communication. Eighty-two percent said work satisfaction improved. Nearly half of patients said their doctor spent less time looking at a computer. This is peer-reviewed analysis of a real deployment — not a vendor whitepaper.

Abridge is live at more than 300 health systems as of June 2026. Nuance DAX and Suki cover similar territory. At Marshfield Clinic, the ambient documentation rollout produced something striking: physicians who had been planning to leave changed their minds after using the tool. Retention is now a legitimate argument for AI scribes, not just productivity.

Vincent Liu, TPMG's chief data officer, puts it plainly: "Both doctors and patients highly value face-to-face contact during a visit, and the AI scribe supports that." The purpose of the tool isn't speed — it's restoring attention to the person in the room.

The honest flip side: benefits were dose-dependent. The top third of users drove 89% of activations and saw more than double the time savings of low-frequency users. Some physicians found editing AI-generated notes took longer than typing from scratch — usually a workflow integration problem, not a fundamental flaw. In-basket time slightly increased even for high users.

It can augment, assist and quantify, but I am not in a place where I give up interpretive conclusions to the technology.
— Theodora Potretzke, Radiologist, Mayo Clinic

Beyond scribes, Mayo Clinic runs more than 250 AI models across its organization, with radiology and cardiology as the heaviest users. AI is entering clinical decision support in ways that are sometimes invisible. Physicians increasingly need to ask: which of my clinical recommendations are already being shaped by an algorithm I didn't choose?

If you're evaluating whether to push for ambient scribes at your organization, Ethan Mollick's Co-Intelligence offers the clearest framework for thinking about human-AI collaboration without losing what makes you clinically irreplaceable.

If You're a Nurse

The nurse story is more complicated.

Hart's sepsis case isn't an isolated incident — it illustrates a structural problem. When an AI recommendation becomes a protocol requirement, the model's error tolerance becomes the patient's risk. The alert may have been right more often than not. The problem was that following it automatically would have harmed this specific patient.

Melissa Beebe, a cancer nurse at UC Davis Medical Center, names the cumulative version of this problem: "You're trying to focus on your work but then you're getting all these distracting alerts that may or may not mean something. It's hard to even tell when it's accurate and when it's not because there are so many false alarms." Alert fatigue is a patient safety issue, not a convenience complaint. A nurse who has learned to dismiss noise is a nurse who may miss the one real signal.

This isn't an argument against alert systems. It's an argument for precision. An alert system that flags a bowel movement as an emergency has failed at design, not just implementation.

AI can be a powerful tool that augments our ability to deliver safe and effective care. These resources can enhance efficiency and support decision-making — but they do not replace professional accountability.
— Justin Gill, DNP, APRN, RN, FNP-C, President, Washington State Nurses Association

Justin Gill, a nurse practitioner and president of the Washington State Nurses Association, represents the honest middle position. He uses AI-assisted tools in his own clinical work for documentation and rapid access to evidence-based recommendations. His verdict: "AI can be a powerful tool that augments our ability to deliver safe and effective care. These resources can enhance efficiency and support decision-making — but they do not replace professional accountability. The responsibility for patient care, and the use of any AI resource, ultimately rests with us as nurses."

On the patient outreach side, Hippocratic AI raised $126 million at a $3.5 billion valuation in late 2025 to expand AI agents that handle pre-appointment preparation and patient navigation — 24/7, multilingual, at scale. These agents are doing work that previously fell to medical assistants, care coordinators, and nurses during off-hours. That's a real workforce impact, even if no bedside nurse loses a shift because of it. Roschelle Fritz, a nursing researcher at UC Davis, offers the necessary counterpoint: "It's the very sick who are taking up the bulk of health care in the U.S. and whether or not chatbots are positioned for those folks is something we really have to consider."

Three questions worth asking your employer right now: What is the false-positive rate on the alert systems in your unit? Is there a documented override protocol you can use without disciplinary risk? Were nurses involved in deployment decisions before go-live?

If You Work in Radiology, Coding, or Revenue Cycle

Geoffrey Hinton said in 2016 to stop training radiologists. He was wrong about replacement and right about direction.

At Mayo, radiologist Theodora Potretzke saves 15 to 30 minutes per kidney image using an AI volume measurement tool she helped design and test. Her assessment is the clearest formulation of the honest position: "It can augment, assist and quantify, but I am not in a place where I give up interpretive conclusions to the technology." Her colleague Francis Baffour describes the broader reality: "A.I. is everywhere in our workflow now" — image cleanup, prioritization, clot detection — but radiologists still own interpretation.

Radiology is the most AI-saturated clinical specialty, and radiologists are more productive, not replaced. The tasks being delegated — measurement, prioritization, routine screening flags — were always a means to interpretation, not the interpretation itself. MIT labor economist David Autor puts it well: predictions that AI will steal jobs "underestimate the complexity of the work that people actually do — just as radiologists do a lot more than reading scans."

The practical implication: radiologists who participate in AI design and validation (as Potretzke did) are better positioned than those who wait for the finished product. The value is in clinical interpretation, communication, and oversight — not in measurement operations the AI already does better.

For medical coders, the picture is more direct. Thirty-six percent of health systems are using AI coding solutions, with 29% year-over-year growth. The direction is clear. Routine high-volume code assignment is moving toward automation. The human role concentrates on exceptions, compliance review, and payer-specific edge cases that require judgment. A coder who knows only routine entry is more exposed than one who understands clinical documentation quality and denial patterns.

Revenue cycle is moving fast in a specific direction: denial prediction sits at 25% adoption, prepopulated technical appeals at 21% with 50% year-over-year growth. Appeal prepopulation growing that quickly means organizations are automating payer interactions — and the professionals who understand the reasoning layer above that automation are becoming AI quality reviewers and exception managers, not being displaced. The data-entry layer is the exposed segment. The compliance and judgment layer is where value concentrates.

What to Actually Do in the Next 12 Months

The single most transferable skill across every healthcare role right now isn't prompt engineering. It's knowing how to evaluate an AI output: what data it used, what it didn't, where it's likely to fail, and when to override it.

Adam Hart's clinical judgment about the dialysis patient is the highest-stakes version of this skill. A coder reviewing an AI-suggested code is a lower-stakes version. Both require the same underlying capability: understanding the tool's purpose and limits well enough to know when the output is wrong.

The practical version for clinicians who aren't building AI systems: learn what questions to ask. What data trained this model? What's the false-positive rate in my patient population? Has this been validated in my specialty and setting? These aren't technical questions — they're professional accountability questions any clinician can learn to ask.

DataCamp's AI Fundamentals or Generative AI for Business track gives any healthcare worker the vocabulary to ask those questions without requiring a data-science career change. The free tier exists. This is professional self-defense, not reinvention.

For those ready to move toward emerging roles: nearly half of EU member states have already created dedicated AI and data-science positions in health. In the US, healthcare AI hiring is concentrated in ambient clinical AI implementation, AI governance, and clinical informatics. Clinical informatics specialists earn around $118,000; informatics directors around $150,000. The Career Essentials in Generative AI by Microsoft and LinkedIn is free, Microsoft-backed, and appears on your LinkedIn profile — a low-cost signal for employers building out these teams.

Three signals worth watching over the next year: draft replies to patient texts had 80% year-over-year growth — the fastest-accelerating category in the 2026 health system survey — which means patient communication roles should expect AI to arrive soon. Revenue-cycle automation is accelerating specifically in appeals. And PwC found that only 27% of organizations report broad impact from digital investments, with integration and adoption as the main barriers — meaning implementation capacity, not model quality, is the actual constraint.

The healthcare workers at most risk aren't those in the highest-skill roles. They're those who either defer entirely to AI recommendations or ignore the technology altogether — for different reasons, both end up performing only the tasks AI already handles adequately.

If you're a physician or APP, push for ambient scribe access and evaluate the output rather than accepting every draft. If you're a nurse, the skill that protects you is clinical override judgment paired with the organizational confidence to use it. If you're in radiology, participate in AI design and testing rather than waiting for a finished product. If you're in coding or revenue cycle, move toward exception management and complex-case expertise — that's where value will concentrate.

For everyone: the workers who understand these tools well enough to question them will shape how they're deployed. That's not a prediction. It's already happening.

Recommended Tools & Resources