The Human in the Loop Must Be a Nurse

I spent eight years as a hospice nurse, which included working in an inpatient unit where we cared for patients on ventilators and heavy drips. Hospice found me by accident, and I stayed by choice. The work was horrible and wonderful at the same time, in ways that are hard to explain to anyone who hasn’t done it.

Then came the LinkedIn message that landed me in my first technology job. A company building hospice charting software wanted to discuss the theory that nurses should be involved in designing software for nurses. It was the smartest thing I had heard in years.

I interviewed the following week. I had been on call the night before and had worked 18 hours straight, so I arrived in scrubs, looking rough, and walked into a glass office building where people were playing ping-pong at lunch. I couldn’t believe they were on the clock. I had spent years eating lunch in a bathroom stall, on the days I ate lunch at all.

The gap between those two workplaces is why I care about how artificial intelligence (AI) arrives in healthcare. The premise behind my interview was right, but the industry still hasn’t absorbed it.

Most of the conversations I hear today remain at the level of models, vendors, integrations, and return on investment. The people who use these tools get far less attention.

A nurse, for example, meets a new system at seven in the morning and must learn it while keeping five patients alive. Nurses who had no voice in the design find the gaps within a week and build workarounds around them. The tool goes half-used, the data it produces is unreliable, and the organization pays the price.

The Workforce Math

Nurses are leaving the profession while demand for their work climbs. Researchers at the University of Pennsylvania’s Center for Health Outcomes and Policy Research surveyed 7,887 nurses who left health care employment between 2018 and 2021. Their findings, published in JAMA Network Open, rank the leading contributing factors as planned retirement at 39 percent, burnout at 26 percent, insufficient staffing at 21 percent, and family obligations at 18 percent.

Burnout and insufficient staffing describe the same shift from two angles. There’s always one more patient to admit and one more hole in the schedule, and the nurse standing there absorbs both.

By 2030, one in five Americans will be 65 or older, and the federal Health Resources and Services Administration projects national shortages of registered nurses (RNs) and licensed practical nurses (LPNs), with the deepest gaps in rural areas. No amount of overtime can close a gap of that size, which is why conversation keeps landing on technology.

A 2025 study in Western Journal of Nursing Research, titled “Scheduling Is Everything,” found that scheduling difficulties feed job dissatisfaction and turnover intention, and that flexibility and nurse involvement in building the schedule change the outcome.

AI Accelerates Whatever Practice it Finds

Nurses are typically creatures of habit. They do things a certain way because that’s how they’ve always done it. And it serves them well at the bedside. Repetition makes it easy for them to spot obscure situations that no software can anticipate. As a result, nurses can be slow to accept new technology.

Convincing nurses to adopt a tool is the easier half of the problem. The harder part surfaces when they start using it. Artificial intelligence (AI), for example, takes whatever nurses are already doing and makes it faster. If the way they work has problems, AI speeds those up, too.

Say a nurse checks on a patient but misses something important. The AI takes that incomplete information and writes up a care plan in a few seconds. Because AI only knows what it’s told, the plan lacks critical details that could impact the patient. Unfortunately, the plan looks official and well-organized, so the nurse is more likely to believe it than her rushed notes.

For years, doctors have used voice-recording software for documentation, and it works well. The software types up what the doctor decides is wrong with the patient, so the thinking part of the situation stays human.

Documentation is where AI can honestly help nurses. Before I built software, my job was training new nurses to chart. I sat with hundreds of them while they learned the software, so I know where the minutes go.

A 2026 study in JAMIA Open estimates that nurses spend roughly 31 percent of a 12-hour shift documenting patient information in flowsheets, nearly four hours of screen time. A multistate JAMA study of 8,581 outpatient clinics across five academic health systems found that AI scribes cut electronic health record (EHR) time by about 13 minutes and documentation time by about 16 minutes per eight hours of patient care.

While getting a few minutes back every shift is a gain worth pursuing, technology companies that sell those minutes as a cure for burnout help repeat a mistake the industry has already made once.

Where Skepticism Comes From

Nurses got burned by EHRs. A generation of us heard that charting would take five seconds and that documentation would fade into the background of the shift. Large companies were built on those claims and reaped substantial profits. The charting still isn’t finished at the end of a shift, so a nurse who hears a new promise about software has good reason to discount it.

These days, I sit in rooms with chief financial officers, chief operating officers, and nurse leaders. The pattern I encounter most often is fear of trying something new. I’ve also owned a staffing budget, so I get the position they’re in.

I worked with a nurse leader who had run an excellent department for 20 years and built her schedules by hand, collecting availability, generating a report, and matching open shifts against it. She had no reason to believe a different approach would work, and I had no evidence to give her beyond other people’s results. After months of conversation, she agreed to try self-scheduling in one part of her organization and measure what happened.

Over the pilot, 325 nurses used the new workflow and filled more than 3,500 shifts themselves. Coverage per nurse roughly doubled. She now has her own numbers, which is the only kind that ever moved her.

Having RN after my name lets me say the hard things in those rooms. A department that has brought in outside help already has a problem it hasn’t solved, and defending the current process wastes everyone’s time at the table.

I once heard a physician say: “There’s a reason we call it the practice of medicine.” Physicians don’t have all the answers. They assess their patient, run tests, imaging, and labs, they talk with the person, make a diagnosis, and then try to implement a treatment. If it doesn’t have the desired outcome, they move on to another treatment option or diagnosis.

Hospitals should treat new AI tools the same way. Try one out on a small scale and be honest about whether it worked. Also, have the results judged by people who can tell a working tool from a good sales pitch.

What Good Looks Like

Many aspects of nursing should be open for AI, provided health systems build and govern it correctly. Here are five conditions that I believe are necessary for successful adoption:

1. Practicing nurses must be on the design team: From the start and have authority over requirements while the design is still open to change. A tool is only as good as the clinical input behind it.

2. Every recommendation must show its reasoning before anyone acts on it: A nurse needs to see how the system reached its conclusion and what evidence supports it. Two major nursing bodies agree.

The American Nurses Association’s 2026 consensus report holds that nurses remain the final accountable decision-makers, and names overreliance on AI output as a material risk to professional judgment. The American Academy of Nursing’s statement, published in Nursing Outlook in March 2026, calls for human-in-the-loop oversight across all AI governance.

3. Start where friction is high and clinical risk is low: Documentation and scheduling both qualify. Trust earned in administrative work carries into higher-stakes uses later.

4. Test on one unit against a real baseline: Be willing to report that it didn’t work. A pilot that can’t fail doesn’t really teach anything.

5. Track human outcomes alongside efficiency: Time at the bedside and turnover belong on the dashboard. Nurses leave over staffing shortages and burnout, so a program that lifts throughput while retention slides has measured the wrong thing.

A Rebuild That Connects Them

The first few seconds in a patient’s room told me a lot, from the patient’s skin appearance to the way a family member stood at the bedside. Eight years working in hospice taught me to read those signals together and know when someone had days instead of weeks left. Families asked me that question directly, and the answer never came from the chart alone.

An AI system reading a chart starts from whatever made it into that chart, which leaves out the part I was carrying in my head. For that reason, final authority must rest with the nurse in the room, no matter how good the model gets.

I left the bedside once I understood I could reach more patients by building software than by caring for them one at a time. I never went back, and I still want the same thing. Care in this country can be far better than it is.

AI can return part of the four hours per shift nurses spend charting and give them more control over when and where they work. Neither fixes staffing, though both lighten the load on the nurses carrying it.

Ultimately, tools built with nurses at the table are the tools nurses will use, and patients will feel the difference at the bedside.