The Nursing Shortage Is a Design Choice, and AI Is Being Used to Entrench It
40 percent of nurses plan to leave by 2029 and hospitals reach for AI. But the shortage is largely a staffing choice, and AI dropped on a broken model just automates it.
Photo by Juan Moccagatta on Pexels
By 2029, 40 percent of nurses, about 1.6 million people, say they intend to leave the profession, and hospitals are reaching for AI as the answer. But the honest reframe is that the shortage is largely a design choice: a New York analysis found appropriate staffing would save 720 million dollars over two years, yet the typical nurse still carries 6 to 10 patients when 4 is the recommended ceiling. When AI is dropped onto that broken model to cut heads rather than cut load, it just automates the dysfunction and adds risk, and the evidence base is thin, only 3 of 1,357 FDA-cleared AI devices were tested on patient outcomes.
We keep talking about the nursing shortage like it is the weather, something that happens to us and that we simply have to endure. I want to push back on that, hard. A shortage of nurses who want to work under current conditions is not the same as a shortage of nurses, and once you sit with that difference, the rush to hand their jobs to AI starts to look less like innovation and more like cost-cutting wearing a lab coat.
It is not a shortage, it is a choice
Look at the numbers and the story reframes itself. Slate reports that 40 percent of nurses, about 1.6 million people, intend to leave by 2029, and that a New York Medicare analysis found appropriate staffing would save 720 million dollars over two years. Sit with that second number. Better staffing would save money, and yet the typical nurse still carries 6 to 10 patients when the recommended ceiling is 4. That is not a supply problem you can blame on nursing schools. That is a decision to run lean, made over and over, and then relabeled as an unavoidable shortage.
AI as the excuse
Here is where it gets uncomfortable. In July 2026, Montefiore in the Bronx laid off 12 nurses and replaced them with AI software for utilization review. Twelve people, gone, swapped for a tool. And hovering over every one of these decisions is a McKinsey estimate that AI in US healthcare could save 200 to 360 billion dollars a year. When a number that large is on the table, the temptation is obvious: point the AI at the payroll. But utilization review done by an experienced nurse and utilization review done by software are not the same product, even when the output looks similar. One carries judgment the other cannot yet replicate, and you only notice the difference when it is gone.
The evidence is embarrassingly thin
If we are going to trade trained nurses for software, you would hope the software is well proven. It mostly is not. Of 1,357 AI medical devices the FDA has cleared, only 3 were tested on patient outcomes, and just 34 were registered for prospective trials. Three. Meanwhile the failure modes are not hypothetical. A sepsis tool at one hospital kept crying wolf on patients who were not septic. In Nevada in 2025, an AI recommended IV fluids for a dialysis patient at real risk from exactly that. Those are the kinds of mistakes a nurse at the bedside catches and a thinned-out unit does not.
What good product leaders do instead
None of this makes me anti-AI. It makes me anti-shortcut. As others have argued, the AI obsession keeps missing the point on nursing. The move that actually works is to point AI at the load, not the headcount: automate the documentation and the busywork so nurses can be nurses, then measure whether they stay. If a tool cuts cost this quarter and drives nurses out next quarter, it lost money you just cannot see on the spreadsheet yet. Augment, measure retention, demand outcome evidence, and keep a human who can overrule the model. That is the whole playbook.
The honest other side
I owe you the other half of this, because it is real. The exact same tools I am wary of can be genuinely great. Ambient documentation that hands a nurse back an hour a shift is a gift. A monitor that catches a patient sliding at 3am is a second set of eyes nobody should refuse. The tool is not the villain. The intent behind the deployment is what decides whether AI relieves overloaded people or just squeezes them harder. Tap between the two below, because it is the same technology in both.
- By 2029, 40 percent of nurses, roughly 1.6 million people, intend to leave the profession, per reporting in Slate.
- The shortage is largely a design choice: typical loads run 6 to 10 patients per nurse when 4 is the recommended ceiling, and proper staffing was projected to save New York 720 million dollars over two years.
- AI is increasingly the excuse, not the fix: Montefiore laid off 12 nurses in July 2026 and replaced them with utilization-review software, with a McKinsey estimate of 200 to 360 billion dollars in annual savings dangled industry-wide.
- The evidence is thin: of 1,357 FDA-cleared AI devices, only 3 were tested on patient outcomes and 34 registered for prospective trials, while errors like sepsis false alerts and an unsafe fluid recommendation show the risk.
- Good product leaders use AI to augment scarce nurses and give time back, measure retention alongside savings, and demand outcome evidence before trusting a tool at the bedside.
Frequently asked
Is there really a nursing shortage?
There is a real gap, with a projected shortage of 267,330 RNs by 2028 and 40 percent of nurses saying they intend to leave by 2029. But much of it is driven by working conditions and staffing decisions, not an unavoidable lack of people.
Why call the shortage a design choice?
Because staffing levels are a choice hospitals make. Typical loads of 6 to 10 patients per nurse exceed the recommended maximum of 4, and a New York analysis found appropriate staffing would save 720 million dollars over two years, so thin staffing is not purely a supply problem.
How is AI being used in nursing?
In good cases, to reduce documentation and catch deterioration early. In worrying cases, to justify cuts, as when Montefiore replaced 12 nurses with utilization-review software in 2026, framing AI as a savings lever rather than a support tool.
How strong is the evidence for clinical AI?
Weaker than the marketing suggests. Of 1,357 FDA-cleared AI devices, only 3 were tested on patient outcomes and 34 were registered for prospective trials, and documented errors include sepsis false alerts and an AI recommending IV fluids for a patient at risk of kidney failure.
What should healthcare product leaders do?
Deploy AI to augment nurses and return time to the bedside, measure retention alongside cost savings, demand outcome evidence over demos, and keep a human able to overrule the model when it is wrong.
Sources
Naveen Kumar
Healthcare engineering and product executive in Pittsburgh. 15+ years building AI-first patient access, a decade at Treatspace.