NEJM Study Finds AI Care Plans Can Cost More Than Physician Orders
Hospitals have spent the past several years treating clinical AI as a route to lower costs, faster throughput, and less administrative drag. A new study in the New England Journal of Medicine complicates that narrative. Across 200 de-identified outpatient and emergency department notes, AI-generated care plans were not cheaper than the plans documented by physicians; in many cases, they were substantially more expensive.
The analysis evaluated 24 systems, including 20 general-purpose large language models from seven developers and four specialized medical AI tools. Each system received the same clinical presentation given to the treating physician and was asked to generate an assessment and treatment plan. The cost of those plans was then compared using 2026 Medicare rates. The comparison was straightforward, and the result was hard to miss: general-purpose models added an average of $314 per visit, while the specialized medical systems recommended care costing 2.7 to 3.5 times as much as real-world physician plans.
This is not a small accounting detail. In a hospital environment, cost inflation from automated recommendations can ripple into downstream testing, specialist use, and longer care pathways before anyone notices the increase.
Diagnostic Testing Drove Most of the Cost Inflation
The main driver of excess spending was diagnostic testing and referrals. In 66.5% of the 200 visits, physicians ordered no diagnostic workup at all. Even in those no-workup cases, every AI system suggested testing in at least some instances. That pattern matters because diagnostic cascades are often the first place where automated suggestions become expensive fast. A single lab panel or imaging order may seem modest in isolation, but repeated across high-volume ambulatory and emergency encounters, the cost burden accumulates quickly.
There is a plausible clinical explanation for the behavior. Large language models are designed to generate complete, defensible-sounding plans, and in medicine that often means erring toward more testing rather than less. In a real clinic, physicians use context that a text prompt cannot capture: the patient’s appearance, the pace of symptom evolution, the exam, prior visits, local resource constraints, and the way uncertainty narrows during a conversation. A model that sees only a written note may interpret incomplete information as a reason to expand the workup instead of stopping at a clinically sufficient answer.
The study’s authors were careful not to equate excess with waste. Nearly a fifth of AI plans aligned with clinical guidelines, and in some cases the systems identified workups that physicians had omitted. The tension is obvious: one set of recommendations may look expensive, yet another may uncover legitimate missed care. Automated medicine can be overinclusive and occasionally useful at the same time.
High recommendation volume is not the same thing as high-value care.
Why Written Prompts Exaggerate AI’s Appetite for Care
The design of the experiment reveals why these systems may look stronger in theory than in practice. The models received a written clinical presentation, nothing more. They could not ask follow-up questions, review bedside observations, or weigh the subtle clinical cues that shape physician decision-making in real time. A clinician at the point of care is not simply reading a note; they are synthesizing narrative history, physical findings, test results, prior trajectories, and the patient’s own priorities.
That gap between prompt-based reasoning and bedside reasoning may help explain why AI plans skewed toward broader testing. Without the friction of real-time uncertainty, models often default to completeness. In health systems trying to reduce avoidable utilization, completeness can become costly very quickly if it is not filtered through a clinician who knows when restraint is safer than escalation.
Richard Carter, chief technology officer at Proximie, framed the problem in operational terms. He argued that the finding is narrow if it is taken as a verdict on all AI value, because clinical AI and operational AI do not mature on the same timeline. Operational systems can show benefit within weeks or months when they automate scheduling, documentation, or workflow tasks. Clinical AI, by contrast, needs longer-term outcome data before hospitals can claim a dependable return on investment. His broader point is practical: AI should guide rather than decide.
Clinical Oversight Remains the Barrier Between Signal and Spend
The NEJM investigators arrived at a similar conclusion. High-value AI-guided care requires deliberate safety-aware prompting and clinician oversight to filter out low-yield recommendations. That framing is important because it places the burden on implementation, not just model selection. A hospital cannot buy a generic or specialty model and assume the output will be cost-sensitive. The institution has to decide where the tool enters the workflow, who reviews the output, and how aggressively the organization permits the model to recommend new testing.
For physicians, the study reinforces a familiar reality: more suggestions are not automatically better suggestions. A plan that looks comprehensive on a screen may be harder to justify at the bedside if it adds cost without improving outcomes. For administrators, the findings raise a sharper question about AI purchasing strategy. If a model expands utilization, then the system’s true price includes not just licensing fees and integration costs, but also the clinical spending generated by its recommendations.
The financial promise of AI in healthcare will likely depend on whether systems can prove they reduce waste without suppressing needed care. This study suggests that, at least in the setting tested here, AI is not yet doing that on its own. It may still be useful as a second set of eyes, a diagnostic prompt, or a decision-support layer. As a replacement for physician judgment, it may be expensive.
For hospitals trying to translate AI investment into measurable savings, the hardest work may be the most old-fashioned part of medicine: deciding when to test, when to stop, and when human judgment should override a machine’s impulse to keep adding orders.