Large language models push care plans toward higher-cost medicine
Clinical-care plans suggested by large language models carried a price tag about three times higher than the plans implemented by human physicians. In a health system where more than 80% of US physicians already say they use artificial-intelligence models in patient care, the cost gap is hard to dismiss as a technical curiosity. It points to a basic bedside problem: if an AI system nudges clinicians toward broader testing, more intensive treatment, or more expensive downstream services, it can reshape spending before a patient ever leaves the exam room.
The concern is not simply that an AI tool makes a different recommendation. The concern is that the recommendation may be systematically more expensive than the judgment of a practicing physician, without any built-in guarantee of better outcomes. In day-to-day care, cost is not an abstract accounting exercise. It determines which tests are ordered, which referrals are made, how quickly treatment is started, and how many patients can be treated inside a limited budget.
Why a cost gap matters at the point of care
AI systems trained on large language patterns can sound authoritative even when they are not anchored to the same constraints that shape clinical decision-making. Physicians weigh comorbidities, local resources, guideline nuance, prior test results, and the possibility that a low-value order set will generate more noise than benefit. An LLM, by contrast, can generate care plans that appear comprehensive while still leaning toward more expensive pathways. In practical terms, that can translate into more imaging, more laboratory work, more consults, or more aggressive interventions than a clinician would choose after considering the full patient context.
A threefold increase in cost is not a rounding error. It suggests a substantial divergence between machine-generated clinical logic and human decision-making at the moment when plans are formed. If those plans are used as decision support, the economic effect could spread quickly because plan generation is one of the earliest steps in care delivery. If they are used more directly, the cost effect could be even larger.
The finding also raises a quieter issue: more expensive does not automatically mean more appropriate. In medicine, over-ordering can create downstream harm through incidental findings, unnecessary procedures, avoidable anxiety, and treatment cascades that start with a single low-value recommendation. A model that amplifies utilization may look efficient on a screen while producing inefficiency at the bedside.
Physicians remain the economic brake on automated care
Human physicians do more than pick diagnoses and treatments. They serve as a filter against excessive escalation. Experience, clinical judgment, and local practice patterns often prevent a seemingly thorough plan from becoming an indiscriminate one. When an LLM-generated plan costs three times as much as physician care, the gap likely reflects more than stylistic difference. It suggests a mismatch between generative output and the disciplined triage that real-world practice demands.
That matters because artificial intelligence is already embedded in clinical workflows across multiple settings. Once a tool becomes routine, small biases can compound across thousands of encounters. Even a modest tendency to recommend more expensive paths becomes expensive at scale. The financial burden would not stay confined to one department or one health system; it would be distributed across payers, institutions, and patients.
The broader clinical question is how to use AI without importing unnecessary cost into routine care. Generative systems may still have value for drafting notes, organizing differentials, or surfacing possibilities a clinician might not immediately consider. But those strengths do not erase the need for oversight when a tool begins to resemble an autonomous planning engine. Any system that influences ordering behavior should be judged not only on whether it sounds competent, but on whether it preserves the restraint that high-value care requires.
Cost control will have to be built into AI-assisted medicine
The price signal in this analysis should push hospitals and developers toward stricter calibration of AI tools before they reach routine use. Cost awareness cannot be an afterthought. If large language models are to be used in clinical decision support, they will need guardrails that measure whether they are steering care toward unnecessary expense, not just whether they produce fluent recommendations.
That may mean tighter prompt design, stronger physician review, or explicit limits on what a model can recommend without confirmation. It may also mean testing models against real-world endpoints that matter to clinicians and patients alike, including utilization, safety, and the downstream consequences of over-treatment. A tool that generates polished answers but inflates spending can still be clinically disappointing.
For now, the clearest signal is financial. LLM-suggested care plans are costing far more than the plans physicians actually implement, and that gap lands at a moment when artificial intelligence is moving from novelty to routine use in medicine. The unanswered question is not whether AI can generate plans; it clearly can. The sharper question is whether those plans can be made economical, restrained, and safe enough to survive contact with real patients and real budgets.