Large language models recommended substantially costlier care than physicians
In a head-to-head comparison across 200 deidentified outpatient and emergency department notes, large language models consistently recommended more expensive care than treating physicians. The gap was not small. Across 20 general-purpose systems, AI-suggested care averaged US$314 per visit, while physician-documented care averaged US$97, a difference of US$216 per visit and roughly 3.2 times the physician cost. Those figures matter because the models were not being asked to manage rare edge cases or elaborate referral pathways; they were responding to ordinary clinical presentations drawn from real-world notes before the physician’s assessment was available.
The pattern was even more striking in the mechanism behind the higher price tag. The excess was driven by two components: US$105 more in diagnostic tests and US$111 more in specialist referrals. New medications were also recommended more often by the models, with 1.14 new drugs per visit versus 0.59 by physicians, but the dominant cost inflation came from testing and referral behavior. In practical terms, the models behaved as if uncertainty should be resolved by ordering more workup, even when the treating clinician had judged that no immediate testing was needed.
Testing at zero-order visits drove the cost gap
A single clinical reality explains much of the divergence: physicians recorded no diagnostic testing in 66% of visits, and on those visits every AI system still suggested testing. The result was a near-universal reflex toward lab panels, imaging, or other diagnostic work in scenarios where clinicians had decided observation, reassurance, or targeted management was enough. In daily practice, that kind of recommendation pattern can turn a simple encounter into a costly cascade, especially when a test triggers follow-up imaging, repeat visits, incidental findings, or specialist referral.
The study design makes that tendency harder to dismiss as an artifact of a single prompt or one model family. Investigators evaluated 24 AI systems using 200 deidentified notes from MTSamples, a public collection of U.S.-format clinical notes contributed by practicing clinicians. Each system received only the presentation recorded before the physician’s assessment and generated its own plan; the treating physician’s documented plan served as the comparator. Orders from both were extracted by three independent LLMs and priced using 2026 Medicare rates. The primary endpoint was the per-visit total cost of recommended care, limited to diagnostic tests and specialist referrals, while new medications were counted rather than priced.
General-purpose and specialized systems both ran expensive
The cost inflation was not confined to broadly deployed general-purpose models. The 20 general-purpose systems averaged US$314 per visit, but the four specialized medical systems were no less costly, running at 2.7 times to 3.5 times physician cost. AI system mean costs also varied widely, from US$209 to US$599, which suggests that model architecture, instruction tuning, or prompt interpretation changes the intensity of recommended care, but not necessarily in a direction that aligns with conservative outpatient decision-making.
This variability is clinically relevant because it shows the problem is not just whether a model can identify plausible diagnoses. The question at the bedside is whether its management plan reflects proportionality. A recommendation can be medically defensible and still be operationally excessive if it adds low-yield testing where clinical pretest probability is low. In outpatient medicine, overtesting has downstream consequences: cost, patient anxiety, false positives, and a larger burden on already stretched specialty services.
Cost-aware prompting lowered the bill but did not normalize it
Two prompt styles were tested: a cost-aware prompt and a parsimonious-but-safe prompt. Both reduced the excess, but neither eliminated it. Under the cost-aware prompt, AI-suggested care still cost 2.5 times as much as physician care. Under the parsimonious-but-safe prompt, the gap narrowed to 1.4 times physician cost. Even the more conservative prompt therefore left a measurable difference, which suggests that cost sensitivity can be introduced into model behavior but not reliably enough to reproduce the restraint shown by human clinicians in this dataset.
The prompt result points to a deeper issue in clinical reasoning by language models. These systems are optimized to generate plausible next steps from text, and plausibility in medicine often leans toward completeness. Yet clinical efficiency depends on recognizing when the safest move is not expansion but restraint. In many ambulatory and emergency settings, the strongest decision is to avoid unnecessary testing because the history, examination, and initial risk assessment do not justify it. A model that repeatedly expands workup in low-risk encounters may appear thorough while still moving against cost-conscious, evidence-based care.
Why the extra workup matters at the bedside
The difference between US$97 and US$314 per visit is more than a billing statistic. It reflects a different philosophy of care. Physicians filtered uncertainty and often stopped at a reasonable threshold; the models tended to continue searching. In a real clinic, that behavior can mean more blood draws, more imaging, more incidental abnormalities, more referrals, and more patient time spent navigating follow-up. It can also shift clinician attention toward problems that are statistically unlikely to matter while burying the more likely diagnosis under a stack of low-yield orders.
The comparison also has an important boundary. These outputs were recommendations, not executed plans. LLMs remain decision-support tools, and final choices belong with clinicians who can integrate exam findings, prior data, patient preferences, and operational constraints. Still, the cost signal is hard to ignore. When models systematically recommend more testing than physicians in visits where no testing was ordered, they are not merely being “complete”; they are defining a higher-cost style of care that has to be judged against the actual yield of each added order.
What this means for clinical deployment
Any effort to deploy LLMs in clinical decision support will need guardrails that go beyond factual accuracy. Cost, pretest probability, and downstream burden have to be part of the output, not an afterthought. A model that names the right differential but repeatedly orders broad panels at low-risk visits could increase spending without improving outcomes. The study’s findings make that tension visible in a way that bedside clinicians will recognize immediately.
The dataset was limited to deidentified outpatient and emergency notes, and the cost model used 2026 Medicare rates for tests and referrals rather than observed real-world billing or patient outcomes. Those boundaries are important, but the directional finding is robust within the study design: AI systems recommended substantially more costly care than treating physicians, mainly because they tested when physicians did not. Any future clinical use will need to answer a blunt question with equal bluntness: can the model recommend care that is not just plausible, but proportionate?