AI Care Plans Cost 3.2x More Than Doctors' Plans: What the NEJM AI Study Found

A study published in NEJM AI (DOI 10.1056/AIoa2600537) found that care plans written by 20 general-purpose large language models averaged $314 per visit versus $97 for treating physicians, about 3.2 times higher, when 200 de-identified outpatient and emergency department notes from MTSamples were priced at 2026 Medicare rates. Diagnostic testing drove most of the gap, at $176 per visit for AI plans against $71 for physicians, and physicians ordered no diagnostic testing at all in 66.5% of cases while the AI systems added tests in 45% to 79% of those same cases. Cost-aware prompts narrowed the difference without eliminating it, and the researchers note the figures reflect recommended care rather than care delivered, with no data on patient outcomes.

A new study in NEJM AI https://ai.nejm.org/doi/full/10.1056/AIoa2600537?ref=explainx puts a dollar figure on something clinicians have worried about for years: when a language model writes the care plan, it orders a lot more. Researchers fed 200 real clinical notes to 20 general-purpose large language models, priced every recommended test, referral and prescription at 2026 Medicare rates, and compared the totals with what the treating physicians had actually documented. The AI plans averaged $314 per visit against $97 for physicians , roughly 3.2 times higher, according to TechTarget's report on the paper https://www.techtarget.com/healthtechanalytics/news/366652874/AIs-recommendations-cost-3x-more-than-physician-care-plans?ref=explainx . That number needs careful reading. It is a cost of recommendations , not a bill that was paid, and it says nothing by itself about whether patients would have been better off. But it is a useful counterweight to the steady stream of "AI beats doctors" headlines, such as Microsoft's Diagnostic Orchestrator results https://explainx.ai/blog/microsoft-diagnostic-orchestrator-science-paper-83-percent-vs-doctors-sdbench-claimed-vs-verified-2026 , which measure accuracy and mostly ignore what the extra workup costs. TL;DR: the numbers and the caveats | Question | Answer | |---|---| | Where was it published? | NEJM AI , paper titled "Comparative Cost of Care Recommended by Large Language Models versus Treating Physicians" DOI 10.1056/AIoa2600537 , DOI 10.1056/AIoa2600537; TechTarget covered it October 27, 2026 | | What was tested? | 20 general-purpose LLMs on 200 de-identified outpatient and emergency department notes from MTSamples | | Headline result | $314 per visit for AI plans vs $97 for physicians, about 3.2x; per-system average ranged from $209 to $599 | | Biggest driver | Diagnostic testing: $176 per visit for AI vs $71 for physicians | | Did prompting help? | Cost-aware prompts reduced the gap but did not remove it | | What it does not show | Patient outcomes, whether extra tests were useful, or what happens when a clinician reviews the AI plan | What exactly did the researchers do? The design is simple, which is part of why it is interesting. Each model received a de-identified note from MTSamples https://www.mtsamples.com/?ref=explainx , a public collection of clinical notes written by practicing clinicians, and was asked to suggest a care plan. The researchers then compared that plan with the plan documented in the same note. To avoid hand-counting orders, they used separate LLMs to pull the orders out of both sets of plans, then priced everything using 2026 Medicare rates. Pricing at one fixed schedule matters: it means the gap reflects how much care was ordered , not different prices at different hospitals. Two assumptions deserve a flag, because both push the AI number up. First, every recommendation is treated as if it would be carried out. Second, the models see a written note, not the patient. A physician who has examined someone often knows which test is pointless; a model reading the note has only the text. Where the extra $217 per visit comes from Medical bags beside a sorting funnel, picturing how extra tests pile up in an AI care plan cost According to TechTarget, diagnostic testing was the biggest single driver. Physicians' plans averaged $71 per visit in testing; across the AI systems it was $176. The pattern is more telling than the total: physicians ordered no diagnostic testing at all in 66.5% of cases, yet the AI systems added tests in 45% to 79% of those same cases. The most common additions were routine, low-cost-looking panels: - a reflexive screening panel - a complete blood count - a comprehensive metabolic panel - hemoglobin A1c testing - thyroid-stimulating hormone testing - a lipid panel None of those is exotic. Each is individually cheap and individually defensible. The cost comes from ordering them reflexively on visits where a physician judged they were not needed. The study also reports that AI recommended prescriptions and specialist referrals two to five times more often than physicians did. "Parallels human defensive medicine" The authors' own framing is that the AI behavior "parallels human defensive medicine," the habit of ordering tests mainly to guard against liability rather than for the patient's benefit. That is a plausible explanation and an uncomfortable one. A model trained on medical text has seen far more descriptions of what could be checked than of the judgment to leave it alone, and a chat model tuned to be helpful and thorough tends to add rather than subtract. It also fits what other clinical-AI work has shown about models and calibration. In Google's AMIE real-clinic study https://explainx.ai/blog/google-amie-lancet-bidmc-real-clinic-study-90-percent-differential-2026 , the strength was a broad differential diagnosis; breadth is exactly what drives up a test list. A system rewarded for not missing anything will look excellent on sensitivity and expensive on cost. Can you prompt your way out of it? The researchers tried. Prompts that told the model to be cost-conscious, or extremely unwilling to spend money, reduced the cost of recommended care. They did not fully curb it, according to TechTarget. I could not open the full paper the NEJM AI page blocked automated access , so check the article for the exact ratios under each prompt. The practical lesson for anyone building clinical tools is that cost behavior is not something a system prompt fully controls. Teams building agents on top of frontier models already know this from other domains: instructions shift tendencies but rarely remove them. If cost matters, it needs its own evaluation, not just a line in the prompt. Is the extra care wasteful? Not all of it A stethoscope curving around a heart, standing for the clinician who stays the final check on AI orders This is the part headline writers skip. The authors say nearly a fifth of the AI plans were aligned with clinical guidelines and caught tests or orders the physician had omitted. So there are two readings, and the data supports a mix of both: 1. Over-ordering: most of the additional workup was not clearly justified, and a note-reading model piles on routine panels. 2. Safety net: some of the additions are real catches, the kind a second set of eyes is supposed to provide. The authors' conclusion, as quoted by TechTarget: "These findings do not argue against clinical AI; they argue that its value depends on how it is deployed… Our data show that currently the clinician who has seen the patient must ultimately decide what is ordered; deploying these AI systems without that learned intermediary could carry the full excess we measured." Why this matters beyond one paper Healthcare affordability is already a political problem, and AI is entering the system from both sides. TechTarget notes that payers and providers are in an "AI arms race," both using the tools to cut administrative burden, which can inadvertently raise spending. If provider-side AI nudges orders up and payer-side AI pushes denials up, the friction compounds. There is also a deployment angle. Tools that draft orders inside the record, such as the ChatGPT and Epic EHR integration https://explainx.ai/blog/chatgpt-healthcare-epic-ehr-integration-2026 we covered in September, sit exactly where this effect would show up. The study does not test those products, and it uses general-purpose models rather than purpose-built clinical systems. But it is a reason for hospitals to measure order volume after rollout, not just clinician satisfaction. It also bears on the push toward more autonomous care, such as the Utah AI prescription pilot https://explainx.ai/blog/nolla-health-utah-ai-initial-prescriptions-acne-first-ai-doctor-2026 we covered earlier. The NEJM AI authors argue that, for broad and open-ended care, the clinician who saw the patient should stay in the loop. Limits of the study - Notes, not patients. The model never examined anyone. Real deployments may feed it richer data, or less. - Cost, not outcomes. A cheaper plan is not a better plan. A more expensive one may have prevented a missed diagnosis. - Physician plans as the benchmark. The documented plan is itself imperfect and reflects local practice; the study does not claim it is the right amount of care. - Public sample notes. MTSamples is a convenient public dataset but not a random draw of modern clinic visits. - Models age fast. Newer releases than those tested may behave differently, though the prompt results suggest the tendency is stubborn. - Secondhand details. I relied on TechTarget's detailed report; I could not open the paper itself. What this means for people building or buying clinical AI If you ship or procure these tools, a few concrete steps follow from the study: | Step | Why | |---|---| | Track orders per visit before and after rollout | The cost effect shows up in order volume, not in satisfaction surveys | | Evaluate cost alongside accuracy | Accuracy-only benchmarks reward thoroughness | | Keep the treating clinician as the approver | The authors' central recommendation | | Test prompts as a mitigation, not a fix | Cost-aware prompting helped but left a large gap | | Review the "extra" orders for guideline value | About a fifth were useful catches, so blanket suppression loses signal | Teams working on agent safety in other domains will recognize the shape of this problem: the system does what it was optimized for, and the unmeasured cost accumulates elsewhere. For teams deploying agents that take real actions, AgentBeam https://agentbeam.com/?ref=explainx is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions, which is the same principle of a checkpoint between a model's suggestion and a consequential order. Bottom line The headline, 3.2 times the cost, is real in the narrow sense that it comes from a controlled pricing exercise on 200 notes and 20 models. It is not proof that AI makes healthcare more expensive in practice, and it is not proof the extra tests were wasted. What it does show is that thoroughness is not free, that prompts only partly tame it, and that the human who saw the patient is still the cheapest filter available. For related reading on how we cover clinical AI claims, see the CNN vs frontier models skin-cancer comparison https://explainx.ai/blog/cnn-beats-gpt-6-astra-claude-fable-5-1-skin-cancer-diagnosis-lynch-2026 and Meta's Muse Health research results https://explainx.ai/blog/meta-muse-health-research-ten-manuscripts-four-weeks-five-million-records-2026 . The study is also being discussed on Hacker News https://news.ycombinator.com/item?id=50287724&ref=explainx , where the thread was small at the time of writing. Related reading Figures are as reported by TechTarget on the NEJM AI paper as of October 29, 2026; check the paper for final numbers.