A study published in NEJM AI in October, led by Dvir Aran of the Technion with Stanford's Jeremy Goldhaber-Fiebert, asked the question most AI evaluations skip. Not whether the model gets the diagnosis right. What it would cost if you did what it said.
The team used 200 de-identified outpatient and emergency department notes from MTSamples, a public collection of clinical transcriptions. Twenty general-purpose language models and four medicine-specific systems each read the patient's presentation and wrote a plan. Every test and referral was priced at 2026 Medicare rates and compared with what the treating physician had documented. Medications were counted but not priced.
The physicians' plans averaged $97 a visit. The general-purpose systems averaged $314, about 3.2 times as much. Extra diagnostics and extra specialist referrals split the gap roughly evenly, and the models suggested nearly twice as many new medications. The medicine-specific systems were no more restrained.
The gap came mostly from visits where the physician ordered nothing. That was about two-thirds of them. In those visits every system still added testing, in 45 to 79 percent of them depending on the system, often a boilerplate panel.
Anyone who has staffed a clinic knows the visit that needs nothing. The blood pressure is fine on the new dose. The cough is a cold. Often the most valuable thing the physician does is decide not to order, and that decision rarely shows up in the note. It shows up as a blank.
A model reading that note sees a blank too. It fills it.
Aran put it plainly: the next advance in clinical AI shouldn't be systems that recommend more, but systems that know what is enough.
The finding I would put in front of a governance committee is the prompting experiment. Telling the systems to be cost-conscious cut the excess to about 2.5 times physician cost. A prompt asking for the most parsimonious plan that was still safe brought it to about 1.4 times. Neither reached the physician.
So how much a tool orders is not a fixed property of the model. It depends on instructions somebody writes, often the vendor, often without the health system ever reading them. That is an order-set decision. Most organizations have a committee for order sets. Very few have pointed it at the prompt.
One pre-deployment question like this lands in The Sarah Matt Briefing every Tuesday. Get it by email.
Ask for resource-use results, not only accuracy. On a typical low-acuity visit, what does the tool suggest next to what your own physicians ordered?
Ask who writes the system prompt, who can change it, and whether you are told when it changes.
Ask what the tool does when the right answer is nothing. Can it say so?
Then measure it yourself. Pull fifty of your own low-acuity visits, run them through the tool, and count what it adds.
These were recommendations, not orders. No patient received the extra tests, and the study measured cost, not outcomes. Nearly a fifth of AI plans were guideline-concordant and caught testing the physician had left out. Orders were extracted from the plans by separate language models, and the comparison was the physician's documented plan, which can under-record what was actually done. The notes came from one public dataset, and the systems saw only the presentation. A tool inside a live EHR, with history and prior results, could behave differently in either direction.
None of that makes the default less important. A suggestion a clinician has to argue down at the end of a 25-patient day is a suggestion that often gets accepted.
I am a surgery-trained physician-executive currently practicing internal medicine in a charity care setting. For my patients a routine panel is not free. It is a bill, a second trip for the draw, sometimes a result that triggers another test nobody can pay for either.
A tool that orders more by default will look thorough in a pilot. In a clinic like mine it lands on the people with the least room to absorb it.
If your AI tool suggested a test on a visit where your physician ordered nothing, who in your organization would ever see that it happened?
Source: Aran D, Goldhaber-Fiebert JD, et al., "Comparative Cost of Care Recommended by Large Language Models versus Treating Physicians," NEJM AI, October 2026; Stanford Health Policy release, November 4, 2026.
Dr. Sarah Matt is a surgery-trained physician-executive currently practicing internal medicine (charity care). She advises health systems and digital health companies on clinical AI implementation, vendor contracting and operational strategy.
Weekly clinical reality on AI, governance, and what actually works inside health systems.