AI's recommendations cost 3x more than physician care plans

Healthcare recommended by AI systems costs significantly more than the care documented by human physicians, a study published in NEJM AI shows. This is primarily because AI suggests tests even when physicians don't order them.  

Fears that AI will send the cost of care skyrocketing have been mounting across the industry for years. New research adds to these concerns, showing that AI-recommended care could indeed raise costs.

As physicians increasingly use AI to support clinical decision-making, understanding the impact of these tools on cost is essential, especially as the healthcare affordability crisis deepens. Already, payers and providers are locked in an AI arms race, both using the tools to reduce their administrative burden, which inadvertently raises medical spending.

The new study shows that clinical AI could be another culprit behind ballooning healthcare spending.

For the study, researchers tested 20 general-purpose large language models using 200 de-identified outpatient and emergency department notes from MTSamples, a publicly available collection of clinical notes from practicing clinicians.

Each LLM received the notes and suggested a care plan, which researchers compared to the physician's care plan based on the same notes. Researchers used independent LLMs to gather orders from both sets of care plans and price them according to 2026 Medicare rates.

Overall, the LLMs' care plans included more diagnostic tests, referrals and medications than the physicians' plans. As a result, the total cost of recommended care per visit was $314 across the 20 general-purpose AI systems, which was 3.2 times the physicians' recommended care cost of $97 per visit. The average cost of care per AI system ranged from $209 to 599.

Diagnostic testing was one of the major cost drivers. Physicians' plans suggested diagnostic testing that amounted to $71 per visit, compared to $176 per visit across the AI systems. Notably, the LLM-suggested testing was concentrated among visits where physicians did not order any tests. The physician did not order diagnostic testing in 66.5% of cases, but the AI systems added tests in 45% to 79% of these cases.

The most common tests suggested by AI were a reflexive screening panel, a complete blood count, a comprehensive metabolic panel, hemoglobin A1c testing, thyroid-stimulating hormone testing and a lipid panel, the study shows.

AI also recommended prescriptions and specialist referrals two to five times more than physicians.

In addition to charting the rise in care costs, the researchers evaluated whether different prompting could rein in AI's excess recommendations. Prompts that asked AI systems to be cost-conscious or extremely unwilling to spend money reduced the cost of recommended care but did not fully curb it.

Researchers concluded that AI-recommended care "parallels human defensive medicine," which refers to physicians ordering tests and procedures primarily to protect themselves from liability rather than for the patient's benefit.

Still, the AI's recommendations were not necessarily wasteful. The study authors noted that nearly a fifth of AI plans were aligned with clinical guidelines and caught testing and orders the physician had omitted.

"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," the researchers wrote.

Anuja Vaidya has covered the healthcare industry since 2012. She currently covers healthcare IT and innovation, including artificial intelligence, digital healthcare, EHRs and interoperability.