AI in health care comes at a price: its care plans cost 3.2 times more than doctors’ plans

Artificial intelligence promised to make health care cheaper, but a study published in the journal NEJM suggests that, for now, it does the opposite: it orders more tests and more referrals than a doctor. Researchers from Technion, Stanford, Tel Aviv University, Kaiser and Rambam Medical Center tested 24 artificial intelligence systems on 200 anonymized outpatient and emergency room case notes, asking each one to propose a care plan for every case.

The result, priced at 2026 Medicare rates, was a difference of $216 per visit: plans suggested by the 20 general-purpose systems averaged $314, compared with $97 for the treating physician’s documented plan. That is 3.2 times more, with per-system averages ranging from $209 to $599. The four medical-specialized systems did not escape the pattern either: their recommendations cost 2.7 to 3.5 times the physician level.

The extra cost split almost evenly: $105 in diagnostic testing and $111 in specialist referrals. The most striking finding came in visits where the physician had ordered no diagnostic tests, 66% of cases: in all of them, every AI system still recommended testing. The systems also suggested more new medications, 1.14 per visit versus 0.59 for physicians, although that cost was not added to the total.

The researchers tried to rein in spending with instructions: asking the system to be cost-aware cut the excess to 2.5 times the physician’s cost, and a “parsimonious-but-safe” instruction brought it down to 1.4 times. Neither version reached the physician’s price, suggesting, according to the report, that cost sensitivity must be engineered into these tools from the start rather than left to the prompt given to the model.

The study has limits its own authors acknowledge: it relies on a single public set of clinical notes, each system saw only the patient’s presentation before the treating physician’s assessment, and it prices recommendations rather than measuring health outcomes or guideline compliance. Its outputs are suggestions, not definitive care plans, and the final decision must remain with the clinician.

Photo: TechRound

Source: TechRound