AI was developed to cut healthcare costs by catching issues early. A new study in the NEJM shows these systems often have the opposite effect, ordering more tests than needed.
Researchers from Technion, Stanford, Tel Aviv University, Kaiser, and Rambam Medical Center tested 24 AI systems. They gave each system 200 anonymized patient notes from outpatient and emergency visits and asked for a care plan.
The AI plans cost far more than those created by physicians. The average AI-suggested care cost was $314 per visit. The physician-actual plan cost $97. That’s a difference of $216, making the AI plans 3.2 times more expensive. Individual AI systems ranged from $209 to $599.
What the Price Tag Represents
The $314 figure is based on 2026 Medicare rates. It covers diagnostic tests and specialist referrals. The study did not include the cost of new medications, even though the AI systems suggested more of them (1.14 per visit vs. 0.59 from doctors).
The study used one public dataset. The AI systems only saw the patient’s initial presentation, not the doctor’s full assessment. The authors note the limitations. These are suggestions, not final care plans. The final decision always rests with a clinician.
A plan costing three times more could be wasteful. Or, it might find problems a rushed doctor could miss. The study only attaches prices to the recommendations.
Why AI Orders More Tests
The extra $216 was split evenly: $105 for diagnostic tests and $111 for specialist referrals. In 66% of cases where the doctor did not order a test, every single AI system still recommended one.
Read Also: South Africa’s DDoS Attacks Surge, Targeting Mobile Networks
This pattern looks like defensive medicine. An AI designed to catch every possible issue will cast a wider net. It’s like an over-eager assistant that approves every request. This behavior isn’t unique to general-purpose AI; four specialized medical systems also recommended plans 2.7 to 3.5 times more expensive than physicians.
Medication Costs and Defensive Medicine
The study did not track patient health outcomes or guideline compliance. It also omitted practical spending details, making it impossible to know which extra steps a clinician might have found necessary.
Specialized medical AI systems showed the same habit. A model built for medicine was no more frugal than general-purpose ones. Across a busy clinic, that enthusiasm adds up quickly.
Can We Make AI More Frugal?
The researchers tried to make the AI systems more cost-conscious. A “cost-aware” prompt reduced the AI cost to 2.5 times the physician’s. A “parsimonious-but-safe” prompt brought it down to 1.4 times.
Both prompts helped. Neither got the AI’s cost down to the human level. This shows that cost sensitivity must be built into the tools from the beginning. Simply asking an AI to be cheaper is not enough.
How to design these systems for cost is still an open question. It could involve tagging each test with its price. Or, it could mean capping spending before the model generates a plan. The study didn’t test these ideas, but the results suggest they are promising.
The future of AI in healthcare depends on the companies building these tools. Systems that consider cost, flag unnecessary tests, and defer to doctors are the only ones that will get close to the $97 physician benchmark. Until then, an AI that never misses anything will recommend everything, driving up costs.
