AI Is Meant To Cut Healthcare Costs – So Why Do Its Care Plans Cost 3.2 Times As Much?

AI was meant to be healthcare’s cost-cutter, the assistant that catches problems early and trims the bill. A study published in NEJM suggests that AI has a habit of ordering extra tests instead.

Researchers from Technion, Stanford, Tel Aviv University, Kaiser and Rambam Medical Center gave 24 AI systems 200 anonymised outpatient and emergency room case notes and asked each one to propose a plan of care.

The result was a $216 difference per visit. Across the 20 general-purpose systems, AI-suggested care averaged $314 compared to $97 for the documented physician plan. That’s 3.2 times the price, with individual system averages ranging from $209 to $599, which makes for a hefty extra bill from a technology sold on saving money.

A Price Tag On Suggestions, Not On Outcomes

Priced at 2026 Medicare rates, the $314 figure covers diagnostic tests and specialist referrals.

It leaves out practical spending, patient health outcomes and guideline compliance, so there’s no way to tell which extra steps a clinician would have judged appropriate. The study also counted new medications, with the AI systems suggesting 1.14 per visit against 0.59 from physicians, though their cost was never added to the total.

The study relies on one public dataset of clinical notes, and each system saw only the patient’s presentation before the treating physician’s assessment. The authors are upfront about the limits: the outputs are suggestions as opposed to care plans, and final decisions should stay with clinicians.

A plan that costs three times as much could be wasteful, or it could catch things a rushed appointment would miss. The study simply attaches prices to the recommendations and leaves the verdict to others.

The Over-Eager Intern Problem

The extra $216 split almost evenly, with $105 coming from diagnostic testing and $111 from specialist referrals. The standout finding centred on visits where human physicians hadn’t requested diagnostic tests, which accounted for 66% of cases. In those cases, every AI system still recommended testing.

The pattern resembles classic defensive medicine. An AI model programmed to eliminate diagnostic blind spots will cast a wider net than a clinician standing at the bedside. It operates on the same logic as an overly agreeable AI assistant: approve everything, add extra steps, cover every angle. Much like a new intern requesting every available lab test without considering the financial impact.

The four specialised AI medical systems didn’t escape the habit either. Their recommendations ran at 2.7 to 3.5 times the physician level, so a model built for medicine wasn’t any more frugal than the general-purpose ones. Across a busy clinic, that enthusiasm stacks up fast.

Telling The Machine To Be Thrifty Only Goes So Far

The researchers tried to rein the systems in. A cost-aware prompt, the machine equivalent of asking someone to go easy on the company credit card, cut the excess to 2.5 times physician cost, and a ‘parsimonious-but-safe’ prompt brought it down to 1.4 times. Both helped, but neither got the models to the physician’s price.

The result makes something clear: cost sensitivity must be engineered into clinical AI tools from the start. Prompting the system to consider expenses noticeably brought the total down, but even this optimised run left AI care costs lingering about 40% past the clinician benchmark.

Defining what cost-sensitive design looks like in practice is still a work in progress. It might involve tagging every suggested order with its real-world price, benchmarking recommendations against typical physician habits for comparable visits or allowing clinicians to cap spending before the model runs. The study didn’t test any of those ideas, but the drop to 1.4 times suggests the direction is right.

Teaching AI To Read The Price Tag

The next chapter of this story will be written by the companies building these tools. Systems that weigh cost, flag when a test is unlikely to change the plan and leave the final call with a clinician are the ones with a shot at closing in on the $97 physician figure.

Until then, 3.2 times is what happens when a system built never to miss anything is left to recommend everything. Cheaper healthcare will have to be a design goal in these tools, because the models don’t arrive with it built in.