Artificial intelligence is increasingly being used to answer medical questions, analyse symptoms and suggest treatment options. However, new research has raised an important concern about its financial impact on healthcare.

A study published in the journal NEJM AI has found that medical care recommendations generated by general-purpose AI systems could cost approximately 3.2 times more than the care plans recommended by human physicians.

Researchers estimated that AI-generated medical recommendations would cost an average of $314 per patient visit, compared with just $97 for physicians’ documented recommendations.

The findings raise questions about whether AI medical tools are recommending more tests, specialist consultations and medicines than patients actually need.

AI Medical Recommendations Come With Higher Costs

The research examined how artificial intelligence systems approach medical decision-making compared with human doctors.

Researchers evaluated 20 general-purpose large language models using 200 anonymised outpatient and emergency department medical records.

Each AI system received clinical information from these cases and was asked to prepare a medical assessment and care plan.

The researchers then compared the recommendations with the original plans documented by the treating physicians.

The results revealed a significant difference in estimated healthcare costs.

AI-generated care recommendations averaged $314 per visit, while physicians’ recommendations averaged $97.

This represents an estimated cost difference of approximately $217 per patient visit.

Importantly, these figures represent the estimated cost of recommended diagnostic tests and specialist referrals using US Medicare rates, not actual patient bills or total treatment expenses.

Why Does AI Recommend More Expensive Medical Care?

One major reason behind the higher costs was the tendency of AI systems to recommend additional diagnostic investigations.

These included:

• Complete blood count tests.
• Blood sugar and diabetes screening.
• Thyroid function tests.
• Cholesterol testing.
• Comprehensive metabolic panels.
• Additional diagnostic investigations.

Such tests can be medically useful when supported by a patient’s symptoms and clinical condition.

However, recommending multiple investigations without a clear medical reason can increase healthcare spending without necessarily improving patient outcomes.

Researchers found that diagnostic testing contributed significantly to the difference between AI-generated and physician-recommended care.

AI Systems Frequently Suggested Tests That Doctors Did Not Order

One of the most interesting findings involved patient visits where doctors had not recommended any diagnostic tests.

Physicians documented no diagnostic testing in approximately 66.5% of the examined cases.

However, AI systems frequently suggested additional investigations in these situations.

Across the evaluated general-purpose models, AI-generated recommendations included testing in 45% to 79% of cases where physicians had not ordered tests.

This suggests that AI systems may sometimes favour more extensive medical investigations rather than recognising when no additional testing may be necessary.

The finding does not establish that every additional test was unnecessary. Some AI recommendations could potentially identify medical issues that were not included in the original physician’s plan.

Nevertheless, it highlights the importance of evaluating whether every recommended investigation is clinically justified.

AI Also Suggested More Medicines and Specialist Referrals

The study identified another important difference between artificial intelligence and human medical decision-making.

AI systems recommended an average of 1.14 new medications per visit, compared with 0.59 in physicians’ plans.

They also suggested more specialist referrals.

Additional medication recommendations can create concerns about drug interactions, side effects and unnecessary treatment.

However, medication costs were not included in the main $314 versus $97 comparison. The primary financial analysis covered diagnostic testing and specialist referrals.

This distinction is important when interpreting the research.

Can Better Instructions Make AI Medical Advice Cheaper?

Researchers also investigated whether changing the instructions given to AI systems could reduce unnecessary recommendations.

They tested different prompts designed to encourage AI models to consider healthcare costs and recommend only appropriate investigations.

The results were encouraging.

When instructed to be cost-conscious, the estimated cost ratio decreased from 3.2 times to approximately 2.5 times the physicians’ recommendations.

A more restrictive instruction encouraging economical but safe care reduced the difference further, to approximately 1.4 times.

This suggests that the way AI systems receive medical instructions can influence their recommendations.

However, reducing costs should not become the only objective. Avoiding necessary tests could also put patients at risk.

The real challenge is finding the right balance between medical accuracy, patient safety and affordability.

Does This Mean AI Is Worse Than Human Doctors?

Not necessarily.

The study does not prove that AI systems are incapable of providing useful medical information.

Artificial intelligence can help healthcare professionals organise medical records, explain complex medical terminology, identify possible health concerns and support clinical research.

In certain situations, AI may also suggest investigations that a physician has overlooked.

However, generating a detailed medical plan is different from examining a real patient.

Doctors can assess physical symptoms, understand medical history, consider individual risk factors and discuss treatment preferences directly with patients.

AI systems working only from written clinical notes may not have access to all this information.

Therefore, their recommendations require careful professional evaluation.

Why Human Doctors Must Remain Involved

The findings reinforce the importance of medical supervision when artificial intelligence is used in healthcare.

A responsible healthcare system should not automatically approve every investigation or prescription suggested by an AI model.

Doctors should evaluate whether a recommendation is appropriate for the patient’s condition and whether the proposed investigation is likely to change the diagnosis or treatment.

Human oversight can help prevent unnecessary spending while protecting patients from missed diagnoses and inappropriate medical decisions.

AI should support clinical judgement rather than independently determine what medical care a patient needs.

What Does This Research Mean for the Future of Healthcare?

The healthcare industry is investing heavily in artificial intelligence to improve efficiency, reduce administrative workloads and support medical professionals.

However, the latest research suggests that technological innovation alone does not guarantee affordable healthcare.

Future medical AI systems will need to become more effective at understanding when additional investigations are necessary and when existing information is sufficient.

Developers will also need to evaluate these systems using real-world patient outcomes, medical safety standards and cost-effectiveness measurements.

Healthcare organisations may increasingly assess AI tools not only by their ability to generate medical recommendations but also by the quality and financial value of those recommendations.

Conclusion

The latest NEJM AI research highlights an important challenge for the future of artificial intelligence in healthcare.

Although AI systems can provide detailed medical recommendations, their suggestions may involve substantially higher healthcare costs than the plans documented by human physicians.

The estimated difference of $314 compared with $97 per visit demonstrates why medical AI must be carefully evaluated before being used independently in clinical decision-making.

Artificial intelligence has considerable potential to improve healthcare, but its success will depend on more than technological capabilities.

The ultimate goal should be safer, more accurate and affordable medical care, with qualified healthcare professionals retaining responsibility for sponsorship