Clinical AI tools recommend testing and referrals that cost more than three times what human physicians actually order.
Hospitals adopting clinical decision-support tools expect lower costs and faster workflows. Recent evidence from an evaluation published in NEJM AI shows the opposite operational reality.
The Cost Gap
Researchers evaluated 24 clinical AI systems across routine visits. The models, including general-purpose LLMs and specialized medical systems, generated care plans averaging $314 per visit. Human physicians documented plans averaging just $97 per visit.
Algorithms default to extreme defensive medicine. Even during basic patient encounters where doctors ordered no workup, AI systems repeatedly requested routine diagnostic tests and specialist referrals.
Operational Friction
High diagnostic accuracy on paper does not translate to financial efficiency in practice. Unfiltered AI recommendations risk inflating health system expenditures under fee-for-service models and burdening clinicians with reviewing unnecessary orders.
Current models lack cost-awareness. Until developers train algorithms to account for clinical necessity alongside risk management, health systems must implement guardrails to prevent AI-driven cost expansion.
