Beyond autonomy: why medicine needs artificial intelligence teammates, not artificial intelligence doctors

Beyond autonomy: why medicine needs artificial intelligence teammates, not artificial intelligence doctors T he medical artificial intelligence (AI) field is racing toward fully autonomous systems that promise to diagnose, treat, and manage patients independently.Yet, real-world deployments tell a different story.Although systems such as CE-verified ChestLink 1 (by Oxipit)-which achieves 99.9% sensitivity in chest X-ray screening-and PathFinder 2 -which outperforms average pathologists in melanoma diagnosis-demonstrate impressive performance, the most sophisticated attempts at clinical autonomy reveal fundamental limitations.Despite outperforming primary care physicians in diagnostic dialogue, Google's Articulate Medical Intelligence Explorer (AMIE) 3 cannot practice medicine.Stanford's Optimization Paradox study 4 demonstrates that combining the best individual AI agents paradoxically creates worse clinical systems, with diagnostic accuracy dropping from 77.4% to 67.7%.Evidence from 2023-2025 suggests the real paradigm shift is not replacing clinicians with autonomous AI agents but developing sophisticated AI teammates that augment human expertise.Our perspective postulates that success requires abandoning fantasy of fully autonomous medical AI and embracing collaborative intelligence that preserves clinical judgment while alleviating cognitive burden and physician burnout.In January 2024, Google DeepMind introduced AMIE, 3 a system that surpassed primary care physicians on 30 of 32 diagnostic conversation criteria.The headlines were sensational: "AI Beats Doctors at Diagnosis." However, buried within the paper was an important caveat: AMIE operated in a controlled, text-only setting with unlimited time, no real patients, and no actual clinical responsibility.Despite its impressive results, AMIE cannot prescribe medications, order tests, or take responsibility for its decisions.This technically advanced system is, at its core, just an elaborate chatbot.This gap between algorithmic success and real-world clinical practice highlights a broader challenge in medical AI.Although technology companies promote autonomous agents that will independently diagnose and treat patients, the gap between computational capability and clinical utility remains substantial.The narrative of inevitable progress-from rulebased systems to machine learning, then to generative AI, and ultimately to fully autonomous agents-ignores growing evidence that autonomy itself may not be the appropriate goal.The true paradigm shift in medicine is not just moving from automation to autonomy but evolving from isolated tools to collaborative partners.This difference is not just a matter of words-it is essential for creating AI systems that clinicians will trust, and that will truly benefit patients.Agentic AI marks a real breakthrough in computing power.Unlike traditional AI, which responds to specific inputs with fixed outputs, agentic systems have three main traits: goal-driven independence (pursuing objectives independently), complex reasoning and planning (breaking down ambiguous tasks into clear steps), and flexible adaptation (learning from environmental feedback).These systems use large language models (LLMs) as reasoning engines, coordinating specialized tools to perform complex, multi-step clinical tasks.The promise appears compelling.ChestLink, 1 the world's first CE-certified autonomous medical imaging AI, analyzes chest X-rays without radiologist involvement, achieving 99.9% sensitivity while autonomously reporting 36.4% of normal cases.PathFinder, 2 a multi-agent