Autonomous AI Agents in Healthcare

The use of large language models (LLMs) and other forms of generative AI (GenAI) in healthcare has surged in recent years, and many of these technologies are already applied in clinical settings. As such, they often qualify as medical devices and must comply with specific laws and regulatory frameworks. Recent decisions have shown that approval is possible for applications with narrowly defined tasks. The next generation of AI systems are broad autonomous AI agents designed to independently carry out complex, goal-directed workflows. They show great potential to support and further improve medical care in the future. AI agents consist of multiple interconnected components, including external databases and computational tools for image analysis, note-taking, clinical guidance, and patient data management. These components are controlled by LLMs that take over decision-making, error handling, and the recognition of completed tasks.

"We are seeing a fundamental shift in how AI tools can be implemented in medicine," says Jakob N. Kather, Professor of Clinical Artificial Intelligence at the EKFZ for Digital Health at TUD and oncologist at the Dresden University Hospital Dresden. "Unlike earlier systems, AI agents are capable of managing complex clinical workflows autonomously. This opens up great opportunities for medicine - but also raises entirely new questions around safety, accountability, and regulation that we need to address," he adds.

Current medical device regulations were designed for static, narrowly focused technologies that maintain human oversight and do not evolve after entering the market. In contrast, new autonomous and broad-scope technologies have fundamentally different characteristics. They demonstrate greater autonomy, adaptability, and scope than previous AI technologies. As they are capable of autonomously executing complex workflows, they present significant challenges for regulators and developers.

"To facilitate the safe and effective implementation of autonomous AI agents in healthcare, regulatory frameworks must evolve beyond static paradigms. We need adaptive regulatory oversight and flexible alternative approval pathways," says Oscar Freyer, lead author of the publication and research associate in the team of Professor Stephen Gilbert, who leads the Medical Device Regulatory Science group at EKFZ for Digital Health, TUD.

As AI agents' capabilities exceed the scope of current regulatory frameworks, the researchers propose several potential solutions to overcome barriers to their implementation.

Immediate adaptations include extending enforcement discretion policies, where regulators acknowledge a product qualifies as a medical device but choose not to enforce certain requirements. Alternatively, a non-medical device classification could be applied to systems that serve a medical purpose but fall outside the traditional medical device regulation.

Medium-term solutions involve developing voluntary alternative pathways (VAPs) and adaptive regulatory frameworks that supplement existing approval processes. These adaptive pathways would shift from static pre-market approval to dynamic oversight using real-world performance data, stakeholder collaboration, and iterative updates. In case of misconduct, devices may be transferred to the established pathways.

The researchers also discuss long-term solutions, such as regulating AI agents in a similar way to the qualification of medical professionals. In this model, regulation would be carried out through structured training processes, in which systems gain autonomy only after demonstrating safe and effective performance.

The researchers note that while tools such as regulatory sandboxes offer some flexibility for testing new technologies, they are not scalable solutions for widespread deployment due to the resources and commitment required by regulatory authorities.

In their article, the researchers argue that meaningful implementation of autonomous AI agents in healthcare will likely remain impossible in the medium term without substantial regulatory reform. VAPs and adaptive pathways are highlighted as the most effective and realistic strategies to achieve this goal. The authors underscore the need for collaborative efforts between regulators, healthcare providers, and technology developers to create frameworks that meet the AI agents' unique characteristics while ensuring patient safety.

"Realizing the full potential of AI agents in healthcare will require bold and forward-thinking reforms," says Stephen Gilbert, Professor of Medical Device Regulatory Science at the EKFZ for Digital Health at TU Dresden and last author of the paper. "Regulators must start preparing now to ensure patient safety and provide clear requirements to enable safe innovation," he adds.

Freyer O, Jayabalan S, Kather JN, Gilbert S.
Overcoming regulatory barriers to the implementation of AI agents in healthcare.
Nat Med. 2025 Jul 18. doi: 10.1038/s41591-025-03841-1

Most Popular Now

Integrating Care Records is Good. Using …

Opinion Article by Dr Paul Deffley, Chief Medical Officer, Alcidion. A single patient record already exists in the NHS. Or at least, that’s a perception shared by many. A survey of...

Should AI Chatbots Replace Your Therapis…

The new study exposes the dangerous flaws in using artificial intelligence (AI) chatbots for mental health support. For the first time, the researchers evaluated these AI systems against clinical standards...

AI could Help Pathologists Match Cancer …

A new study by researchers at the Icahn School of Medicine at Mount Sinai, Memorial Sloan Kettering Cancer Center, and collaborators, suggests that artificial intelligence (AI) could significantly improve how...

AI Detects Early Signs of Osteoporosis f…

Investigators have developed an artificial intelligence-assisted diagnostic system that can estimate bone mineral density in both the lumbar spine and the femur of the upper leg, based on X-ray images...

AI Model Converts Hospital Records into …

UCLA researchers have developed an AI system that turns fragmented electronic health records (EHR) normally in tables into readable narratives, allowing artificial intelligence to make sense of complex patient histories...

AI Sharpens Pathologists' Interpret…

Pathologists' examinations of tissue samples from skin cancer tumours improved when they were assisted by an AI tool. The assessments became more consistent and patients' prognoses were described more accurately...

AI Tool Detects Surgical Site Infections…

A team of Mayo Clinic researchers has developed an artificial intelligence (AI) system that can detect surgical site infections (SSIs) with high accuracy from patient-submitted postoperative wound photos, potentially transforming...

Forging a Novel Therapeutic Path for Pat…

Rett syndrome is a devastating rare genetic childhood disorder primarily affecting girls. Merely 1 out of 10,000 girls are born with it and much fewer boys. It is caused by...

Meet Your Digital Twin

Before an important meeting or when a big decision needs to be made, we often mentally run through various scenarios before settling on the best course of action. But when...

NHS National Rehabilitation Centre to De…

The new NHS National Rehabilitation Centre will deploy technology to help patients to maintain their independence as they recover from life-changing injuries and illnesses and regain quality of life. Airwave Healthcare...

AI Finds Hundreds of Potential Antibioti…

Snake, scorpion, and spider venom are most frequently associated with poisonous bites, but with the help of artificial intelligence, they might be able to help fight antibiotic resistance, which contributes...

AI can Accelerate Search for More Effect…

Scientists have used an AI model to reassess the results of a completed clinical trial for an Alzheimer’s disease drug. They found the drug slowed cognitive decline by 46% in...