Column

Editorial

This month, I had the privilege of participating in a Healio Exclusive with Allan Gibofsky, MD, JD, MACR, FACP, FCLM; Kenneth G. Saag, MD, MSc; and Grace C. Wright, MD, PhD; on the emergence and use of ambient AI in the exam room.

As many of you know, I am an enthusiastic supporter of ambient AI and now use it in virtually every patient encounter. Along the way, I have learned a great deal, not only about the technology itself but also about how it can reshape the clinical experience. I encourage you to read the full discussion, which highlights many of the nuances, opportunities and differing perspectives surrounding its use.

At this point, the evidence is increasingly clear. Clinical studies have demonstrated ambient AI can improve efficiency by substantially reducing documentation burden. More importantly, it has the potential to enhance the experience of both clinician and patient by freeing us from the tyranny of the keyboard and allowing us to refocus on the human connection at the center of the clinical encounter.

I have much more to say about that topic, but for now I would like to focus on what comes next.

To begin, I direct your attention to a thoughtful narrative review published in June in The Lancet Rheumatology by Venerito and colleagues on the next phase of AI in our field: agentic AI. Unlike the large language models that have already entered clinical care and research — tools such as OpenEvidence, which has become one of my closest companions in clinic — agentic AI represents a fundamental shift. These systems are not merely passive assistants that respond to prompts. Rather, they are capable of operating autonomously, coordinating multiple tasks across disparate data sources, and potentially taking action within the care pathway itself.

Agentic AI systems can integrate information from laboratory data, imaging studies, pathology reports, pharmacy records, patient-reported outcomes and clinical documentation. Depending on how such systems are designed and regulated, they may not only identify important clinical signals, but also initiate downstream actions.

Venerito and colleagues argue that rheumatology may be uniquely positioned to benefit from these technologies. Unlike critical care specialists, who often must make immediate decisions in rapidly evolving circumstances, rheumatologists manage chronic diseases that unfold over years. Our specialty depends on the integration of longitudinal data from multiple domains: laboratory trends, imaging findings, serial examinations, patient-reported outcomes, and treatment responses.

Consider, for example, the patient with inflammatory arthritis. An agentic system could continuously monitor laboratory results, imaging studies, medication adherence and patient-reported symptoms from messages to the doctor or remote monitoring, synthesizing these data into actionable insights. It could identify subtle evidence of disease progression, flag concerning trends, recommend intervention and perhaps eventually initiate predefined care pathways, particularly in settings where rheumatology expertise is limited or unavailable.

Beyond direct patient care, agentic AI may autonomously perform many of the administrative tasks that consume our time today, including prior authorizations, clinical trial screening, treatment pathway management and ongoing monitoring of patient communications.

The future sounds promising. Yet it also raises profound questions.

For many years I have been fascinated by the architecture of clinical reasoning. Rheumatologists rely heavily on what psychologist Daniel Kahneman, PhD, described as “system two” thinking: deliberate, analytical and effortful reasoning that underpins complex diagnostic and therapeutic decision making.

Now imagine training in an environment where agentic AI performs an increasing proportion of those complex cognitive tasks. How will trainees develop the expertise required to supervise these systems if they are never forced to engage in the reasoning themselves? How will they recognize when an autonomous recommendation is flawed? The challenge extends far beyond the hallucinations that currently dominate discussions of generative AI.

We must also consider the emergence of interconnected networks of autonomous agents communicating with one another and executing clinical actions across health care systems.

For that reason, I was reassured by the authors’ vision of a tiered framework for supervision and regulation that spans every level of the health care ecosystem. They argue that rheumatologists must develop both the declarative knowledge to understand these systems and the procedural knowledge to determine when autonomous execution is appropriate — and when it is not. In other words, we must remain firmly in the loop, even as the capabilities of AI continue to expand.

One day, we may look back on today’s debates about ambient AI-generated notes as the relatively simple era of artificial intelligence in medicine. If so, these may prove to be the halcyon days before a much more transformative wave arrives.

The beginning of the beginning is already here. The question is not whether agentic AI will influence rheumatology, but whether we will be prepared to guide its development, govern its use, and ensure that it ultimately serves our patients rather than simply our workflows.

That’s my take. What’s yours? Please share your thoughts with me at calabrl@ccf.org or at rheumatology@healio.com.

For more information:

Leonard H. Calabrese, DO, is the Chief Medical Editor of Healio Rheumatology, Professor of Medicine at the Cleveland Clinic Lerner College of Medicine of Case Western Reserve University, and RJ Fasenmyer Chair of Clinical Immunology at the Cleveland Clinic.