Imaging Pearls ❯ Deep Learning ❯ Agentic AI
|
-- OR -- |
|
- An agentic system, in contrast, may operate like a full - fledged clinician charged with caring for the patient as a whole. Having defined shared health goals with the patient, it could be tasked with reducing their 10 -year atherosclerotic cardiovascular disease risk score. It might calculate current risk, order follow -up laboratory tests, ensure that the patient was prescribed and is taking statins, recommend a smoking cessation program, and enroll them in supplemental employee insurance that covers nicotine patches.
A Typology of Generative Health Care Artificial Intelligence — Definitions and Policy Implications
David Blumenthal , Vivian S. Lee
NEJM AI 2026;3(6) - Agentic AI systems can operate at scale. Programmed at the level of a chief medical officer, such a system could be designed to achieve target blood pressure control across a health system’s patient population, thereby optimizing value -based reimbursement. If granted the autonomy and access, it might independently decide to scan all electronic health records, identify patients with blood pressures that exceed target levels, and prompt clinicians to take patient-specific steps. With clinician agreement, it could prescribe medications, schedule tests and appointments, and inform patients, repeating this process — under human supervision — until population -level goals are achieved.
A Typology of Generative Health Care Artificial Intelligence — Definitions and Policy Implications
David Blumenthal , Vivian S. Lee
NEJM AI 2026;3(6) - AI agents pose different policy and management challenges. As agents produce actions as well as information, they offer both greater risks and greater benefits than the LLMs on which they are built. An agent that is trained and marketed to accomplish a specific clinical task, such as minimizing cardiovascular risk, will likely be classified by the U.S. Food and Drug Administration (FDA) as Software as a Medical Device (SaMD) and be subject to FDA regulation. If the FDA were not to invoke this authority, it would likely face legal challenges demanding that it enforce current law.
A Typology of Generative Health Care Artificial Intelligence — Definitions and Policy Implications
David Blumenthal , Vivian S. Lee
NEJM AI 2026;3(6) - Agentic systems raise additional issues that require alternative channels for review and oversight. One is that they may be trained to accomplish diverse tasks, such as comprehensive care for patients or complex organizational goals. In the former case, they may pose unprecedented regulatory challenges for the FDA because of the technical challenges in conducting premarket assessment of devices that undertake a wide range of clinical activities and that can change and learn over time. Novel regulatory or oversight pathways may be needed to deal with the capabilities of agentic systems.
A Typology of Generative Health Care Artificial Intelligence — Definitions and Policy Implications
David Blumenthal , Vivian S. Lee
NEJM AI 2026;3(6) - The distinctions among LLMs, agents, and agentic systems matter because their inherent properties, capabilities, and uses raise different quality and safety issues — and different challenges for oversight. Some forms may be subject to minimal direct oversight. For others, oversight may vary with use case. Relevant factors include clinical versus nonclinical, level of direct patient risk, and existing regulatory authorities. As they are used more widely, increased scrutiny of the underlying values embedded in these models — values that influence their recommendations and behaviors — may prompt further ethical review. In addition, agent -to -agent interactions will create new challenges as patient, provider, pharmaceutical, and insurer agents interact across administrative matters and clinical care.
A Typology of Generative Health Care Artificial Intelligence — Definitions and Policy Implications
David Blumenthal , Vivian S. Lee
NEJM AI 2026;3(6) - Understanding—or interpretability—of AI need not mean grasping every line of code or every neural-network parameter. Just as we study human behavior at multiple levels, from neuroscience to psychology to sociology, AI principles and operations can be explored and understood at varying levels. Full mechanistic understanding may remain elusive, but science is rarely all-or-nothing; partial understanding is still useful. What makes such understanding urgent is not a demand for completeness but a practical need: As capabilities accelerate, even imperfect causal insights into AI systems may let us detect risks early and intervene before harm compounds.
A narrowing window to understand AI
Eric Horvitz and Robert West
Science 392 (6802), June 2026. DOI: 10.1126/science.aei3167 - One trend challenging understanding is the rise of AI-directed AI design. AI systems are now designed and refined by AI systems through recursive cycles that can outpace human understanding and unfold in high-dimensional spaces that resist intuition. The result isgrowing operational opacity: Performance improves, while insight into how it is achieved diminishes. To promote human insight and control, AI systems that contribute to their own design should produce explanations and tools that make their architecture and operation intelligible to humans. Otherwise, opacity may emerge as an unintended consequence of the design process itself.
A narrowing window to understand AI
Eric Horvitz and Robert West
Science 392 (6802), June 2026. DOI: 10.1126/science.aei3167 - More subtle is the possibility that we will gradually lose interest in understanding and guiding AI. As AI systems become deeply embedded in human environments, they may respond to preferences but also shape them. Systems optimized for engagement or approval may reduce friction and discourage scrutiny. Over time, curiosity and skepticism may erode, leading to neglect and acceptance. Preserving human agency must therefore remain a central goal. It is not enough to monitor how AI systems behave. We must also understand how they shape human goals and judgment, and ensure that people retain the capacity and motivation to question, audit, and guide them.
A narrowing window to understand AI
Eric Horvitz and Robert West
Science 392 (6802), June 2026. DOI: 10.1126/science.aei3167 - The goal is not just more capable AI, but AI that is more intelligible, accountable, and aligned with human aims. The window for achieving that future is narrowing. Without sustained efforts to keep AI intelligible, we may come to depend on systems that we can neither adequately understand nor effectively guide—transforming the relationship between people and the systems they create.
A narrowing window to understand AI
Eric Horvitz and Robert West
Science 392 (6802), June 2026. DOI: 10.1126/science.aei3167
- Interestingly, while near-term administrative investments remain a priority to achieve financial recovery, executives continue to illustrate a strong interest in clinically focused, and clinically-adjacent AI investment domains signaling an eventual transition toward more investments at the point of care. When asked to select their top 3 domains for future AI investment, 47% chose clinical decision support, 43% chose clinical documentation and/or patient engagement/ communication, and 40% chose patient flow optimization. These domains represent the top 5 priorities for future investment, outside of revenue cycle management, which was also selected by 47% of executives .
At the Frontier: Gauging Health Care’s Readiness for Agentic AI Innovation
Peter Durlach, C.V.P.,1 Kees Hertogh,2 Wesley Adams, M.P.H.,3 Emma Feeney,3 and Coltin Ball, M.P.H.3
NEJM AI DOI: 10.1056/AI-S2501336 Published: December 24, 2025 - Alongside current use cases and investment domains, respondents see significant potential for agentic AI to close the capacity gap between patient care demands and constrained human resources. Thirty-one percent of respondents viewed agentic AI as an emerging enabler of future transformation. As expected, active deployment of agentic tools remains limited (3%). Instead, leaders remain focusedon pilots (43%) and exploratory efforts (33%). Over the next five years, leaders anticipate that agentic AI will evolve from concept to catalyst, enhancing productivity, expanding workforce capacity, and enabling more adaptive, resilient care delivery.
At the Frontier: Gauging Health Care’s Readiness for Agentic AI Innovation
Peter Durlach, C.V.P.,1 Kees Hertogh,2 Wesley Adams, M.P.H.,3 Emma Feeney,3 and Coltin Ball, M.P.H.3
NEJM AI DOI: 10.1056/AI-S2501336 Published: December 24, 2025 - One academic health system piloting a multi-agent tumorboard application demonstrated both the promise and the challenge of this new paradigm. The initiative connected radiology, pathology, and genomic analysis agents to reason collaboratively across patient data, allowing community hospitals to access expertise previously confined to academic settings. Yet it also surfaced unresolved questions about responsibility: Who ultimately approves the recommendations an agent produces, and how are reasoning pathways validated for clinical safety?
At the Frontier: Gauging Health Care’s Readiness for Agentic AI Innovation
Peter Durlach, C.V.P.,1 Kees Hertogh,2 Wesley Adams, M.P.H.,3 Emma Feeney,3 and Coltin Ball, M.P.H.3
NEJM AI DOI: 10.1056/AI-S2501336 Published: December 24, 2025 - Despite the widespread recognition of agentic AI’s potential, most health systems have not deployed agents at the enterprise level. The readiness paradox, where 60% of executives identify capability-building as both the most critical enabler and the top implementation challenge, reveals the health care industry knows its destination but lacks effective frameworks for transformation. Our findings suggest the barriers to agentic AI deployment are interdependent: workforce readiness depends on clear governance frameworks, which only scale with mature data infrastructure. To bridge the agentic innovation-implementation gap, health systems must address governance, data, and workforce transformation in concert. Those who successfully integrate these three imperatives will move to operational scale, while those that approach them as siloed initiatives will remain stuck in pilot.
At the Frontier: Gauging Health Care’s Readiness for Agentic AI Innovation
Peter Durlach, C.V.P.,1 Kees Hertogh,2 Wesley Adams, M.P.H.,3 Emma Feeney,3 and Coltin Ball, M.P.H.3
NEJM AI DOI: 10.1056/AI-S2501336 Published: December 24, 2025