Agentic AI can help hospitals prepare for unprecedented weather

Abstract

As climate change intensifies, healthcare systems will increasingly face unprecedented climatic emergencies surpassing assumptions underlying their emergency protocols. While conventional scenario-based preparedness and contingency planning will remain relevant for most events, they may falter in unexpected crises. In such cases, a threshold-based framework could become essential for hospital response and resilience. Emerging AI agent technology offers opportunities to develop systems that could drive this paradigm shift in hospital climate preparedness.

Over the course of the 21st century, climate change is anticipated to significantly increase the frequency and intensity of extreme weather events¹. This trend will likely bring not only more unprecedented anomalies, but also a rise in “perfect storm” scenarios—instances where multiple extremes coincide². Although rare, the most severe climatic emergencies could be considered “climatic black swans” (after Taleb’s black swan³): unforeseen, high-impact phenomena that exceed the assumptions underpinning existing emergency preparedness plans, thereby jeopardizing the stability of critical public-serving systems. Even infrastructures and facilities that are well prepared for “extreme but reasonable” climatic scenarios could find themselves overwhelmed by climatic black swans—events that would be anything but reasonable.

One striking example was the “Heat Dome” event that scorched the Pacific Northwest in the summer of 2021⁴. This once-unthinkable heatwave brought record-shattering temperatures of almost 50 °C in Canada, surpassing all “extreme but reasonable” heatwave scenarios for which critical public-serving systems were prepared. Consequently, various infrastructures were damaged or compromised: roads and rail tracks buckled, electrical insulation in high-voltage systems melted, and bridges had to be sprayed with water to prevent structural damage, in what turned out to be the most expensive natural disaster in Canadian history⁴. The most tragic consequence, however, was the loss of human life, including the death of patients in hospitals and long-term care facilities⁵. The intense heat severely strained healthcare facilities, exacerbating existing medical conditions, overwhelming staff, and overloading intensive care units^6–8. The crisis was amplified by the lingering effects of the COVID-19 pandemic on healthcare systems, which were still recovering from staffing depletion and personnel burnout. In some hospitals, essential equipment such as CT and MRI machines failed due to the breakdown of their cooling systems, and indoor temperatures in various departments reached unbearable levels, impairing staff performance and worsening patient conditions⁹.

Post-event reviews indicated that many decision makers, operating under extreme stress, did not activate available emergency protocols in time, because the heatwave did not resemble any scenario for which they had trained⁹. For instance, only one hospital declared a mass casualty event; others delayed their response and consequently failed to meet surging demands in their emergency departments⁹.

As the global climate continues to change, failure-intolerant systems such as hospitals and other healthcare facilities must revise their emergency preparedness plans, acknowledging that a climatic emergency defying all projections will eventually occur.

Climate resilience in healthcare

The World Health Organization (WHO) urges healthcare facilities to improve their resilience to climate change by identifying structural, non-structural, and functional weaknesses that may compromise their ability to withstand natural disasters¹⁰. These recommendations call for general preparedness and better planning to strengthen a facility’s resilience to a broad range of potential climatic hazards¹⁰. The WHO also states that “building climate resilience at the health facility level requires an understanding of current and projected climate conditions¹¹.” These projections are essential for pre-disaster risk assessments, which guide the development and implementation of preparedness and response programs that are tailored to specific climatic hazards¹².

Relying on projected extreme weather scenarios is a common approach in emergency planning, which typically involves preparing and training for one or several plausible extreme events¹³. For instance, the National Oceanic and Atmospheric Administration’s (NOAA) new tool for helping communities plan for extreme heat is based on a single reference scenario per community, one that must be validated by a local weather or climate expert to ensure it is realistic¹⁴.

While scenario-based contingency planning (SBCP) is popular, practical and cost-effective, it has inherent limitations. It functions well when events remain within—or close to—the bounds of predefined reference scenarios, i.e., the “extreme but reasonable” events that lead the organization’s preparedness efforts. However, when a crisis evolves rapidly and significantly exceeds the reference scenario, SBCP provides little guidance, forcing managers to improvise under pressure.

Hospitals and climatic black swans

Hospitals are probably the most complex healthcare facilities—and therefore particularly challenging to manage during emergencies—since they function as intricate ecosystems comprised of interdependent subsystems, infrastructure, equipment, supply chains, and personnel. While redundancy is built into many of their components, each element, including its backups, has its own operational thresholds beyond which performance deteriorates or fails altogether. Often, these thresholds are not fixed values but ranges, with an increasing propensity for failure as conditions worsen. During a thermal black swan event, for example, multiple components may simultaneously exceed their thermal thresholds, triggering cascading malfunctions that could significantly interrupt hospital activity.

In essence, climatic black swans are deeply surprising and therefore demand rapid, high-stakes decision-making. Ideally, hospital administrators would have real-time access to detailed knowledge of each component’s thermal and other operational thresholds, allowing them to proactively identify and manage vulnerabilities. Yet, even with such knowledge, effective and timely crisis analysis would remain difficult, especially when lead time is limited. So how can hospitals prepare for “unknown unknowns”–events beyond the scope of any preparedness plan?

The promise of agentic AI

The recent advent of agentic AI, a new paradigm in the field of AI, offers an opportunity to transform the way hospitals deal with the growing uncertainties of climate change. Agentic AI systems are based on individual or multi-agent architectures that incorporate memory, autonomous reasoning, planning, the use of external tools, and decision-making capabilities, to pursue complicated goals with minimal human oversight¹⁵. At the forefront of the agentic AI revolution are frameworks like Microsoft’s recently introduced Azure-based toolkit for developing and supervising multi-agent systems¹⁶. Such frameworks incorporate sophisticated monitoring and advanced safety features, which are essential for bridging the trust gap often associated with AI deployment in sensitive environments such as healthcare facilities.

Already, AI-driven predictive modeling and real-time AI-powered data analysis are beginning to transform emergency response by enhancing early warning systems, optimizing resource allocation, and augmenting decision support systems¹⁷. However, AI-driven emergency management tools rely on models that are trained on historical data, thus limiting their effectiveness in unprecedented scenarios, where their reliability becomes questionable¹⁷. To better support decision-making during climatic black swans, AI-enhanced emergency management must evolve beyond the confines of traditional SBCP and the reliance of foundation models on past data.

This is where agentic AI can make a critical contribution. Like novel AI-driven emergency management tools, agentic AI can curate and analyze vast amounts of data, integrate external information sources, and make autonomous real-time decisions. However, its unique value in the context of climatic black swans lies in its ability to step in when SBCP loses relevance and foundation models become unreliable. In such cases, an agentic AI system could employ a more adaptive “threshold-based planning” (TBP) approach, grounded in a comprehensive database of thermal and other operational thresholds for all—or most—hospital components. A multi-agent AI system¹⁸ would use this database (Fig. 1B) to predict the behavior of each component under any thermal or other environmental condition. By continuously monitoring internal hospital systems, sensor data, and weather forecasts (Fig. 1C, D), such a system could anticipate failures, issue early warnings, and detect cascading risks—even in the face of unprecedented events (Fig. 1E). Unlike conventional SBCP or AI-based systems, which are inherently constrained by the assumptions and foresight of human designers, a TBP multi-agent AI system would operate independently of event likelihood, offering robust performance under any thermal or other environmental condition.

Fig. 1. A resilience-enhancing AI framework that could enable hospitals to prepare for and manage unprecedented, unforeseen extreme weather events (“climatic black swans”).

A An underlying multi-agent subsystem that models all hospital components and functions and their interconnections (components = equipment, materials, infrastructure systems, capacity, etc.), to create a blueprint guiding all other agents in the system. B A comprehensive database of probable operational thresholds for optimal performance and failure limits, for all or most of the hospital’s components. The database also includes existing emergency protocols and contingency plans. C Continuous monitoring of hospital system and sensor data (e.g., temperature readings, inventory, staff rota, occupancy). D Real-time updated weather forecast. E A multi-agent subsystem integrating diverse data, issuing timely warnings, and predicting issues and malfunctions. F Alerts from subsystem E inform hospital management and administrative staff decisions for preparations (G). When the climatic black swan event begins (H), crisis management (I) is continuously supported by insights and alerts from subsystem E. J Starting from the initial early warning, an agentic system collects data from all parts of the system. This may include written communications and telephone conversations among hospital staff. Post-event, this data is analyzed and distilled into an accessible report evaluating the hospital’s emergency response, identifying successes and areas for improvement.

The TBP concept has been recently applied in the humanitarian field, where frameworks such as Anticipatory Action and Forecast-based Financing demonstrate how predefined thresholds and triggers can facilitate proactive disaster response. These approaches link climate and weather forecasts to specific thresholds that, when crossed, automatically release funds or activate interventions before crises escalate and observable damage occurs^19,20. In this sense, a TBP multi-agent AI system would apply similar impact-based forecasting logic, using predefined thresholds to trigger early action.

Manually creating a threshold database for numerous hospital components would be daunting, due to the amount of work required and the potential lack of relevant manufacturer specifications, especially for legacy systems. When operational thresholds for medical equipment must be assessed in the absence of manufacturer data, established risk management practices recommend gathering data from similar devices or systems and applying cross-functional expert judgment²¹. AI agents can accelerate and expand this process by autonomously aggregating regulatory filings, adverse event reports, technical documentation, patent applications, and information from comparable devices or systems^21,22. If the necessary information remains unavailable, an AI agent may contact the manufacturer directly or escalate the case to human experts for assessment. When the agentic system identifies critical components with particularly narrow or uncertain operational thresholds, it may recommend targeted testing to empirically determine actual thresholds, which could exceed conservative manufacturer declarations²³. Eventually, this database will be made available to the agentic system using structured RAG (Retrieval-Augmented Generation), a technique that grounds large language models (LLMs) in accurate domain-specific data²⁴.

TBP multi-agent AI systems could be adapted for climate resilience across a wide range healthcare facilities and organizations. The degree of autonomy granted to such systems would depend on organizational needs and on AI technology maturity. However, current implementation would probably face several challenges¹⁵: consistent alignment of AI agent goals with organizational goals, integration with legacy systems, and the ethical concerns and liability issues that might arise in high-stakes contexts such as healthcare systems.

Perhaps one of the most significant challenges for integrating an AI-driven TBP system into hospital climate preparedness is the opaque, “black box” nature of AI models. During an emergency, where mistakes may have life-or-death consequences, limited explainability (the capacity for AI systems to make their decision processes understandable to humans) and lack of transparency could make it difficult for human operators to understand, trust, or appropriately override AI guidance²⁵.

The WHO highlights explainability as essential for safe and effective use of AI in healthcare, enabling clinicians and administrators to understand and validate system outputs²⁶. More broadly, international guidelines such as the UNESCO Recommendation on the Ethics of Artificial Intelligence²⁷ emphasize the importance of transparency and human-centered oversight in all AI applications. However, to earn trust in hospital settings, agentic AI must not only provide explainable outputs, but also demonstrate to healthcare managers and stakeholders robust, reliable performance and alignment with ethical and governance frameworks, through testing in controlled simulations and drills.

A new path to resilience

The core strength of a TBP multi-agent AI system lies not in simply extending conventional SBCP, but in its proactive capacity to detect latent vulnerabilities and orchestrate real-time responses to unforeseen events. Rather than supplanting SBCP, the TBP framework will complement it: as emergencies deviate further from “extreme but reasonable” conditions, TBP multi-agent AI systems will dynamically adapt, refine, and even override existing action plans.

Thus, while SBCP will remain vital for preparing and training for anticipated climatic scenarios, TBP multi-agent AI systems could allow various complex and critical public-serving systems to endure environmental conditions once considered unimaginable.

Author contributions

M.G. conceived the original idea for the article and drafted the initial manuscript. M.G. and C.R. revised and refined subsequent versions. All authors have read and approved the manuscript.

Data availability

No datasets were generated or analysed during the current study.

Competing interests

The authors declare no competing interests.