My Basement Flooded — and It Wasn’t an Information Problem

My smartphone received a weather alert from my city: possible heavy rain. The storm began overnight. I wasn’t especially concerned — my sump pumps were only five years old and had always worked well. The city website warned of possible flooding. Search engines and chatbots offered general recommendations. Yet no one addressed the only question that mattered: what should I do right now?

By morning, my basement was flooded with sewage. It was not an information problem. It was a failure of decision support.

Experiences like this are increasingly common — not because people lack information, but because our systems fail to translate information into timely, context-specific decisions. This is where AI agents enter the picture.

The Real Problem: Cognitive Overload

Current AI is powerful at prediction. It can empower weather forecasts, flood maps, sewer capacity models, and city alerts. All the information related to my flood existed. The problem wasn’t ignorance — it was overloaded.

We used to talk about information overload: too much data to find what matters. Now we face something harder: cognitive overload — too much data to judge and act. During the storm, I had to answer: Is this rain normal or dangerous? Should I move things now or wait? Is my sump pump enough? Do I call someone? What matters most in the next 30 minutes? These are judgment calls, not search problems. And humans are bad at judgment under stress.

Now scale this up. Imagine a public health control center during a major flood. Data is streaming in. Alarms are accurate. But the problem is no longer finding information — it’s deciding which signal matters most, which action to take first, and what can safely be ignored. Predictions are not prioritized. No search engine helps you make public health policy under pressure.

From Data-Mining AI to AI Agents

What more can AI do? In the flooding example, imagine a system that continuously monitors both weather and your home context, reasons about your risk, prioritizes actions, acts across time — before, during, and after the heavy rains — and coordinates alerts, insurance, and city services. This system tells you: here is what you should do next, and why.

That is agentic behavior. A classic AI agent can sense the environment (perception), decide what matters (reasoning), and do something in the world (action). This loop — perceive, think, act — is the foundation of the “agent” concept in computer science. But is it enough for messy, real-world public health situations? Early AI agents performed well on controlled tasks but struggled in chaotic human environments. They didn’t understand personal constraints, especially those of vulnerable groups. We need something more.

The Public Health Upgrade: Three Key Features

Public health AI agents need three features beyond the classic perceive-think-act loop.

Compliance: Is this action consistent with laws and guidelines, and will real people follow it? A flood evacuation order is useless if elderly residents can’t physically leave, don’t understand the alert, or fear abandoning their homes. Compliance in public health must include both legal permission and practical adoption. A flooding response agent should know that issuing a digital evacuation alert to a neighborhood of older adults with low digital literacy isn’t truly compliant — it’s compliant on paper only.

Efficacy: Will this action prevent disease, promote health, and save lives — right here, right now? Opening an emergency shelter means nothing if the bus route is flooded and residents can’t get there. High efficacy in a flooding scenario might look simpler: a phone call or door knock the evening before peak rain, focused on moving valuables upstairs, checking sump pumps, and identifying who needs physical help. More alerts and more dashboards do not equal more lives saved. Efficacy is about fitness, timing, and access under uncertainty.

Autonomy: Can the agent act safely on its own while ensuring individuals retain choice and dignity? A flooding agent might monitor local water levels, sewer capacity, and known vulnerable households. If a resident’s basement is at high risk, the agent sends a timely alert to the resident or caregiver without waiting for a call for help. It might schedule sandbag deliveries or arrange volunteer wellness checks. But the resident always decides how to respond. Autonomy here means independence from constant supervision — not independence from human judgment.

The Observer’s Insight

These three features — Compliance, Efficacy, Autonomy — shift AI from risk prediction to risk reduction. They become the logic inside public health agents.

But let me be clear: public health AI agents are not early Artificial General Intelligence. They are domain specific. A flooding response agent can monitor weather, water levels, and hospital capacity. It can recommend deploying shelters or rerouting emergency services. It cannot decide whether to reallocate budgets or override elected officials. Its intelligence ends exactly where human responsibility begins.

My flooded basement taught me something that years of research had already shown me in theory: Public health does not lack evidence. It lacks the systems to turn evidence into action now. AI agents — designed with compliance, efficacy, and autonomy — are how we start building those systems.

Has a flood, storm, or emergency ever left you overwhelmed by information but unsure what to do? What would a “living system” have changed for you? Reply or leave a comment — I’d love to hear your story.

— Min Wu, PhD
Associate Professor, Zilber College of Public Health
University of Wisconsin-Milwaukee