Making faster, better decisions when managing crises is literally your job
Published on April 21, 2025
Disaster management teams are stretched thin—and looking to AI for help. Could what works for them help the rest of us too?
The Swannanoa is a small river just outside Asheville, NC. Normally, it’s about two feet deep. So when a video of a house being carried downstream hit TikTok in late September of last year, people began to grasp just how serious the situation with Hurricane Helene had become. A shipping container floating in the brown water folded around a telephone pole like tinfoil. Just behind it, what looked like an entire apartment building went bobbing along. It gave rise to that uncanny, brain-tingling sensation, seeing horrors usually confined to far away places—tsunamis are supposed to stay in Indonesia—show up in the American backyard.
Helene dumped six months of rain on Asheville in three days. In climatology, there is an effect known as orographic lift where mountains force air to rise and condense, dropping massive amounts of precipitation. The steepness of the topography swelled rivers like the Swannanoa beyond recognition in a matter of hours. Heavy rainfall in the days ahead of Helene’s arrival meant that the ground was already saturated and unable to absorb any more. A region not known for its flooding watched as entire buildings were swept away. At its highest the Swannanoa rose to 27 feet. The death toll in North Carolina alone was 106.
By nature, disaster response work is reactive. Aid organizations, governments, and communities scramble to meet the moment. In the best of scenarios, it’s complex and challenging. Disaster management teams must coordinate and collaborate across multiple organizations, manage limited resources, and make decisions with incomplete information. It’s just like any other office—you have to find data and execute on it—except in this case, there are lives on the line.
Dropbox Dash: Find anything. Protect everything.Climate change is only making everything harder. More disasters means less time to prepare, and current emergency management systems are buckling under the intense, increasing pace of fresh crises. In an effort to keep up, emergency managers—like knowledge workers more broadly—have been turning to AI-powered tools for help. The hope is that AI can take on some of that rising cognitive load, helping responders see patterns faster, allocate resources more effectively, and make better decisions in real time.
With power and telecommunications down, Helene’s first responders relied on handwritten notes to communicate. Pack mules brought in medical supplies. Residents in Black Mountain set up makeshift message boards to track missing persons and coordinate aid. The Mayor of Canton, NC, 20 miles to the west, put it succinctly: "This is 2024, and we are having to deal with this crisis with 1990s technology. [...] It’s heartbreaking.”
“We are talking about augmenting existing processes with AI," says Dr. Mostafavi.
AI won’t stop disasters like this from happening, and it won’t replace the humans who respond to them. But it might help emergency managers see a few hours further into the future, cut through the noise, and make faster, better decisions. That could mean a truckload of food getting to the right neighborhood sooner, an evacuation route staying open just long enough, or one less car getting swept away. In a field where every second counts, even small improvements could mean the difference between life and death.
Dr. Ali Mostafavi, an AI researcher at Texas A&M University, is optimistic about AI’s potential here. His lab has already done some promising work on predicting floods. But he’s appropriately skeptical, too. “Some people go from zero to 100 and think about replacing all the firefighters with AI. No.” said Dr. Mostafavi. Disaster chatbots don’t interest him either.
So what does success look like to Dr. Mostafavi? “We are talking about augmenting existing processes with AI.”
~ ~ ~
As with almost every function in our modern society, emergency management has been professionalized. You can now get an MS in Emergency and Disaster Management at Georgetown. Jacksonville State offers a Doctor of Science in Emergency Management. On its surface, disaster relief presents as a labor-intensive set of tasks: search and rescue, evacuation, food and medicine distribution. But just one layer beneath that is a dense web of logistics. That means planning, processes, workflows. As Tim Manning, former Deputy Administrator at FEMA, put it on a podcast recently: “The way I like to describe emergency management to people who don’t understand it: We tend to be the process experts.”
In California, the Department of Forestry and Fire Protection (CAL FIRE), uses over 1,100 cameras throughout the state to identify new hot spots. An artificial intelligence tool called ALERTCalifornia developed by researchers at UC San Diego scans for smoke or other signs of fire. But as Battalion Chief David Acuña explained, everything still has to be verified by a human.
“When the AI identifies what it believes is smoke or fire, it sends an email to the relevant emergency command center,” Chief Acuña said. “And you know, AI learns. But when it first came out it was identifying, ‘Hey, there’s a fire.’ And we're like, ‘Well, yes. Technically that is a fire, but that's also the sun.’ So, let's ignore that.”
AI tools, no matter how powerful, are only useful if they integrate seamlessly into people's exisitng workflows—without slowing anyone down.
Some of the processes they have place are holdovers from the pencil and paper era, Chief Acuña said. That’s part of what makes AI adoption so challenging. The problem isn’t that AI can’t help—it’s that emergency management isn’t built to absorb new technology quickly.
“Information overload is a big issue during times of crisis,” explained Dr Mostafavi. “One of the reasons that AI tools are not widely adopted is because this becomes another source of data and information that humans need to digest during disasters.” The field runs on experience, intuition, and split-second decision-making—and in a crisis, that means responders rely on what they already know works.
AI tools, no matter how powerful, are only useful if they integrate seamlessly into existing workflows—all the apps and tools people already use—without slowing anyone down. If a system requires extra training, adds another layer of complexity, or delivers insights that don’t immediately translate into action, it’s likely to be ignored. Dr Mostafavi’s goal is to bridge that gap.
~ ~ ~
The majority of flash flooding deaths are vehicle-related, according to the National Weather Service. Drivers are particularly vulnerable because roads can quickly become impassible. People often underestimate how quickly a car can be carried away, even by relatively shallow water. (Hence the slogan with its grim little rhyme: "Turn Around, Don't Drown.")
That’s part of what led Dr. Mostafavi to begin work on the Flood Genome. It’s an AI model built by the Urban Resilience.AI Lab at Texas A&M University, which uses technology to reduce the social, economic, and physical impacts of disasters. The Flood Genome predicts how flooding propagates in roads. The idea is that if emergency management teams know which roads are in danger of flooding in 6-12 hours, they can divert traffic from those areas ahead of time. The Flood Genome also pulls in 50 years of flood data and uses graph analytic methods to suggest evacuation patterns. The goal is to provide near-real-time information to decision-makers as a situation is unfolding.
“Each event has a unique, first-of-a-kind nature,” Dr. Mostafavi said. “Each event catches emergency managers by surprise. That's one of the main challenges.”
This is why many new crisis technologies are built around a single goal: achieving “situational awareness.” That might mean predicting flood spread, monitoring evacuation and rescue efforts, allocating resources, or rapid damage assessment using satellite imagery.
AI helped a food bank get resources to impacted areas faster—a small but meaningful example of using AI to cut through uncertainty in a crisis.
In the days leading up to January’s LA fires, for example, Chief Acuña and his team received a number of notices from the National Weather Service. “There was a red flag warning,” he said. “And then it was upgraded to a PDS, a Particularly Dangerous Situation.” So CAL FIRE started moving resources—”engines, crews, and heavy equipment”—up and down the state as needed. “We sent a bunch of resources down to Southern California ahead of time, just in case there was an issue,” Chief Acuña recalled. “And so they were staged in Southern California when this all started.”
In every case, the key is reducing potentially lethal blind spots—whether AI is involved or not. That’s the animating purpose for the solutions that come out of Dr. Mostafavi’s research. Those solutions are then developed and scaled into products for public sector adoption, where they’ve already had some promising real-world applications.
During Hurricane Beryl, a tropical cyclone that reached Category 5 last summer, two million people in Houston lost power for several days. Food access was a pressing issue. So Dr. Mostafavi partnered with the Houston Food Bank to help coordinate distribution with AI.
“For the first 72 hours they had no clue about which areas are struggling with food until they start getting 211 calls and start getting information from grocery stores. That takes time and during those hours people suffer,” Dr. Mostafavi explained. “We monitored disrupted access to food based on our graph analysis model and provided daily updates to emergency management partners about which areas seemed to be struggling the most.”
The data helped the food bank get resources to the impacted areas faster. It was a small but meaningful example of how AI could cut through uncertainty in a crisis—turning guesswork into slightly-more-informed-work. More importantly, the work made a real difference in people’s lives. Dr. Mostafavi’s team received a thank you note from Houston Food Bank a few days after the power came back on.
The Author
.png/_jcr_content/renditions/hero_square%20(1).webp)
.png/_jcr_content/renditions/hero_wide%20(1).webp)
.png/_jcr_content/renditions/hero_square%20(3).webp)
.png/_jcr_content/renditions/blog%20(1).webp)
.png/_jcr_content/renditions/hero%20(1).webp)
.png/_jcr_content/renditions/hero_wide%20(1).webp)
.png/_jcr_content/renditions/1080x1080%20(1).webp)
.gif)
.png)
.png)
.png)
.jpg)
.jpg)
