When it comes to artificial intelligence, nothing better illustrates the technology’s promise and limitations than its approach to forecasting the weather.

Along with calculating missile trajectories, weather prediction was a major motivation behind the invention of computers. Today, with the explosion of AI, it’s become more accurate than ever — but can it ever achieve perfection?

Chaos theory suggests not. There are simply too many variables affecting global weather patterns, and neither human nor machine will ever fully understand the ways in which they interact. What modern-day science can do, though, is assess the risk of extreme weather events to supply chains, so that businesses can prepare for any eventuality.

ClimateAi, creator of a software platform for climate risk modeling, is applying AI to the problem of accurate weather forecasting and its impact on retail sales. The company’s model, known as FICE, for Foundational Intelligence for Climate & Economy, was designed “to quantify the timing, duration and magnitude of demand spikes and suppressions related to weather,” says data science lead Dave Farnham. It combines traditional sales-related inputs such as seasonal demand patterns with detailed information on local weather and geological conditions.

The company views the tool as a kind of AI-driven “Waffle House Index,” the informal metric that pins the severity of a storm to the number of Waffle House restaurants that remain open within an affected area. It’s used by the Federal Emergency Management Agency (FEMA) and other aid responders to assess the required level of disaster-response services.

ClimateAi says FICE can be especially valuable in helping food and beverage retailers keep shelves stocked, and distributors to plan labor and inventory-management strategies, in the face of weather-related supply disruptions. Included in the data that feeds the model is information on air and soil temperatures as well as crop conditions.

Farnham says FICE functions both as a “forward- and backward-looking” model. In the case of the first, it analyzes demand trends across grocery stores within a given region, to help determine how much product to keep in stock at each location. In the second, it can draw on historical data to aid in longer-term financial planning and analysis. In the process, he says, FICE seeks to “disentangle” all of the elements that lead to variability in weather patterns and consumer demand.

The data comes from multiple sources, including ClimateAi’s own store of intelligence, government information, and such macroeconomic data as unemployment rates and energy prices — in short, just about anything that’s likely to impact consumer behavior. In addition, Farnham says, the company draws on a third-party data provider to track brand sales and credit-card activity across more than 100 sectors. “It allows us to actually see how much people are spending.”

The power of AI makes it possible to analysis massive amounts of weather-related data. “Computational increases make it easier to fit large models,” Farnham says. But FICE isn’t using generative AI or the large language models that underlie it. Instead, it’s relying on machine-learning systems and “agentic” workflows to allow users to query the model.

As sophisticated as all this sounds, the FICE model is “still in a learning curve” in terms of being able to correctly interpret the data. (Indeed, the very term “machine learning” underscores how it becomes more accurate with experience.) Farnham says it’s important that FICE not be a “black box” — a system that’s unable to explain the basis for its insights. Customers, he adds, “want some indication of what’s driving that.”

Climate Ai also provides long-term projections of climatic conditions, but that’s where the limitations of any predictive system, whether human- or machine-driven, come to the fore. Farnham acknowledges AI’s “fundamental limits” in this regard, leaving forecasters to couch their insights in the form of probabilities rather than precise outlooks.

“We’re trying to provide tools for folks to make decisions in the context of uncertainty,” he says. “Our job is to keep shrinking that uncertainty in terms of better and better forecasts. But it’s never going to be perfect.”