AI Cuts Food Waste 18% at Cub, Fuel Emissions 2% at Heidelberg Plant

Produce shrink at Midwest grocery chain Cub fell 18 percent after a full rollout of AI-driven demand forecasting from San Francisco software firm Afresh — one of a handful of deployments now putting hard numbers behind the AI-for-sustainability debate.

The results, reported by Cub's chief information officer, followed a three-month pilot during which Afresh's platform lifted sales by more than 2 percent and cut inventory holdings by 7 percent. Cub, which historically relied on employees' experience for forecasting, ordering and store operations, is now on track to avoid more than 1,200 metric tons of greenhouse gas emissions, 43 million gallons of water and 2 million pounds of food waste annually.

That is built, operating capacity — not a pledge. And it frames a judgment call sustainability practitioners now face: do the carbon emissions and other impacts of artificial intelligence outweigh the potential benefits? The calculus remains unclear, partly because some impacts, such as emissions generated during training runs, get omitted from some calculations.

The deployments below suggest that uncertainty shouldn't stop practitioners from exploring the tools.

What did AI do at the cement plant?

Heidelberg Materials wanted to cut emissions at its Mokra cement plant in the Czech Republic, in an industry where 30 percent of a plant's emissions come from fuel combustion. It deployed software firm Gigaton's AI platform on top of the plant's existing advanced process control systems, adjusting fuel mix, kiln speed and oxygen levels.

The goals: lower fuel costs, lower emissions, better clinker quality. In the first month of continuous operation, the plant recorded:

• a 4 percent cut in average fuel cost;
• a 2 percent reduction in fuel-derived emissions.

Lower heat consumption also let the team shift the fuel mix away from heavy fuel oil toward cheaper, less carbon-intensive sources. One month of data is early, but the direction is clear.

How far along are Unilever's digital twins?

AI-enabled digital twins — virtual replicas of physical systems fed by real-time shop-floor data — are the second working application. This June, Unilever and Accenture announced a multi-year partnership to scale digital twins across manufacturing of soap, detergent, deodorant and food, targeting 40 live-data twins that reduce the process variabilities behind waste, stoppages and excess energy use.

Two twins already run in production:

• At Raeford, North Carolina, a twin of deodorant stick production for Dove, Degree and Axe predicts 95 percent of process flow restrictions and delivers a 20 percent waste reduction.
• At Haldia, India, a twin optimizing fan speeds, temperature settings and moisture controls has cut thermal energy use over two years.

The 40-twin target is the number to track: it separates an announced program from the two facilities already reporting results.

What comes next: nature-based risk analysis

The third application is still emerging. A method known as Nature Adjusted Economic Valuation (NAEV) uses AI to help extractive, manufacturing and other resource-intensive companies assess their reliance on ecosystem services — water, timber — and the financial hit they would take if those services became unavailable.

AI compiles and analyzes vast volumes of data from field studies and computational ecosystem models, projecting how far ecosystems sit from degradation and exhaustion. The output feeds models of revenue loss, business disruption costs and weighted average cost-of-capital impacts. Pilot studies are underway and early results show potential; before widespread AI use, such analyses would have been prohibitively time-consuming and expensive.

The takeaway across all three fronts is straightforward: measured against avoided waste, fuel and water, the upside of AI use looks likely to outweigh its own impacts — but the pilot results, not the promises, are what carry the argument so far.