One Third of Everything We Grow

Roughly 1.3 billion tons of food is lost or wasted globally every year. That figure, from the FAO’s 2011 report and updated periodically since, is so large as to be almost meaningless as a mental image. To make it concrete: it’s enough food to feed the approximately 800 million people who are currently classified as chronically hungry, multiple times over. Every year.

The 1.3 billion tons breaks into loss (waste occurring before the retail stage, primarily in production, post-harvest handling, and processing) and waste (occurring at the retail and consumer stage). Lower-income countries lose a higher share in production and post-harvest; higher-income countries waste more at the retail and consumer end. The aggregate is roughly split between these two categories, which matters because the solutions are different.

Post-harvest loss in sub-Saharan Africa is primarily a storage and infrastructure problem: crops spoil because cold storage doesn’t exist, or is too expensive, or isn’t accessible to smallholder farmers. Consumer food waste in Europe and North America is primarily a behavioral and incentive problem: people buy more than they eat, retailers overstock to ensure full shelves, and expiration date labeling conventions are systematically conservative in ways that cause people to discard edible food.

AI is useful for the logistics and prediction aspects of both categories. It is less useful for the behavioral and infrastructure ones.

The commercial case for AI food waste reduction in retail is strong and has been the primary driver of adoption. A grocery chain with $5 billion in annual fresh produce sales and a 12% shrink rate (industry average for produce) is losing $600 million annually to waste. A system that reduces shrink by 20%—to around 9.6%—saves $150 million. Against that value, significant software investment is justified.

The AI tools deployed in retail fresh food management work primarily on demand forecasting and markdown optimization. Better demand forecasting—predicting how much of each item to order, by store, by day—is the most direct lever. An overordered produce section is a waste problem in waiting. AI that improves forecast accuracy at the item-store level by 20-30% compared to baseline systems (which are often simple seasonal averages or vendor-managed inventory defaults) reduces the structural over-ordering that drives shrink.

Markdown optimization is the correction mechanism when over-ordering has already occurred. The decision of when to discount approaching-expiry items, and by how much, to maximize the probability of selling them before they spoil rather than discarding them, is a dynamic pricing problem. Algorithms can run this optimization continuously across a full product catalog in a way that human buyers cannot. Wasteless and Afresh are among the companies that have built specialized grocery AI for this purpose; Afresh’s clients have reported shrink reductions of 25-30% in pilot deployments.

The upstream supply chain waste is harder to address through AI because it’s more distributed and more heterogeneous. Field-level food loss—produce that is discarded during or immediately after harvest because it fails cosmetic quality standards—is estimated at 20-40% in some fresh produce categories. A large share of this is driven by retailer specification requirements (diameter minimums, color uniformity standards) that have no bearing on nutritional quality or food safety but reflect consumer preferences in affluent markets.

AI-powered grading systems—computer vision sorting lines that evaluate produce on dozens of parameters simultaneously—have the interesting dual-use potential here. They can be used to enforce tighter cosmetic standards (eliminating more “imperfect” produce), or they can be used to enable more flexible grade tiers that find buyers for produce that would previously have been discarded. Which way this technology cuts depends on the business decisions made by the companies operating the sorting lines.

The restaurant industry has been the most dramatic commercial deployment of food waste AI in the last five years. Kitchens are environments with high labor costs, fast-changing inventories, and enormous variation in demand by day and time. The combination of point-of-sale data analysis, inventory tracking, and AI demand forecasting specifically for kitchen prep quantities has shown consistent waste reduction in pilot deployments.

Too Good To Go, the Danish-founded app that allows restaurants and retailers to sell surplus food at reduced prices at end of day, has expanded from a consumer behavior app to an AI-integrated food management platform. Their tools analyze historical sales patterns and weather data to predict which days will generate surplus and prompt chefs to adjust prep quantities proactively. The surplus prediction capability is more valuable than the surplus disposal capability—catching the problem before the food is cooked is better than discounting it after.

Winnow, a British startup now operating in over 2,000 commercial kitchens globally, deploys connected food scales that photograph and weigh discarded food in commercial kitchens, categorizing waste by dish and type through computer vision. The system feeds back to kitchen managers, showing precisely where waste is occurring at sufficient granularity to change behavior. Their reported client results consistently show 40-70% reduction in kitchen food waste within 12 months of implementation.

These numbers—40-70% reduction—are impressive but should be read with context. Kitchen food waste in large commercial operations was often occurring without any systematic tracking or measurement. When you first start measuring something and then manage to what you measure, large percentage improvements are normal. The absolute baseline matters. A kitchen wasting $3,000 of food per month and reducing that to $1,000 is a meaningful win; a kitchen that was already carefully managing waste and achieves a 10% improvement has a less dramatic headline number but may represent more fundamental change.

The food waste problem at the consumer level is where the highest-income-country waste accumulates but where AI has the least purchase. The behavioral economics of household food waste are clear enough: people over-purchase because they’re optimistic about their cooking plans, they’re bad at estimating quantities, and the psychological cost of throwing away food that they paid for is lower than the cost of a potentially wasted shopping trip.

Supermarket apps with AI-powered recipe suggestions calibrated to the user’s existing inventory—“you have half a head of cabbage and some chicken that expires tomorrow; here are three recipes”—address this at the margins. Meal kit services, which send precisely portioned ingredients for specific recipes, nearly eliminate this waste category for the meals covered. Consumer behavior apps that track fridge contents and remind users about expiring items have been tried and have thin adoption, because the cognitive overhead of maintaining the tracking exceeds the motivation for most users.

The honest assessment is that consumer food waste in high-income countries is substantially a product of cheap food. When food costs 10-15% of household income, as it does in the United States, the financial pain of discarding it is insufficient to change behavior at scale. In countries where food costs 40-50% of income, people waste much less at the consumer level—not because they’re more virtuous, but because the economics produce different behavior.

AI can improve the efficiency of logistics and enable better prediction up and down the supply chain. It can’t change the price signal that makes waste cheap. It can’t build the cold chain infrastructure that would prevent post-harvest loss in lower-income countries. It can’t change retailer buyer specifications that reject cosmetically imperfect food.

The 1.3 billion tons is not a technological problem in the main. It’s a combination of an infrastructure deficit in poorer countries, perverse incentives in rich-country retail, and a consumer psychology that is shaped by decades of agricultural policy designed to make food inexpensive.

AI tools can take meaningful bites out of specific parts of this waste flow—enough to save money, enough to reduce some of the environmental cost. But the full-system change required to cut food waste in half globally requires policy changes (ending cosmetic standards for retail, pricing food to reflect its environmental cost, investing in cold chain infrastructure in low-income regions) that AI cannot substitute for and that have to be won in political processes.

The number that puts this in context: if food waste were a country, it would be the third-largest emitter of greenhouse gases in the world, after the US and China. The food wasted globally every year required land, water, energy, and fertilizer to produce—and that resource investment was wasted along with the food.

Addressing it is not optional for a serious climate strategy. AI-driven efficiency improvements in the supply chain, retail, and food service sectors are the tractable near-term contribution. The deeper structural changes—in how food is priced, how it’s graded, how it’s distributed, and how it’s protected between harvest and consumption—are the durable ones.

One third of everything we grow doesn’t get eaten. That’s not because we lack algorithms. It’s because we’ve built a food system that made waste the economically rational choice at multiple points along the chain. Changing that requires changing the chain itself, not just adding more intelligence to manage the waste it generates.