We devote entire articles to models that gain three points on a coding benchmark. Meanwhile, agriculture applies these technologies across vast areas, with measurable effects on yields, inputs, and the environment. It's arguably one of the sectors where the real impact is greatest, and one of the least talked about.
The contrast is striking. Agriculture ticks every box for what AI does well: visual recognition, prediction from historical data, optimisation under constraints. And it does so with an immediately quantifiable benefit.
What already works
Disease and pest detection. Photographing a leaf and getting a diagnosis is an image recognition problem, exactly what these systems excel at. The point isn't to replace the agronomist but to detect early, over large areas, what a human eye can't inspect everywhere.
Targeted weeding. This is arguably the most spectacular application. Systems mounted on machinery identify a weed in real time amid a crop and treat only that precise spot. The herbicide reduction reported for such devices is often in the tens of percent. The benefit is both economic and environmental, which is rare.
Within-field rate modulation. Adjusting fertiliser or water inputs according to the actual needs of each zone in a field, based on satellite or drone imagery. This moves away from the uniform treatment approach.
Forecasting. Yields, optimal intervention dates, weather risks. These are time-series problems where machine learning has been effective for a long time.
Livestock. Early detection of lameness or drops in feed intake through video analysis, allowing intervention before the animal is visibly sick.
Three reasons. The benefit is immediately quantifiable: fewer inputs, higher yields, it shows up on the farm's accounts. The risk of error is recoverable: a bad diagnosis on one field can be corrected, unlike a medical error. And the sector is already equipped with sensors, with decades of agronomic data. These are exactly the conditions we described in our article on jobs for automation to take hold quickly.
The limits, which are real
Access. These tools require hardware, a connection, and often a subscription. They benefit first the farms that can invest, which may widen the gap between large and small operations. This ties into the adoption divide we covered for businesses.
Data dependence. A model trained on one climate, soil type, and set of varieties transfers poorly elsewhere. It's the same gap between lab and field we noted in medicine. A system that performs well in the Beauce region won't necessarily do so in southern Tunisia.
The question of data ownership. This is the most sensitive and least discussed point. Yield, practice, and soil data constitute a considerable asset. When they flow back to an equipment or input supplier, the farmer can find themselves in a weak position in a negotiation where their counterpart knows more about their own operation than they do.
What this says more broadly
This sector illustrates something that media coverage of AI systematically misses: the most useful applications are almost never the most visible.
A system that halves herbicide use across thousands of hectares produces a greater environmental and health effect than much of what we talk about every week. It makes no headlines, because there's no conversation, no spectacular feat, no philosophical debate.
It's a useful reminder for evaluating this technology: the question isn't what it can do that's most impressive, but where it solves a problem someone actually had. On that score, a camera that tells a weed from a sugar beet plant is probably worth more than many an announcement.