From John Deere's weed-targeting cameras to Merck's livestock wearables, the AI actually deployed on farms today is narrow and sensor-heavy — and it keeps running into the same wall: rural connectivity, capital cost, and fields that don't hold still for a model.
In 2021, John Deere put a camera-and-AI system called See & Spray into commercial corn and soybean fields. The pitch was narrow and concrete: instead of blanket-spraying an entire field with herbicide, cameras mounted on the boom identify individual weeds in real time and fire nozzles only at those spots. Deere's own field data puts the herbicide reduction commonly in the 50-70% range depending on crop, weed pressure, and field conditions — not because AI made farming smarter in some abstract sense, but because a narrow computer vision problem (weed vs. crop, at highway speed, in variable light) turned out to be tractable enough to ship.
That's the pattern worth paying attention to across agriculture: the AI that's actually in fields today is not a general "farm brain." It's a series of narrow, sensor-heavy systems bolted onto very specific, very expensive pieces of equipment, each solving one problem that used to require a human eye or a fixed schedule. Four areas account for almost all of the real deployment: precision spraying and weeding, crop and field monitoring, autonomous tractors, and livestock health monitoring. Each has shipped real product; each also runs into the same wall — rural connectivity, capital cost, and the fact that outdoor environments don't hold still for a model the way a warehouse or hospital does.
The computer vision problem here is genuinely well-suited to deep learning: distinguish crop plants from weeds, at speed, across lighting conditions and growth stages, using cameras mounted on a moving sprayer boom. Deere's See & Spray (built on the 2017 acquisition of Blue River Technology) runs this inference on embedded compute at the boom, triggering individual nozzles rather than spraying the whole width of the machine. The economic case is straightforward enough that adoption doesn't require anyone to believe in AI generally — herbicide is a real input cost, and cutting it by half while maintaining weed control pays for the hardware over a few seasons on large acreage.
Carbon Robotics took a different approach with its LaserWeeder: instead of targeted chemical spraying, it uses computer vision to aim lasers that thermally destroy weeds, eliminating herbicide for that pass entirely. This matters for organic operations and for growers facing herbicide-resistant weed populations, where the chemical toolkit is shrinking. The tradeoff is throughput — laser weeding is slower per acre than spraying, which limits it to higher-value row crops rather than commodity corn and soybeans at scale.
Both systems share a limitation worth naming honestly: they work well on well-defined row crops with visible contrast between crop and weed. Mixed cover crops, dense canopy, and non-row planting systems are harder problems that the current generation of products doesn't fully solve.
Above the plant level, platforms like Bayer's Climate FieldView and Taranis apply computer vision and ML to satellite imagery, drone imagery, and in some cases leaf-level aerial scans to flag disease, pest pressure, and yield variability before it's visible to a walking scout. Taranis's pitch specifically is resolution: its aerial imaging is fine-grained enough to identify early-stage pest damage or fungal infection at a scale a human scout would miss walking the field, and it does this across enough acreage that manual scouting simply isn't an option.
The value proposition is earlier intervention — catching a fungal outbreak or nutrient deficiency in the window where a single targeted pass fixes it, versus discovering it after yield loss is already locked in. The honest limitation is that these systems are diagnostic, not autonomous: they surface an alert, and a human agronomist still decides and executes the response. FieldView and similar platforms have also become a genuine point of friction around farm data ownership — who owns the yield and imagery data collected off a grower's own field, and what the platform vendor can do with it in aggregate, is an unresolved commercial and legal question that shapes adoption as much as the technology does.
John Deere's autonomous 8R tillage system, unveiled at CES 2022, is the most visible example: six pairs of stereo cameras for 360-degree obstacle detection, GPS-guided path planning, and no driver in the cab for tillage passes. The scope is deliberately narrow — tillage is a repetitive, well-bounded task on a field the operator has already surveyed, which is a much more tractable autonomy problem than, say, general road driving. Deere has expanded the approach incrementally rather than claiming full autonomous farming.
Monarch Tractor took a different market: a fully electric, autonomous-capable tractor aimed initially at vineyards and orchards, where rows are narrow, well-defined, and repetitive — again, a bounded environment that favors current autonomy tech over open-field chaos.
The adoption barrier for autonomous tractors isn't primarily the AI — it's capital and liability. These machines cost significantly more than their manually operated equivalents, and a farm's ROI calculation has to account for who's liable when an autonomous machine damages equipment, crops, or property. Deere has also been in a long-running right-to-repair dispute with farmers and the FTC over locked-down diagnostic software on its equipment, which is a reminder that the barrier to agricultural AI adoption is often as much about who controls the machine as what the machine can perceive.
The fourth area is less visible but genuinely deployed: wearable and environmental sensors paired with ML models to flag illness, estrus (heat cycles), and stress in cattle before visual symptoms appear. Merck Animal Health's SenseHub and Connecterra's Ida platform both fall in this category — collars or ear tags stream activity and rumination data, and a model flags deviations from an individual animal's baseline that correlate with early-stage illness or optimal breeding windows.
The economic case is again narrow and real: earlier illness detection reduces antibiotic use and losses, and better heat-cycle timing improves breeding efficiency — both are line items a dairy or beef operation can measure directly. The limitation is that these are anomaly-detection systems tuned to individual animal baselines; they don't diagnose a specific condition, they flag "this animal's behavior changed" and route to a human for follow-up.
| Application area | Maturity | Primary adopters | What still blocks wider rollout |
|---|---|---|---|
| Precision spraying/weeding | Commercial, scaling | Large row-crop operations | Upfront hardware cost; struggles on non-row/mixed canopy |
| Crop/field monitoring | Commercial, widely used | Mid-to-large operations with digital ag stack | Data ownership disputes; still requires human-in-the-loop response |
| Autonomous tractors | Early commercial, narrow tasks | Large acreage (tillage), vineyards/orchards | Capital cost, liability, right-to-repair/software lock-in disputes |
| Livestock health monitoring | Commercial, growing | Dairy and beef operations | Sensor cost per animal; connectivity in remote pasture settings |
The connective tissue across all four is that agriculture's hardest constraint isn't model capability — it's the physical and economic environment the model has to operate in. Rural broadband and cellular coverage remain patchy in exactly the regions with the most acreage, which limits real-time cloud inference and pushes vendors toward edge compute on the machine itself (part of why Deere runs See & Spray inference on the boom rather than round-tripping to a server). Capital cost concentrates adoption among large operations that can amortize six-figure equipment purchases over thousands of acres, while smaller farms are largely priced out of anything beyond monitoring subscriptions. And because outdoor fields don't offer the controlled lighting, fixed camera angles, or bounded object sets of a warehouse or factory floor, every vision system here has had to be engineered around a narrower task than "understand the field" — weed vs. crop, healthy vs. diseased leaf, normal vs. anomalous gait — rather than solved as a general perception problem.
None of this is a story about AI failing in agriculture. It's a story about AI succeeding exactly where the problem was scoped tightly enough to ship: one sensor, one narrow classification, one machine, one measurable input cost. The operations getting real value today aren't the ones waiting for a general farm intelligence — they're the ones that found the specific, bounded task where a model beats a human eye or a fixed schedule, and built the hardware to act on it in the field, not in the cloud.