Proven, promising, promised.

01How we grade.

• Proven at scale: used on millions of acres or animals, with measured results, at least some of them from university or extension researchers rather than the company selling it.
• Promising: works in trials or in specific niches; independent evidence is limited or mixed.
• Promised: mostly vendor claims, pilots, or research prototypes.

The grades follow the evidence ladder: a company's own figures count, but they cannot carry a grade alone. Two rows rest on published figures: spot spraying and GPS guidance. The other eight grades are our reading of what we could find (company material, pilots, research reviews), and each one says where the evidence is thin. This map is a snapshot. We re-grade it every October, and the date at the top of the page says when it was last done.

02The map at a glance.

03Proven at scale.

Camera-guided spot spraying.

A sprayer with cameras along the boom looks at every patch of ground, a model decides whether it is looking at a weed, and only the nozzles over weeds open. John Deere's See & Spray is the best-known example. Deere reports the system covered more than five million acres in the 2025 season and cut non-residual herbicide use by an average of nearly 50%, about 31 million gallons of mix. Those are the company's figures, and the release does not publish its method or baseline. Deere's yield claims (an average of 2 bushels an acre in soybeans, up to 4.8) come from trials it sponsored.

The "proven" grade rests on two university field studies. In 2024, Iowa State University engineers ran a See & Spray machine over five soybean fields, 415 acres in all, and measured herbicide savings from 44% to 91% depending on weed pressure, 76% on average. A three-year University of Arkansas soybean trial found 43% to 59% less herbicide than broadcast spraying, and warned that the least sensitive setting let the weed seedbank grow quickly. Neither report says who paid for the work, but both teams measured the savings themselves, on working fields.

GPS guidance and auto-steer.

Rarely called AI, but it is the foundation every other field machine builds on. USDA found 70% of large U.S. crop farms used auto-steer in 2023. Its data is position: where the machine is, to within an inch or two, and where it has already been. See the numbers for how far precision agriculture has spread, and to whom.

04Promising.

Image-based pest, disease, and plant identification.

Phone apps and camera systems identify common weeds, pests, and diseases well. They are weaker on local varieties, unusual conditions, and anything they were not shown during training: a model trained mostly on one region's crops can stumble on another's. Researchers who reviewed deep learning in agriculture name the cost of collecting and labeling training images as a central bottleneck (Li and colleagues, Computers and Electronics in Agriculture, 2023).

Evidence behind this grade: research reviews and wide everyday use; we found no large independent study of accuracy on working farms.

Livestock monitoring.

Cameras, collars, and ear tags that flag a lame cow or an animal off its feed are widely sold. Published evidence of better outcomes differs from product to product. The data underneath is a continuous record per animal, plus health events a person has labeled.

Evidence behind this grade: our judgment from product material and pilots; we found no independent study of outcomes across many farms.

Yield forecasting.

It works where long histories exist and is useless without them. A forecast is only as good as the seasons of yield, weather, and imagery it learned from.

Evidence behind this grade: our judgment; results depend so much on local history that no single study settles it.

Greenhouse climate control by AI.

The research base is strong. In Wageningen University's Autonomous Greenhouse Challenge, teams let AI run a greenhouse crop against expert growers, and the AI-managed crops matched or beat them under the trial's conditions (see the future). Commercial results so far come mostly from large, heavily instrumented operations, because that is where the dense, calibrated data already exists.

Evidence behind this grade: a research competition plus large-grower reports; results have not been reproduced widely at commercial scale.

AI advice by chatbot.

This area is moving fast and has little outcome evidence so far. Digital Green, a nonprofit that serves smallholder farmers, tested advisory models against facts written by agricultural experts, using crops grown in Bihar, India, in a February 2026 preprint (not yet peer-reviewed). It is one study, but a careful one. Its best system scored about 57% on that test (a measure of how closely its answers matched the experts' facts), and it was trained on a larger mixed set of about 130,000 examples, some of them machine-generated. Trained on the roughly 12,000 expert-checked examples alone, it scored about 52%. General-purpose models with no farm training scored between 22% and 48%. The weakest topics were the ones with the least training data. Whether this kind of advice actually improves yields or income has not been shown in a rigorous trial yet. The first one, in Kenya, runs to March 2027 (details).

AI for business planning.

Budgets, scenarios, loan packages, and strategy. Asked in August 2026 which of their skills AI could improve most, U.S. producers picked strategic planning more than anything else (the numbers). It runs on the farm's own financial and operating records.

Evidence behind this grade: what producers expect, from the Purdue survey; we found no study of results.

05Promised.

Autonomous field robots.

Weeding, harvesting, and scouting robots are real, and some are on farms. They are expensive, mostly aimed at high-value specialty crops, and many "products" are still pilots or research machines. Some are edging toward promising.

Evidence behind this grade: our judgment from product material and pilots.

"AI runs the farm."

We found no published evidence of a fully autonomous farm manager running a commercial farm.

06What the map shows.

Read down the third column. Both proven rows sit on purpose-built data systems: millions of labeled images and a camera on every section of boom in one, satellite positioning in the other. In this map, the grade follows the data.