The First Randomized Test of an AI Farm Advisor Found It Changed How Farmers Fight Pests, Not How Much They Harvested

Plenty of companies now promise that an AI chatbot can put an agronomist in every farmer’s pocket. Until this month, there was surprisingly little hard evidence about what happens when farmers actually get one.

A new paper posted to the research site SSRN in early October offers some of the first. Economists Ali Aouad, Alp Süngü, Niyazi Taneri and Canberk Ucel ran a pre-registered randomized controlled trial with 1,935 smallholder rice farmers in the Philippines and India over one growing season. In a LinkedIn post announcing the paper, Ucel described it as, to the authors’ knowledge, the first randomized evaluation of AI advisory for smallholder farmers.

The short version of the result is encouraging and sobering at the same time. The AI advisor changed how farmers managed their fields, especially when it came to pests, weeds and disease. It did not change their yields or costs within that season.

How the experiment worked

Farmers in the treatment group got access to an AI agricultural advisor through WhatsApp or Facebook Messenger. According to the paper’s abstract, the tool was developed with local experts and promoted sustainable rice-growing practices. In his own post about the study, Süngü said the team built the chatbot, which they call AgriLLM, over years of iteration and in partnership with leading agricultural firms.

The control group is what makes the design unusually clean. Instead of getting nothing, those farmers received a similar conversational AI chatbot that simply did not discuss agriculture. That way, any difference between the groups could be traced to the farming advice itself, not to the novelty of chatting with an AI.

The researchers tracked how farmers used the advisor over the season, then measured outcomes with a detailed end-of-season survey. Ucel wrote that the quality of farming practices was measured with a new instrument inspired by the World Management Survey, a tool originally designed to grade how well businesses are run.

What changed in the fields

According to the abstract, deliberate changes to farming practices rose by about a fifth, measured as an effect of 0.10 standard deviations. The quality of farming operations improved by 0.08 standard deviations. Those are modest numbers, but in a single season and a large randomized sample, they are real shifts.

The gains clustered in one area. Farmers with the AI advisor reported moving toward more integrated management of pests, weeds and diseases, an effect of 0.11 standard deviations, and were more likely to adopt new practices in those areas. Crop protection and problem response were also the topics farmers raised most often in their conversation logs.

That pattern says something about how people use these tools. Ucel noted that farmers tended to consult the advisor when an immediate problem showed up in the field rather than proactively, and the improvements were concentrated in exactly those reactive, problem-solving practices.

One result came from a less direct route. Farmers with the advisor were 7 percentage points more likely to report avoiding crop-residue burning, the practice of torching leftover straw after harvest. Ucel wrote that the advisor tended to bring up alternatives for handling residue during related conversations, rather than because farmers asked, and may have reinforced existing government and extension efforts to discourage burning.

What did not change

The study found no statistically significant effect on yields or costs in that growing cycle, and no effect on the adoption of technologies or inputs other than pesticides. Other sustainable practices the advisor promoted did not take hold in a measurable way.

Süngü gave a concrete example. The chatbot increased farmers’ interest in direct-seeded rice, a method promoted as more sustainable, but that interest did not turn into adoption, possibly because the method requires machinery or other resources. “This doesn’t mean AI can’t improve farmers’ productivity or support sustainable farming,” he wrote, “but it suggests information alone may not be enough.”

The authors draw a similar line in the abstract. AI advice, they conclude, can improve the management decisions farmers can act on with resources they already have, particularly when a problem is time-sensitive. Changes that depend on money, equipment or other inputs are a different matter. Ucel added that the team did not find evidence that low-resource or disadvantaged farmers engaged with or benefited from the tool differently than others.

How it compares with older phone-based advice

Digital farm advice is not new. A 2025 working paper from the nonprofit Precision Development evaluated Ama Krushi, a free voice-based service in the Indian state of Odisha that sends regular audio messages and offers an automated hotline. In that randomized rollout among 13,675 rice farmers, access improved knowledge and adoption of recommended practices, produced a modest increase in rice yield and harvest, and sharply reduced the likelihood of crop loss on average. The benefits were concentrated in areas hit by certain weather shocks, where harvests rose by up to 9 percent.

The new study tests something different. A voice service pushes advice on a schedule. A chatbot waits for the farmer to ask. Ucel’s post points to exactly that distinction: because an AI advisor operates on demand, farmers used it when trouble arrived, and that is where it helped most.

What it means for home gardeners

Nobody in this study was growing tomatoes in a San Tan Valley backyard, and rice paddies in the Philippines and India are a long way from a raised bed in the Sonoran Desert. Still, many gardeners now type plant problems into a chatbot, and the findings line up with some sensible habits.

Use AI where timing matters. The clearest gains in the trial came from fast, practical responses to pests, weeds and disease. If aphids show up on your kale or a squash leaf starts to wilt, describing the symptoms in detail and asking for possible causes and a step-by-step response is a reasonable use of the tool.

Ask about the whole system, not just a spray. Farmers in the trial shifted toward integrated management, which combines monitoring, cultural fixes, physical controls and targeted treatments instead of reaching for one product. Asking a chatbot what to try before and alongside a pesticide tends to produce more useful answers.

Expect advice, not resources. The study suggests information alone did not get farmers to buy new equipment or overhaul how they plant. The same is true at home. A chatbot can suggest drip irrigation or a shade cloth, but it cannot install either, so pick the suggestions you can actually carry out this week.

Check it against local expertise. The advisor in this study was built with local experts and tuned for specific crops and regions. General-purpose chatbots are not. In Arizona, the University of Arizona Cooperative Extension publishes free guides written for low-desert conditions, and comparing an AI answer with one of those guides is a quick way to catch advice meant for a wetter, cooler place.

A careful first step

The trial covered one season, two countries and one crop, and its authors are careful not to overclaim. What it adds is something the field has been missing: a clean test showing that well-built AI advice can change what growers do when a problem appears, even if bigger harvests take more than a good answer.

For a technology that often arrives wrapped in sweeping promises, that is a useful, grounded result. It suggests the best farm and garden chatbots will be judged less by how much they know and more by whether they help someone make the right move on the day the bugs arrive.