Digital agricultural extension effectiveness (RCT evidence)

Digital Agricultural Extension Effectiveness: RCT Evidence

Digital agricultural extension works, but not in the way most programme plans assume. Across the randomised trials, advice sent by text, voice or app reliably shifts which inputs farmers buy — by roughly 20 percent in relative terms. Average yield gains sit near 4 percent, and income effects are often too small to detect. The effectiveness of digital agricultural extension depends less on how much information a service sends than on whether the recommendation changes from farm to farm.

Precision in the message is not the same as precision in the recommendation.

How effective is digital agricultural extension?

Digital agricultural extension changes input decisions by a small but consistent amount. The largest single test is the 2025 paper by Fabregas, Kremer, Lowes, On and Zane in the American Economic Journal: Applied Economics, which pooled six text-message trials covering more than 128,000 farmers in Kenya and Rwanda. Farmers who received messages were more likely to follow the input recommendation, with an odds ratio of 1.22 (95 percent confidence interval 1.16 to 1.29).

In absolute terms that is a 2 percentage point increase, measured from real purchases — coupon redemptions at agrodealers, order records from One Acre Fund — rather than survey answers, which overstated the effect by about 19 percent.

The earlier review by Fabregas, Kremer and Schilbach, published in Science in 2019, put the pooled figures at a 4 percent yield gain and a 22 percent rise in the odds of adopting a recommended input. Those numbers remain the most quoted in the field. They are closer to the ceiling than the floor.

Adoption moves. Yield moves less. Income moves least.

The three outcomes separate cleanly, and in the same direction in nearly every trial.

Cole and Fernando's 2021 study in the Economic Journal followed 1,200 cotton farmers in Gujarat, India, for two years on Avaaj Otalo — a voice service combining weekly push calls with a hotline answered by agronomists. Farmers changed where they got information and which inputs they bought. The published paper reports no systematic gains in yield or profitability, and puts willingness to pay below the per-farmer operating cost at study scale.

A larger, more recent trial points the same way. The 2025 evaluation of Ama Krushi in Odisha, India, by Cole, Goldberg, Harigaya and Zhu with Precision Development, enrolled 13,675 rice farmers. Across two seasons, treated farmers harvested 4.1 percent more rice and recorded 1.7 percent higher yield per hectare. Profit effects were positive but not statistically significant.

For a programme leader: expect adoption to move, expect yield to move by low single digits, and do not build the business case on measured income.

Precision in the message is not precision in the recommendation

Programmes usually plan to improve weak results by making messages more specific. The trials say that does not work.

Two of the East African experiments randomised farmers between a general message and a detailed one naming the local soil acidity level, the recommended lime quantity, its cost and the expected yield gain. The detailed version raised knowledge. It did not raise purchases. A third project added a phone call from a field officer on top of the texts and found no additional effect; only 13 percent of farmers offered a call requested one, and fewer than 1 percent of One Acre Fund clients ever used the free hotline. Repetition helped slightly. Framing and detail did not.

Now contrast the 2021 trial by Arouna, Michler, Yergo and Saito in the American Journal of Agricultural Economics. AfricaRice's RiceAdvice tool, used by extension agents to build a field-specific nutrient plan in Nigeria, raised rice yields by 7 percent and profits by 10 percent — without increasing total fertiliser use. The advice was not more detailed. It was different, farm by farm.

The cost of context has to be smaller than the decision it changes

A worked example from the Kenyan lime trials. Applying 10 kg of agricultural lime raises maize yield by about 10.3 kg, which the researchers valued at roughly USD 2.10 in profit after application, harvest and transport costs. An individual soil test in the same districts cost USD 15 to 20.

That test needs about seven seasons of the gain it refines to pay for itself, before the farmer buys any lime. On One Acre Fund's own trial plots, lime raised maize yields by a similar amount whether the soil tested above or below pH 5.5 — the threshold the test exists to resolve. Area-level soil data, already collected and effectively free per farmer, was close enough.

Per-farm precision is worth buying when it changes a decision worth more than the measurement costs.

What does digital extension cost per farmer compared with a field visit?

A three-message lime programme cost between USD 0.003 and USD 0.03 per farmer per season. On the most conservative assumption, inducing one farmer to try lime cost about USD 1.50. Farmer field days run by the Kenya Agriculture and Livestock Research Organization cost about USD 9 per farmer who attended, and USD 38 to 46 per farmer induced to experiment — even after assigning only one fifth of the event's cost to the lime message. Per 10 kg of lime actually bought, the text programme cost USD 0.25 and the field day USD 2.80.

Scale changes the arithmetic again. Fabregas and colleagues estimated a marginal benefit-cost ratio of up to 46 to 1 at bulk messaging prices, against 8 to 1 at medium scale. Ama Krushi, at nearly 7 million farmers by the end of 2023, carries an estimated long-run ratio of 12 to 1 through 19 to 1. Below a few hundred thousand users, fixed costs dominate and none of these ratios hold.

Digital advice adds to agent time rather than replacing it

The substitution question has a clearer answer than the yield question.

Fernando's 2021 follow-up on the Gujarat experiment found the service raised the returns to farmer-to-farmer conversation rather than displacing it. The 2024 trial by Baul, Karlan, Toyama and Vasilaky in the Journal of Development Economics tested Digital Green videos across 280 villages in Bihar, India, as a supplement to human extension sessions; effects on output were positive at the median, and largest when the video explained the extra labour a practice required. And in Nigeria, RiceAdvice raised yields because agents used it to collect field conditions. Remove the agent and the tool has no input.

The extension agent to farmer ratio explains why this matters. Abate and colleagues' 2015 paper in Food Security recorded one extension worker per 476 farmers in Ethiopia, the densest system in Africa, against 1:2,500 in Tanzania. In the United States, Wang's 2014 analysis in Choices found Cooperative Extension staffing fell from 17,009 full-time equivalents in 1980 to 13,294 in 2010, while a 2026 study in Agricultural and Resource Economics Review found the federal funding share fell from 42.4 percent in 1973 to 17.7 percent in 2024.

No digital service closes that gap by answering more questions. It closes it by making each remaining hour of agent time count for more.

Where the returns actually concentrate

Average effects hide both the value and its limits. In Odisha, farmers in districts hit by excess rainfall harvested 9.4 percent more than controls, and those in low-rainfall districts 11.6 percent more. Access cut the chance of losing more than half a rice crop by 10 percent overall, and by 26 percent for pest and disease losses. Where rivers flooded, the service showed no effect at all.

Digital advisory earns its return in the seasons and on the farms where standard guidance is wrong. That is an argument for systems that know the field, not systems that send more messages. Valora Earth is built on that premise: plot location, soil, crop stage and local weather are assembled into the farm's context first, so the recommendation changes with the farm rather than the wording around it.

Frequently asked questions

Does digital extension work better than in-person extension?

It is cheaper per farmer, not more effective per contact. Kenyan trials put text messaging at about USD 1.50 per farmer induced to try a new input, against USD 38 to 46 for farmer field days. Effect sizes per farmer are similar or smaller. The advantage is reach and cost, not depth.

What is a realistic extension agent to farmer ratio?

Ethiopia runs roughly one agent per 476 farmers, the densest public system in Africa. Kenya is near 1:1,000, Malawi 1:1,603, Tanzania 1:2,500 and India close to 1:5,000. United States Cooperative Extension staffing fell from 17,009 full-time equivalents in 1980 to 13,294 in 2010 as federal funding share declined.

How should a cooperative measure whether a digital advisory tool is working?

Measure input purchases from administrative records, not survey responses — surveys overstated effects by about 19 percent in the East African trials. Then segment by season conditions. Returns concentrate among farmers facing weather shocks or pest pressure, so a flat average across a normal year will understate value.