Improving smallholder agriculture with generative AI: The impact of Farmer.Chat in Kenya

How to cite

Abate, Gashaw T; Boukaka, Sedi; Deutschmann, Joshua; Bernard, Tanguy; Spielman, David. 2025. "Improving smallholder agriculture with generative AI: The impact of Farmer.Chat in Kenya." AEA RCT Registry. AEARCTR-0017035. https://doi.org/10.1257/rct.17035-1.2

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Gashaw T Abate* , Sedi Boukaka , Joshua Deutschmann , Tanguy Bernard , David Spielman

* Primary investigator

Registry ID
AEARCTR-0017035
DOI
https://doi.org/10.1257/rct.17035-1.2
First registered
Last updated
Status
On going
Trial dates
2025-09-01 – 2027-03-31
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/17035

Abstract

This study evaluates the impact of an AI-powered agricultural advisory chatbot, Farmer.Chat, developed by Digital Green to enhance knowledge, technology adoption, productivity, and income among smallholder farmers in Kenya. Traditional agricultural extension systems in many low- and middle-income countries face persistent capacity and reach constraints, leaving farmers with limited access to timely, tailored information. Digital advisory tools have begun to bridge this gap, yet rigorous evidence on the effectiveness of AI-enabled solutions remains limited. We will implement a cluster-randomized controlled trial (RCT) in 600 villages in Nakuru County, Kenya, to assess the causal effects of access to Farmer.Chat on farmers’ knowledge, uptake of recommended agronomic practices, yields, and household income. The experiment will also include a cross-randomized video intervention designed to stimulate learning and behavioral change. Within villages treated with FarmerChat onboarding, we embed two additional randomized, exploratory treatments: an expanded onboarding for other farmers in the village, and an experience sharing/feedback session with Digital Green. The trial will be conducted across two agricultural seasons. Findings will provide some of the first causal evidence on the role of generative AI in agricultural extension, informing future efforts to integrate AI tools into scalable, farmer-centered advisory services.

Intervention

There are two cross-randomized interventions:

(1) Farmers in the Farmer.Chat treatment will have the application installed on their phone.
(2) Farmers in the video treatment will be shown videos to increase aspirations and spark thinking of new ways of doing things, possibly supplemented with text messages and/or phone calls.

Within villages randomized to receive the FarmerChat treatment, we randomize and embed two additional exploratory treatments, alongside a group with no additional interventions. These interventions will be conducted several months after the sampled farmers were onboarded.
(a) Expanded onboarding: Digital Green will conduct expanded village-level onboarding sessions to FarmerChat, in line with their usual in-person expansion strategy
(b) Experience sharing and feedback: we will organize sessions for sample farmers and other FarmerChat users in the village to share their experiences with each other, and provide feedback to Digital Green

Experimental design

We use a cluster randomized controlled trial (RCT) with a 2x2 factorial design, assigning treatments at the village level to minimize potential spillover effects. The first intervention, Farmer.Chat, involves guiding farm households in treatment villages to download the app and giving some advice to craft queries. The second intervention is a video nudge: a subset of both treatment and control villages will be randomly assigned to receive a short video featuring a local farmer, aimed at lowering cognitive barriers, increasing aspirations, and promoting innovative thinking. Pending additional budget, we may reinforce the video nudge treatment with text messages and/or phone calls.

Within villages randomized to receive the FarmerChat treatment, we layer two additional embedded, exploratory treatments, alongside a group with no additional interventions: expanded onboarding, and experience sharing/feedback sessions. 100 villages are randomized to each of these two embedded treatments, with an additional 100 villages held out to receive no additional intervention. These exploratory treatments will be conducted several months after the initial onboarding of sampled farmers in treatment villages.

Design details

Randomization method
Randomization was done in office by a computer.
Randomization unit
Village
Clusters
600 villages
Observations
3000
Arms
150 villages control, 150 villages Farmer.Chat, 150 villages video nudge, and 150 villages both treatments. Within the 300 villages randomized to receive FarmerChat, we randomize 100 to receive expanded onboarding, 100 to receive an experience sharing/feedback session, and 100 to receive no further intervention.
Primary outcomes
Access to and use of agricultural extension Knowledge and adoption of good agricultural practices Agency / decision-making capabilities
Secondary outcomes
Productivity and production decisions (e.g. agricultural portfolio choice) Income and consumption Food and nutritional security

Cited by

No citation data available for this registration yet.

Keywords