Digital Data-Driven Extension: AI Chat Assistants for Smallholder Farmers

Thesis Abstract

The proliferation of information and communication technologies (ICTs) has not fully translated into accessible, timely, and actionable agricultural guidance for smallholder farmers in many regions, resulting in suboptimal adoption of improved practices and limited productivity gains. This study investigates how AI-driven chat assistants can deliver data-informed extension services that are contextually relevant, timely, and scalable for smallholders, addressing persistent gaps in knowledge transfer, advisory reach, and user engagement. The aim is to evaluate the effectiveness of AI chat assistants as a digital data-driven extension channel and to elucidate the mechanisms through which such tools influence decision-making, adoption of improved practices, and yield outcomes. Specific objectives are (i) to design a domain-specific AI chat assistant integrated with a weather, pest, and market intelligence backend; (ii) to assess user acceptance, trust, and perceived usefulness among smallholder farmers and extension agents; (iii) to quantify the impact of AI-assisted advisory on farmer knowledge scores, technology adoption rates, and yield; (iv) to identify contextual determinants (plot size, literacy, access to mobile devices, and network quality) that modulate tool effectiveness; and (v) to develop a practical framework for scalable deployment and governance of AI-powered extension services. The study adopts a mixed-methods research design comprising a quasi-experimental field trial and an embedded qualitative inquiry. The population includes smallholder farmers across three agro-ecological zones within a defined rural district, with a target sample of 600 households for the quantitative strand and 36 in-depth interviews for the qualitative strand. A stratified random sampling approach ensures representation by farm size, crop system, and gender. Data collection employs (a) pre- and post-intervention surveys to measure knowledge, adoption, and yields; (b) system log analytics to capture user interactions, query types, response times, and engagement depth; (c) structured observation of advisory sessions; and (d) semi-structured interviews and focus group discussions to explore perceived value, trust, and barriers. Instruments include a validated knowledge assessment tool for agronomic practices, an adoption checklist aligned with regional extension messages, and a user experience questionnaire anchored to the Technology Acceptance Model (TAM) with extensions for trust and usefulness. Validity and reliability are ensured through pilot testing of the AI chat assistant, triangulation across multiple data sources, and Cronbach’s alpha checks for survey scales. Data analysis proceeds with (i) descriptive statistics and reliability analysis; (ii) difference-in-differences estimation using a regression framework to identify causal impacts on knowledge scores, practice adoption, and yield, controlling for baseline covariates; (iii) multivariate regression to examine determinants of tool usage and adoption intensity; (iv) ANOVA to compare outcomes across zones and farm sizes; (v) thematic analysis of interview and focus group data to illuminate perceived benefits, trust, and barriers; and (vi) structural equation modeling to test the hypothesized pathways from tool engagement to outcomes, mediated by perceived usefulness and trust. The study further employs a conceptual framework integrating Technology Acceptance Theory with Diffusion of Innovations and a data-driven feedback loop for adaptive advisory content. Anticipated findings include (a) significant improvements in knowledge scores and adoption rates for climate-smart practices among users of the AI chat assistant compared with non-users; (b) modest but meaningful yield gains in participating plots, with larger effects in medium-sized farms and zones with reliable connectivity; (c) positive relations between engagement depth, perceived usefulness, and trust with adoption intensity; and (d) context-specific moderators such as literacy and device access that shape effectiveness. The research contributes to knowledge by integrating AI-enabled natural language processing with extension theory, offering empirical evidence on the viability and constraints of digital data-driven extension in smallholder contexts, and presenting a deployment framework that addresses governance, data privacy, inclusivity, and scale. Practical recommendations include investing in offline-enabled AI capabilities for low-connectivity areas, co-design with farmer groups to ensure relevance and trust, establishing governance and data ethics protocols, and developing a phased rollout with continuous performance monitoring. The study concludes that AI chat assistants, when embedded in a data-informed extension architecture and supported by local capacity building, can enhance advisory reach, accelerate knowledge transfer, and improve productive outcomes for smallholder farmers.