Generative AI and Chatbot-Based Advisory Systems in Agricultural Extension: A Comprehensive Review with Insights from Rajasthan, India

Abstract

Agricultural extension in most developing countries suffers from a persistent ‘last-mile’ problem: too few trained personnel, uneven coverage of remote and arid regions, and advice that rarely matches the specific agro-climatic conditions of an individual farm. Generative artificial intelligence (GenAI), and in particular large language model (LLM)-based conversational chatbots, has recently been put forward as a scalable complement to human extension networks. Drawing on a structured narrative synthesis of 30 peer-reviewed, pre-print, and institutional sources identified through systematic database and grey-literature searches, this review synthesises the fast-growing literature on GenAI and chatbot-based advisory in agriculture, tracing its evolution from rule-based expert systems and SMS/IVR advisory to retrieval-augmented, multilingual, voice-first conversational agents. It examines the enabling technologies — large language models, retrieval-augmented generation, and multimodal, multilingual natural language processing — and reviews global and Indian deployments, including Digital Green’s Farmer.Chat, IFPRI’s Generative AI for Agriculture (GAIA) initiative, and India’s newly launched Bharat-VISTAAR platform, inaugurated in Jaipur, Rajasthan, in February 2026. Evidence on farmer trust, adoption, and socio-technical barriers is critically evaluated, with particular attention to the semi-arid, drought- and locust-prone agro-ecology of Rajasthan and to unresolved challenges around low-resource Indic and dialectal language coverage, data sovereignty, and human-oversight requirements. Concretely, the reviewed evidence shows automatic-speech-recognition error rates for Indian languages climbing sharply outside Hindi, multilingual large language models measurably more prone to factual error in low-resource Indic languages than in English, and a 2026 producer survey in which just under half of respondents report weekly use of general-purpose AI tools while only about a quarter say they trust its agricultural recommendations without independently cross-checking them; a large-scale randomised evaluation of non-GenAI digital advisory in Odisha, by contrast, found only modest average yield gains that were markedly larger for farmers facing weather shocks — a reminder that platform scale and demonstrated impact do not automatically track one another. The review identifies persistent gaps — thin longitudinal impact evidence, negligible coverage of regional dialects such as Marwari and Mewari, unresolved liability questions for erroneous advice, and weak integration with local Krishi Vigyan Kendra (KVK) networks — and sets out a research and implementation agenda for context-specific, human-in-the-loop GenAI advisory suited to arid and semi-arid smallholder agriculture.