AgriLLM: How CGIAR is developing an AI-powered agricultural advisory service for global South

CGIAR, in partnership with the United Arab Emirates' AI71, has initiated the development of an AI agricultural advisory platform based on a large language model (LLM) tailored to agriculture, aptly named AgriLLM.

But this is not just another LLM. AgriLLM is an ambitious project aimed at equipping the agricultural community, including researchers, policymakers and smallholder farmers, with tailored AI tools and models. Unlike general-purpose AI tools, AgriLLM is rooted in scientific rigor, grounded in CGIAR knowledge and local realities, and tailored for people who rarely feature in the AI revolution: those who grow, raise, and fish for the food that feeds the world.

To kick-start this ambitious effort,CGIAR's Digital Transformation Accelerator(DTA) convened two workshops at ILRI, Nairobi that aimed to gather a foundational dataset of high-quality question and answer (Q&A) pairs that reflect the concerns, language, and context of key user personas - smallholder farmers seeking localized, practical advice; extension agents needing field-ready guidance; researchers looking for synthesis and analysis support; and policymakers requiring concise, data-driven insights.

The first workshop was held on 9 June, bringing together over 25 in-person participants from ILRI. The second workshop, held on 12 June 2025 brought together over 70 virtual and in-person participants from CGIAR centers: ILRI, theCenter for Agricultural Research in the Dry Areas(ICARDA), the International Crops Research Institute for the Semi-Arid Tropics (ICRISAT), the International Potato Center (CIP), the International Institute of Tropical Agriculture (IITA), the International Rice Research Institute (IRRI), AfricaRice, the International maize and Wheat Improvement Center (CIMMYT), and partners such as the World Vegetable Center among others. The participants ranged from scientists, data specialists, extension agents, economists, communications, to digital innovation experts.

Generating the Q&A pairs

Participants generated the Q&A pairs in a structured three-phase process. First, they generated broad agricultural topics relevant to various user personas. Then, they created realistic, needs-based questions from the perspectives of those personas, such as farmers, extension agents, and policy makers. Finally, teams collaborated to create clear, evidence-based answers for each question, including citations when possible.

The participants exceeded expectations, with the first workshop producing over 360 Q&A pairs and the second generating more than 500.

Jean-Baka Domelevo Entfellner, Head of Data and Research Methods Unit, ILRI, Lina Yassin, Product Lead, CGIAR, and Mahmoud Alaoui from AI71, and Ram Dhulipala, interim Director of the DTA facilitated the workshops. Together, they walked participants through the vision, architecture, and strategy of AgriLLM.

Dhulipala emphasized the goal of grounding the model in trusted, domain-specific knowledge as this ensured that AgriLLM is more relevant, accurate, responsive and better equipped to provide high-quality, context-specific responses for real-world agricultural challengesto the real-world needs of users such as smallholder farmers, extension workers, researchers, and policymakers especially in the global South.

Toward COP30 and beyond

These sessions are part of a wider effort to gather 500 Q&A pairs per CGIAR center. The ultimate goal? To launch the first version of AgriLLM's AI-powered assistant with region-aware, role-specific responses that can support farmers, advisors, and decision-makers alike. A working chatbot prototype is expected to be showcased at COP30, positioning CGIAR and its partners at the forefront of AI-powered agricultural transformation.

Next steps

Yassin outlined the planned follow-up activities from the workshops as follows:

• Collect additional Q&A pairs from CGIAR, and initiate Q&A pairs collection with FAO.
• Finalize post- processing and curation of the human-generated Q&A pairs; expand and diversify the training / test sets (e.g., by adding typos to the questions on purpose); and proceed to a second round of fine tuning.
• Deploy the fine- tuned model on the AI Assistant; start developing context-aware retrieval capabilities (location-specific, crop-specific); and develop onboarding flow to extract information about the user

It's still early, but the seeds have been sown. And if these workshops are any indication, AgriLLM could be one of CGIAR's most transformative tools - for that mother in Ghana planting cassava, for the extension agent advising on fall armyworm, for the policymaker designing drought insurance.