By Najeeb G. Abdulhamid¹, Elizabeth “Liz” Ankrah¹, Abiodun Ogunyemi², Mark Perry³, Merja Lina M. Bauters², Jona Repishti⁴, Steven Sam³, Samuel Chege Maina¹, Millicent Ochieng¹, Mercy Muchai¹, Stephanie Nyairo¹, Rikin Gandhi⁴, Jacki O’Neill¹ (see author affiliations)
Reported extension worker to farmer ratios in Africa often reach the thousands, including estimates as high as 1 to 10,000 in Nigeria [1]. This gap is more than a staffing shortage because it represents a systemic information crisis. Farmers need timely, contextual guidance on pests, inputs, and climate volatility, and AI is increasingly positioned as a way to extend advisory capacity at scale. Yet our core conclusion from the AfriCHI 2025 workshop in Cairo is clear. Agricultural AI will only deliver value if it is designed and evaluated for field utility rather than just model accuracy [2].
Field utility functions as a practical test of whether a system produces guidance that is usable, trustworthy, and actionable under real conditions of farming, including constrained infrastructure, mediated access, and linguistic diversity. When we optimize primarily for benchmark accuracy, we risk scaling systems that look strong in the lab but fail at the critical moment of decision in the field.
The Epistemic Gap and When AI Contradicts Reality
A recurring failure in agricultural AI is not that a system is wrong in the abstract, but that its logic conflicts with how farmers actually diagnose and decide. We refer to this as epistemic alignment, a concept that measures whether a system’s assumptions match the diagnostic frameworks of the people it serves.
A case discussed at the AfriCHI workshop makes this concrete. Conventional computer vision pipelines are often trained to recognize pests from high quality images of the insect itself. In practice, many subsistence farmers rarely see the insect directly. They diagnose through indirect cues such as frass, which is insect waste, and specific leaf damage patterns [2]. When an AI system cannot recognize these cues, it does more than miss a classification. It contradicts lived expertise. That contradiction can undermine trust and reduce willingness to reuse the tool regardless of its laboratory precision.
The design implication is not simply to collect more data in a generic sense. It is to represent the right evidence, including the local cues farmers already trust, and to build feedback mechanisms that allow farmers and intermediaries to correct system assumptions over time.
The Myth of the Individual User
Many tools are still designed around assumptions of a single person using a single device. In African agriculture, advisory is frequently a collective resource. In models such as the TGI Outgrower framework, information flows through lead farmers and cooperatives who relay questions, interpret recommendations, and coordinate action [2].
The effective user is often a human network rather than the device owner. For AI to work in these settings, systems must support mediated use, including shared device workflows, handoffs, and verification routines that reflect how advisory actually travels through a community.
Language as Infrastructure
Africa is home to high linguistic diversity that is commonly described as 1,000 to 2,000 or more languages, many of which are primarily oral [3]. In this context, text heavy interfaces can function as a form of digital exclusion. Agricultural AI therefore needs interactional intelligence, representing the ability to recognize ambiguity, request missing context, and guide users through multi-turn clarification.
Tools such as FarmerChat illustrate why multi-turn dialogue is a safety requirement [2]. A farmer’s initial statement that a crop is dying requires the system to probe for symptoms, crop stage, recent inputs, local conditions, and constraints before recommending action. In agricultural advising, one shot answers are not only low quality but can also be risky.
A Practical Agenda for the Next Phase of AgTech
We propose four sector shifts to guide future development.
- First, we must evaluate for feasibility. This means assessing systems on clarity and actionability, including whether advice is workable under labor, cost, and household constraints rather than focusing only on whether it is technically correct.
- Second, we must design for the collective. Intermediaries, shared device realities, and verification practices should be treated as first class design requirements rather than edge cases.
- Third, we must communicate uncertainty. As climate variability disrupts traditional heuristics, systems should support seasonal updating and make uncertainty visible. They must avoid overconfident recommendations when critical context is missing.
- Fourth, we must invest in shared public resources. We need a coordinated effort to create public agricultural corpora, including ethical multimodal datasets that reflect African linguistic and diagnostic realities.
Conclusion
The future of agricultural AI in Africa will not be decided by who has the most parameters. It will be decided by who can make advice travel safely into real decisions under real constraints while respecting the social systems through which agriculture actually functions.
Listening comes before speaking. In agricultural AI, that is not a slogan but a fundamental design requirement.
Author Note. The AfriCHI 2025 Workshop Consortium on Advancing Sustainable Agricultural Practices in Africa with AI includes Microsoft Research Africa, Digital Green, Brunel University London, and Tallinn University Estonia.
Authors:
Najeeb G. Abdulhamid¹, Elizabeth “Liz” Ankrah¹, Abiodun Ogunyemi², Mark Perry³, Merja Lina M. Bauters², Jona Repishti⁴, Steven Sam³, Samuel Chege Maina¹, Millicent Ochieng¹, Mercy Muchai¹, Stephanie Nyairo¹, Rikin Gandhi⁴, Jacki O’Neill¹
Affiliations:
¹Microsoft Research Africa, Nairobi, Kenya
²Tallinn University, Estonia
³Brunel University London, United Kingdom
⁴Digital Green
Further reading
- Full workshop report: https://doi.org/10.5281/zenodo.21391960 (opens in new tab)
- Strategic policy brief: https://doi.org/10.5281/zenodo.21394971 (opens in new tab)
- Technical white paper: https://doi.org/10.5281/zenodo.21395685 (opens in new tab)
References
[1] Oluoch M, Kitanaka M. Advancing agriculture extension models in Africa: bridging the gap for effective delivery of technologies and innovations. African Journal of Food, Agriculture, Nutrition and Development (AJFAND). 2024;24(3): Editorial. (See Introduction section statement on extension worker-to-farmer ratios). https://doi.org/10.18697/ajfand.128.ED138 (opens in new tab)
[2] Full Workshop Report on Outcomes and Strategic Insights from AfriCHI 2025 (opens in new tab)
[3] Harvard University, African Language Program. Introduction to African languages. n.d. https://alp.fas.harvard.edu/introduction-african-languages (opens in new tab)
Artificial intelligence