In the AI era, agricultural research needs better questions, not more tools

by Eliot Jones-Garcia, Senior Research Analyst, IFPRI, for CGIAR Digital Transformation Accelerator.

I’ve been to more AI conferences this past year than I can count.  While they all highlight important topics in the “AI – food systems space,” I have also begun to ponder whether agricultural research institutions might be at risk of mistaking AI activity for AI strategy.

Research organizations want to remain relevant in this fast-moving space, so many are building chatbots, platforms, dashboards, advisory tools, data infrastructures… The temptation is understandable. Donors are interested. Technology firms are entering the sector.  

Do public agricultural research organizations enjoy a strategic comparative advantage in building AI tools?  The answer is arguably no: digital products require software engineering, product management, user support, maintenance, cybersecurity, business models, and continuous iteration. If research institutions move too far in this direction, they risk drifting away from their public-good mandate, competing with local technology providers, and producing tools that are expensive to maintain, quickly outdated, and weakly adopted.  This is not to say that CGIAR should never build technology. It is that building should be the exception, not the reflex.

The harder question is not what we can build. It is what we should build, what others are better placed to build, and what kind of knowledge only research institutions can provide.

CGIAR and other agricultural research organizations’ comparative advantage lies elsewhere: they have valuable scientific knowledge, datasets, trusted relationships, and deep experience across farming systems. These are powerful assets in the AI era.  

A stronger role for agrifood systems researchers may be to build the knowledge infrastructure that makes better technology possible: high-quality science, standards, benchmarks, governance principles, and safeguards for responsible data use. These are areas where public research can inform and create value for entire innovation ecosystems, rather than launching one more project-based tool.

But even here, caution is advised.

Too many discussions about AI in agriculture quickly devolve into discussions about data. Who owns it? Who stores it? Who governs it? Who can access it? Who is represented? These are important questions, but they can also create the comforting impression that research institutions will remain central to the AI era simply because they hold data. That assumption avoids a more difficult question: do farmers, extension agents, and national agricultural organisations actually need what is being built?

A recent paper makes this tension visible. When asked who would benefit most from aggregated farm data, only 16% of farmers identified themselves as the primary beneficiaries. The finding suggests that many farmers perceive agricultural data as generating value primarily for commercial actors rather than for those who produce it – a figure that should stop us in our tracks.  

If farmers do not see the arrangements through which agricultural data are collected, controlled, and converted into value as serving their interests, then calling those arrangements human-centred does not make them so. It may be inclusive in language but exclusionary in practice, technically sophisticated but poorly aligned with how decisions are actually made, or designed to answer a question that no farmer or adviser is asking.

This is the problem with the “if we build it, they will come” logic that still shapes too much of the discussions about agricultural technology. It is not problem-driven. It starts from institutional resources and works outward, rather than starting from farmers’ decisions, constraints, and aspirations. Data may contribute to the solution, but it cannot substitute for understanding the problem.

Human-centred design (HCD) is often presented as the corrective to this tendency. Yet it can reproduce the same logic when users are brought in only after the problem, technology, and desired outcome have already been defined. HCD is a creative approach to technology design that prioritizes users’ needs, but in reality, it is often treated as an instrument for scaling the adoption of tools scientists already believe are correct. Users are consulted to refine the interface, validate the concept, or reduce friction. Their role is to help a predetermined innovation travel further.

If HCD is going to matter in the AI era, it needs to become more than a method for making existing tools easier to adopt. It needs to become a research agenda, examining what tools – if any – people need, what functions those tools should perform, and which features would make them useful within the realities of everyday decision-making.  

The appeal of building tools, platforms, and data infrastructures is clear: creating and launching new tools brings visibility. But these other “softer” questions for research organisations are where they can actually have the greatest impact in the fast-moving AI age.  

AI systems will only be useful in agriculture if they are built around real decision contexts. A farmer deciding whether to apply fertilizer, sell maize, plant early, seek credit, trust an advisory message, or share farm data cannot simply be seen as an end user of a particular tool. Farmers are operating within a social, economic, and institutional environment, which must be understood and incorporated into an AI tool, if it is to be useful. AI tools which ignore that environment, do so at their peril; they risk scaling the same old design failures faster.

Agricultural research institutions do not need to control last-mile delivery. They do not need to build a multitude of apps. They do not need to become platform companies. Instead, their most important contribution is to help the wider innovation ecosystem understand farmers better.

They can produce evidence on user needs across different farming systems; support participatory design processes that treat farmers and local organisations as co-creators rather than data sources; evaluate not only whether a tool was adopted, but whether its design was inclusive and accountable; and identify where AI adds value, where simpler tools are better, and where no technical solution is appropriate. Much of this can be done through familiar research methods: surveys, interviews, focus groups, observation, and participatory workshops.

This division of labor would also be conducive to a better relationship with the private sector. Technology firms and start-ups may be better placed to build and maintain products. Local SMEs may be better placed to adapt tools to local markets. Extension providers may be better placed to deliver advice. Farmer organizations may be better placed to define priorities. Public research can strengthen all of them by making the underlying knowledge base stronger and more inclusive.

Launching new platforms brings visibility. But ultimately, public value comes from making agricultural innovation, including AI tools, more useful to the people it claims to serve.

I do not see AI solving the relevance problem in agricultural research. It may even make it worse, by rewarding speed, scale, and technical novelty over a deeper understanding of users and their contexts.  

In the AI era, agricultural research should not start with technology. It should start with the question that too often comes last: who is this the technology actually intended for?

---

Learn more:  

• Wyche, S., Olson, J., & Karanu, M. (2019). Redesigning agricultural hand tools in western Kenya: Considering human-centered design in ICTD. Information Technologies & International Development, 15, 97–112.
• Abera, W. (2026) Where does authority reside in AI-enabled agricultural advisory systems? CGIAR’s partnership architecture as the knowledge layer for AI-enabled agricultural advice. Accra (Ghana): International Center for Tropical Agriculture (CIAT). 16 p.
• Emily Duncan, Sarah-Louise Ruder, Krishna Bahadur KC, Hannah Wittman, Evan Fraser, The political economy of agricultural data governance: Farmer perspectives on trust in data sharing, Journal of Rural Studies, Volume 123, 2026, 104076, ISSN 0743-0167.
• Rashid Parvez Khan, Saurabh Gupta, Regina Birner, Dhirendra Mani Shukla, Thomas Daum, Agri-tech startups in India: Knowledge recombination under resource constraints, Journal of Rural Studies, Volume 125, 2026, 104226, ISSN 0743-0167.