Before I studied collective intelligence or AI, I studied crops. This post is about what I learned from the older system, and what I think the newer one is forgetting.
The Extension Agent Knew Your Name
Agricultural extension is one of the oldest knowledge distribution systems in the world. The basic model has been running for over a century: a trained agent visits farming communities, observes local conditions, listens to what farmers are dealing with, and offers advice adapted to what they’ve seen.
The system was never efficient. One agent might serve hundreds or thousands of farmers. Travel was slow. Record-keeping was patchy. The advice wasn’t always right. In many countries, extension services have been underfunded for decades, and in some regions they’ve collapsed entirely.
But extension got something structurally right that most people overlook: the advice was distributed through a human network, and that network introduced variation. Different agents had different training, different experience, different relationships with the communities they served. An agent in Oyo State didn’t give identical advice to an agent in Kwara State, even when they were drawing from the same manual. The advice was filtered through local observation, personal judgment, and conversation with the farmers themselves.
This meant that across a region, farming communities were never all doing the same thing at the same time for the same reason. The system was inefficient, yes. But that inefficiency was also a form of resilience.
The Phone Knows Your Location
Now compare what’s arriving to replace it.
AI-powered agricultural advisory platforms are scaling across Sub-Saharan Africa and South Asia. Farmers receive planting recommendations, pest identification, fertiliser guidance, and market timing through their phones. The tools are genuinely useful. They reach farmers that extension services never could. They operate in local languages, they’re available at any hour, and they don’t require a government salary to keep running.
But the architecture is different in a way that matters.
When a thousand farmers in the same district open the same app and ask the same question about when to plant maize, they get the same answer. The answer might be good. It might be the best answer available given the data. But it is one answer, delivered uniformly, to people whose fields, soils, microclimates, and risk tolerances are not uniform at all.
The extension agent, for all the system’s flaws, would have told farmer A something slightly different from farmer B, because farmer A’s plot drains poorly and farmer B is near a river. The app doesn’t know this. Or if it does, it adjusts at the level of GPS coordinates, not at the level of “I walked your field last season and saw what happened.”
What Gets Lost
Three things disappear when you move from distributed human advice to centralised algorithmic advice.
Local knowledge stops flowing upward. Extension agents didn’t just deliver information. They collected it. They noticed what farmers were doing that worked, carried those observations to the next community, and sometimes carried them back to researchers. The system was a two-way channel. AI advisory platforms are, for the most part, one-way. The farmer receives a recommendation. The platform receives usage data. But the farmer’s observation that a particular local variety outperformed the recommended one in heavy clay soil, that knowledge has no path back into the system.
Risk gets concentrated instead of spread. When every farmer in a cooperative follows the same recommendation and the recommendation is wrong, they all fail together. When advice is varied, even imperfectly, some farmers will have planted differently, harvested earlier, chosen a different variety. The community has a fallback. Anyone who has watched a farming community recover from a bad season knows that recovery depends on the households that did something different having something to share.
Trust becomes binary. Farmers trusted their extension agent (or didn’t) based on years of relationship, observed competence, and whether the agent’s previous advice had worked. That trust was granular and earned. Trust in an AI app is more binary: you use it or you don’t. And when it fails, the failure isn’t attributed to a specific agent whose judgment was off. It’s attributed to “the app,” which makes it harder to diagnose what went wrong and easier to either abandon the tool entirely or, paradoxically, to assume the failure was the farmer’s fault for not following instructions precisely enough.
What Extension Understood About Knowledge
I spent my undergraduate years studying crop and environmental protection. The curriculum was built around a basic insight: agricultural knowledge is local. A recommendation that works in one agroecological zone can fail in the next one. Soil, rainfall, temperature, pest pressure, market access, labour availability, cultural practice: all of these vary at scales that national or even regional models struggle to capture.
Extension systems, at their best, respected this. They were slow because local knowledge is slow. It accumulates through seasons, not through training epochs. It lives in the memory of farmers who have watched their fields for decades, not in datasets scraped from satellite imagery.
AI advisory tools fill a gap that neglected extension services left open, and for many farmers they represent the first access to any structured agricultural guidance at all. But the people building these systems would do well to understand what the older system knew about the structure of agricultural knowledge, and to design with that understanding rather than against it.
The extension agent who walked between farms and listened wasn’t just delivering information. That agent was maintaining something: the diversity of approaches within a community, the flow of local knowledge back toward researchers, the kind of trust that’s earned over seasons rather than downloaded. An AI advisory system that understood this would look different from what most platforms currently offer. Not worse. Not slower. Just designed with the knowledge that a community’s resilience depends partly on its members making different choices.