In development circles, agricultural extension is often treated as settled infrastructure. Governments hire agents, agents visit farmers, farmers adopt improved practices, yields rise. The logic is intuitive, the model is decades old, and the results are consistently disappointing.

Impact evaluations across sub-Saharan Africa and South Asia tell a stubborn story. Public extension services routinely reach fewer than one in ten farmers in a given year. Among those reached, adoption of recommended practices is modest, and productivity gains are often statistically indistinguishable from zero.

This is not a failure of effort or funding. Countries have poured billions into extension reform, from training-and-visit systems to decentralized farmer field schools. The persistence of underperformance suggests the problem is structural, not operational. Understanding why traditional extension keeps failing—and what alternatives are outperforming it—reveals broader lessons about how development interventions actually reach the people they intend to serve.

The traditional extension model emerged from a specific historical context: the American land-grant university system of the late 19th century, later exported through colonial administrations and postwar development agencies. Its architecture assumes a linear flow of knowledge—from researchers to extension agents to farmers—correcting an information deficit that supposedly holds productivity back.

In practice, this means a government ministry employs agents assigned to geographic zones. Agents receive periodic training on approved technologies—improved seed varieties, fertilizer regimes, pest management protocols—and are expected to disseminate these to farmers through field visits, demonstrations, and group meetings. Ratios of one agent per one thousand to three thousand farmers are typical, and often optimistic.

The model rests on three assumptions that empirical work has steadily undermined. First, that farmers lack information rather than facing binding constraints in credit, insurance, or markets. Second, that centrally validated technologies are appropriate across heterogeneous agroecological zones. Third, that agents are effectively supervised, adequately resourced, and motivated to reach remote clients.

When any of these assumptions fails, the system's outputs collapse. When all three fail simultaneously—as they typically do—extension becomes what evaluators politely call a low-impact expenditure. The remarkable thing is not that the model performs poorly, but that it has persisted so long despite performing poorly.

Takeaway

Interventions inherit assumptions from the contexts that created them. When those assumptions travel across borders without scrutiny, the intervention travels but the results do not.

Extension systems fail through a predictable set of mechanisms, well documented across country case studies and randomized evaluations. The first is coverage. Agents concentrate visits on accessible, better-off farmers—those near roads, those already using improved practices, those who resemble the agents themselves. Marginal farmers, women, and remote communities receive little or nothing.

The second is content relevance. Recommendations flow from national research stations calibrated to average conditions that describe no specific farm. A blanket fertilizer recommendation across a district with variable soils, rainfall, and cropping systems will help some farmers, harm others, and confuse most. Meta-analyses of extension impact show effect sizes that shrink dramatically when researchers account for heterogeneity of context.

The third is accountability. Extension agents report upward to ministries, not downward to farmers. Farmers cannot fire an unhelpful agent or reward a good one. Monitoring focuses on activities performed—visits made, meetings held—rather than outcomes achieved. Predictably, agents optimize for what is measured.

The fourth is the underlying binding constraint problem. Even when information reaches farmers and is locally appropriate, adoption requires cash, tolerable risk, and functioning input and output markets. Extension addresses none of these. Telling a farmer to apply more fertilizer when she cannot afford it, cannot insure against drought, and cannot reliably sell surplus grain is not helpful advice.

Takeaway

Information is rarely the binding constraint in poverty. When a program addresses a non-binding constraint, it can be well-implemented and still change nothing.

The evidence base for alternatives has grown considerably over the past fifteen years, and three approaches stand out. The first is farmer-to-farmer learning, formalized through models like lead farmer networks. Randomized trials in Malawi, Uganda, and India show that information transmitted through trusted peers—particularly peers selected for social centrality rather than technical status—reaches more farmers and produces higher adoption rates than agent-delivered advice.

The second is ICT-based extension. Voice-based advisory services in local languages, delivered via basic mobile phones, have shown persistent effects on knowledge and modest but real effects on practices and yields. Precision Agriculture for Development's work in India documented sustained increases in recommended practice adoption at a fraction of the per-farmer cost of conventional extension. The mechanism is not novelty but scale, repetition, and content tailored to local conditions.

The third is private extension, particularly when embedded in output market relationships. Contract farming arrangements, cooperative-based advisory services, and input dealer networks internalize accountability: the extension provider bears cost when advice fails. Evidence from Kenyan dairy and Ghanaian cocoa suggests these arrangements outperform public services on both reach and impact, though with obvious limits in commodities and populations that lack organized value chains.

None of these alternatives is a universal replacement. Each has failure modes—elite capture in farmer networks, digital exclusion of the least literate, and the exclusion of subsistence farmers from private models. But each addresses at least one of the structural weaknesses that cripples public extension, and their costs per behavior change are consistently lower.

Takeaway

Effective alternatives rarely deliver the same service more efficiently. They usually restructure who bears cost, who holds accountability, and what counts as success.

Agricultural extension is not failing because governments lack commitment or agents lack skill. It is failing because the model itself misdiagnoses the problem it was built to solve, and because its accountability structures point in the wrong direction.

The alternatives that work share a common feature: they align incentives between the information provider and the farmer, whether through peer trust, market relationships, or repeated low-cost interaction. They do not assume the farmer's ignorance; they respect the farmer's judgment.

The broader lesson extends beyond agriculture. Development interventions inherited from earlier eras often persist by institutional inertia long after evidence suggests better approaches. Progress requires the willingness to retire what does not work, even when it has always been done.