When AI Works for Farmers

Approximately 40 percent of the global workforce is employed in agriculture. The majority of those workers are smallholder farmers — people who cultivate less than two hectares, who lack access to extension services, who make planting, input, and selling decisions based on personal experience and social networks rather than data. These farmers are the people whose productivity improvements would have the largest impact on global food security, rural poverty, and development outcomes. They are also among the last people that standard AI product development processes would identify as a target market.

The AI tools that have worked for smallholder farmers in Asia look nothing like enterprise software. They sound like neighbors.

Kisan AI services, variously delivered through IVR (interactive voice response) systems, WhatsApp bots, and purpose-built mobile apps, have deployed across India, Bangladesh, Indonesia, and Vietnam over the past four years with a consistency of design that reflects accumulated lessons about what actually works. The common elements: voice-primary interface (most effective with intermittent literacy and older phones), local language (not just Hindi but the specific dialect of the specific agricultural region), hyperlocal data (not average national weather but the forecast for the specific district, the pest pressure in the specific crop zone), and integration with trusted social infrastructure (delivered through the cooperative or the input supplier rather than as a standalone app).

Agritech India’s Kisan Helpline, which operates at scale in Uttar Pradesh and Bihar, handles approximately 3 million agricultural queries per month through a voice interface in multiple languages. The queries range from pest identification (“what are these spots on my tomato leaves?”) to market price information (“what’s the mandi price for wheat today in Varanasi?”) to weather-based planting advice (“is it safe to plant paddy this week?”). The AI components of the system — the natural language understanding, the crop advisory logic, the market data aggregation — are built on relatively modest models by frontier standards, but they are tuned with extraordinary care for the specific knowledge domain and user environment.

The evidence on impact is mixed in the specific and positive in the aggregate. Randomized evaluations (which agricultural development programs have a better track record with than most technology programs) have found that access to AI advisory services correlates with higher yields for farmers who use the services consistently, reduced pesticide overuse, and more effective timing of irrigation in water-stressed environments. The effect sizes are not transformative — we are talking about 10 to 20 percent yield improvements, not the kind of transformative leap that technology press coverage implies. But applied across hundreds of millions of smallholder farmers, 10 to 20 percent yield improvements are enormously consequential for food security and rural income.

The pest identification application is the one where AI most clearly outperforms the counterfactual. Plant Village, the open dataset of crop disease images created by Penn State University researchers, trained models that can identify a large number of crop diseases from smartphone photos with accuracy comparable to trained agronomists. The original research was published in 2016; the application has been deployed and improved by dozens of organizations across South and Southeast Asia over the following decade.

What the deployment experience revealed is that the model accuracy in lab conditions was significantly higher than the model accuracy in deployment conditions. The images that farmers take with cheap smartphones, in field conditions, with inconsistent lighting and background, are not the images the model was trained on. The gap between benchmark accuracy (the model is right 90 percent of the time on test dataset images) and deployment accuracy (the model is right 60 percent of the time on images taken by farmers in actual fields) is the difference between a useful tool and a misleading one.

The organizations that built effective deployments solved this by building feedback loops: when the model gives an uncertain or incorrect diagnosis, the farmer can report the outcome, and that data is used to retrain the model on deployment-representative images. This requires the organizational infrastructure to collect feedback, the data labeling capacity to turn it into training data, and the engineering capacity to iterate on the model. These requirements are not glamorous, and they are not what most AI companies’ core capabilities are. They are the reason that many technically impressive agricultural AI pilots failed to scale while a smaller number of technically adequate but operationally disciplined deployments did.

Indonesia provides the most interesting case study in agricultural AI at national scale. The country has approximately 36 million smallholder farm households, diverse agroecological zones ranging from lowland Java to highland Sumatra to the outer islands, and a government that has been willing to experiment with digital agricultural services through the Ministry of Agriculture’s information systems. The SIMLUHTAN platform, the Petani digital cooperative network, and several private-sector services have collectively created an ecosystem of digital agricultural services that reaches a meaningful fraction of Indonesian farmers.

What makes Indonesia distinctive is the role of the government agricultural extension system. Indonesian extension workers — Penyuluh Pertanian Lapangan, or PPLs — number approximately 40,000, organized into a national system that covers essentially every agricultural area. Digital tools that work for extension workers are tools that can reach farmers through trusted intermediaries with local knowledge, language facility, and established relationships. The most effective Indonesian agricultural AI deployments have not replaced extension workers; they have augmented them, giving the PPLs tools for more consistent advice, better market information, and access to specialist expertise for diseases and pests outside their personal experience.

This augmentation model — AI as a tool for frontline workers rather than a replacement for human contact — appears consistently in the successful agricultural AI deployments across the region. Vietnam’s agricultural cooperatives using AI crop advisory services. Thailand’s rubber-tapper support programs. Bangladesh’s Krishi Gorbo project for rice farmers. The common thread is not the AI itself but the organizational structure through which it is delivered: human networks with established trust and local knowledge, enhanced by AI tools that extend their capabilities.

The implication for development AI strategy is one that the technology industry finds difficult to absorb. The most impactful AI deployments in agricultural development are not the most technically sophisticated products. They are products designed with obsessive attention to the constraints and capabilities of the specific user population, delivered through existing social and organizational infrastructure, with feedback mechanisms that allow continuous improvement in deployment conditions.

This is not how technology companies build products. Technology companies build products for a generalized user, deploy them at scale, and measure success by user numbers and engagement metrics. Effective development AI requires building for a highly specific user (a smallholder rice farmer in the Mekong Delta), deploying through specific organizational infrastructure (the village cooperative or the government extension service), and measuring success by development outcomes (yield, income, reduced pesticide use) that are harder to track than app downloads.

The funding structures for this kind of development AI also look different from technology investment. Impact investors, development banks, government procurement, and philanthropic grants — not venture capital seeking unicorn returns. The organizations best positioned to do it are NGOs and government agencies with development mandates and local organizational infrastructure, not AI startups with exit ambitions.

The agricultural AI that works for farmers in South and Southeast Asia offers a vision of what AI can accomplish in high-stakes, resource-constrained environments when it is designed for those environments rather than retrofitted from products designed for different users. The yields are modest in technology terms. The development impact is significant.

The lesson for the broader question of AI and development is that the technology is rarely the limiting factor. What limits effective deployment is organizational infrastructure, local knowledge, appropriate product design, and the patience to iterate in deployment conditions rather than benchmark conditions. Those are human capabilities, not algorithmic ones.

The farmers who are getting better advice from AI-augmented extension workers in Bihar and Sulawesi and the Irrawaddy Delta are not benefiting from the frontier models that dominate AI press coverage. They are benefiting from carefully deployed, aggressively localized, operationally disciplined AI products that look unremarkable from the outside and produce consistent results in practice.

That is exactly what the development world needs more of.