Bridging data divides at scale through tech & AI
By Kailash Raj Pandey (‘24 Strategy Fellow)
When new technologies emerge, they don’t just create commercial value. They also offer important opportunities to address social challenges in low-income countries, whether by reducing income gaps, increasing access to education, or improving health outcomes. Very often, the social sector struggles to harness the full promise of new technologies.
But this is starting to change. A growing number of groundbreaking social-sector organizations have found success using technology to drive meaningful change in low-income environments. This blog explores two such organizations that The Agency Fund is supporting. Both work in the context of agricultural extension services, or efforts to help farmers in low-income contexts adopt more productive and climate-resilient farming practices: Digital Green, which connects farmers to frontier AI technology, and Precision Development (PxD), which does similar work through traditional mobile technologies.
Connecting rural farmers with conversational AI
Digital Green works in India, Ethiopia, Kenya, and Nigeria – which are collectively home to over one-third of the world’s 570 million smallholder farmers. Since 2008, the organization has supported government-run agricultural extension systems by creating farmer-led digital videos that are screened in community groups. RCTs conducted in Ethiopia and India found that Digital Green’s community video approach had impressive results: crop yields increased by up to 17% and income increased by up to 24%.
Over the past year and a half, Digital Green partnered with Microsoft, OpenAI, and other major tech partners to develop Farmer.Chat, an AI-powered chatbot that leverages large language models (LLMs) for agricultural extension services. After trialing Farmer.Chat, Digital Green found that the LLMs can achieve entirely new levels of scale (in the first two months, 15,000+ farmers used it for the first time, sending 300,000 queries) and can potentially reduce costs by a remarkable 100x (from $35 per farmer under traditional agricultural extension to $0.35 with the LLM-powered tool; by comparison, the cost of Digital Green’s community video approach is $3.50 per farmer).
In addition to the cost reductions, Farmer.Chat can minimize language barriers – which has limited the scalability of the community video approach. Farmers can now access context-relevant advice in multiple languages (including Hindi, Swahili, and regional languages) and formats (i.e., text, video, and voice). Farmer.Chat product is built on an agile tech stack that integrates with local language translation datasets and services in each country. In addition to technological advancements, field workers using the tool have reported higher confidence in the advisory services they provide.
In our discussion with Digital Green, we uncovered several important insights in their development process:
Pilot Strategy: Enabling early learning and product improvements is a key challenge for any early stage product. To achieve this, Digital Green piloted Farmer.Chat with extension workers before scaling to farmers (e.g., clickable question suggestions, voice features). Interviews with pilot participants revealed four main use cases for the bot: improving extension workers’ knowledge, helping them answer farmers’ questions, helping them facilitate group Q&A sessions, and helping them train farmers to use the bot directly. They also found that early adopters helped build trust and promoted the product within their communities, demonstrating the value of a phased rollout.
Feedback: Farmer.Chat leverages continuous user feedback and real-time data to drive uptake and results. Automated, real-time dashboards track usage behavior, engagement, and retention, while feedback mechanisms (e.g., thumbs up/down, open responses) highlight content gaps. Telegram-based group chats with farmers supplement this process, enabling quick product refinements. For example, user demand led to the integration of crop diagnostics, image upload feature, and weather information. Likewise, a partnership with Plantix enables image analysis for pest and disease detection.
Product iteration: To mitigate risks of the bot giving erroneous advice, Digital Green tested a curated Retrieval Augmented Generation (RAG) system using historical data, vetted content through experts, local weather information, and soil databases. The team then designed evaluation systems to analyze farmer questions and bot responses for accuracy and effectiveness using a Reinforcement-Led Human Feedback process. Additionally, they are establishing a data trust to fine-tune an agriculture-specific LLM that integrates with WhatsApp, Telegram, and a standalone app, ensuring scalability while maintaining reliability and, ultimately, quality advisory for smallholder farmers.
Innovation through traditional mobile technologies
Similar goals are pursued through more traditional mobile technologies by Precision Development (PxD). PxD is a global nonprofit organization that leverages data-driven, technology-based approaches to scale innovations that millions of farmers can use to improve their lives. They provide digital agricultural extension services directly to farmers using simple technologies like voice and SMS on basic-feature phones that are accessible to nearly every farmer in the world. Their goal is to improve the productivity and livelihoods of 100 million smallholder farmers in low- and middle-income countries. PxD’s tech-enabled programs serve as a bridge between cutting-edge, highly-scalable, and cost-effective agricultural innovations and the farmers who need them, currently reaching more than 18 million farmers (at a cost of $0.94 per user).
PxD’s Ama Krushi tool, for instance – meaning “Farmer’s Friend” in the Indian state of Odisha – provides farmers with free customized information (weekly advice and answers to queries) through automated voice messages as well as an automated, interactive phone hotline. Serving more than 7 million farmers as of 2024, Ama Krushi also focuses on relevant advisory content developed in local dialects and customized with real-time, local information such as pest observations and weather. PxD’s impact evaluation showed that the service helped farmers cope with common weather shocks such as excess or inadequate rainfall. In areas affected by excess rainfall, farmers with access to Ama Krushi had 9% higher agricultural production compared to areas without access; in areas affected by inadequate rainfall, the likelihood of severe crop loss was 21% lower for farmers with access to Ama Krushi. PxD estimates that they deliver $9-15 of additional farmer profits for every $1 invested in their services.
In addition to traditional voice and SMS technologies, PxD also leverages remote sensing, AI-based weather forecasting, WhatsApp- and Telegram-based chatbots, data science and analytics, and other frontier technologies to deliver world class digital agriculture services. PxD focuses on identifying evidence-based use cases to select the right technology to address farmer problems and deliver impact in each context. PxD is currently advising India’s Ministry of Agriculture and Farmers’ Welfare on their AI strategy, evaluating a range of AI-based digital agriculture use cases, and developing an LLM-based farmer-facing solution to improve the customization of voice-based advice to farmers in local dialects.
Some of PxD’s main strategies are highlighted below:
Developing a data ecosystem to enable innovation: PxD’s voice and SMS tools emphasize flexibility, localization, and customization. Initially, farmers were profiled via call center interviews using government data on recipients of extension services or input subsidies. PxD improved its database to include more profile information (e.g., geo-location, existing practices, demographics, and crop/livestock value chains). Responding to feedback, PxD also integrated weather data from the Indian Meteorological Department, enabling tailored advice through tools like Ama Krushi.
Continuous iteration and scalability: By operating at scale, these tools produce large amounts of user feedback data, which PxD utilizes for A/B testing, human-centered design, and systematic experiments and trials. Over 60 A/B tests and several randomized evaluations have been conducted. In Kenya, for example, it found that small tweaks to a message (i.e., using urgent language) increased farmer engagement by three percentage points.
Building trust and sustainability through local partnerships: With Ama Krushi, PxD embedded the service within the state Department of Agriculture and Farmers’ Empowerment; through its “Build-Operate-Transfer” model, it then transitioned the financing and operation of the service to the state government in 2022 to ensure long-term sustainability. Furthermore, PxD developed intentional strategies to expand the reach of Ama Krushi to female farmers by engaging with local women’s self-help groups, increasing the female user base from 6% to 24% between 2019 and 2022. Similarly, in Ethiopia, PxD has partnered with the Agricultural Transformation Institute to improve the government’s voice-based digital agricultural extension platform and to transfer features of PxD’s services to the government platform.
What this means for tech in the social sector
International development has a mixed track record of leveraging tech and data effectively, but these stories show what is possible. NGOs are surrounded by opportunities to build not only useful technologies, but ones that generate and leverage user data. By taking user feedback seriously, they can begin to unlock technology’s potential to cut costs, increase engagement, and reach wider audiences. Technology is also critical for scaling personalized and localized services that may be required for agricultural development.
Two lessons are clear. The first is that multiple technologies hold promise. While exploring AI, organizations should also recognize the potential of voice- and SMS-based services as well as other legacy technologies that are accessible on basic-feature phones, particularly for reaching low-income and rural populations. Second, technology and data are not magic wands. They must be paired with strong feedback mechanisms, user-centered design, and an approach to product improvement that prioritizes rapid learning and iteration. Acquiring cutting-edge tools is important – but leveraging technology for social impact is just as much about how organizations use them.