How AI and chatbots can boost crop yields

Solving the world’s food crisis has no simple answers, but it’s clear that innovation is needed to even begin addressing the problem. A big hurdle in farming, however, is that an innovation that works well in one area may fail in another due to different soils, microclimate conditions, and other factors. Context is key — making it challenging to replicate successes. One answer may lie in artificial intelligence, which offers a way to pool disparate sources of information and data to give more comprehensive, location-specific advice on improving agricultural methods. Digital Green, which became an independent nonprofit when it was spun off from Microsoft Research 15 years ago, has been developing new tools with that issue in mind. It began as a video-based service in India for agricultural extensions — a term referring to the provision of advisory services for farmers to help improve yields and income. The creation of videos featuring local extension workers and farmers talking about their practices has been a strong motivator for others, said John Collery, Digital Green’s chief product officer. “Farmers are excited to see somebody from their community talking to them via that platform about best practices to apply.” Having built a wealth of information from creating thousands of locally relevant videos in dozens of languages, the organization is now harnessing this in new ways with AI. The service Farmer.CHAT, for example, aids real-time communication between governments and farmers, and has been developed in collaboration with the startup Gooey.AI, alongside partnerships with the governments of India and Ethiopia, the Food and Agriculture Organization, and Microsoft. The conversational Telegram bot integrates content based on call center logs, transcribed training videos, and farmer feedback in local languages. Farmers can send queries via text or audio message and receive advisory responses with links to helpful content. Devex spoke to Collery about the potential for AI to deliver highly context-specific advice. This conversation has been edited for length and clarity. What’s the background and ethos of Digital Green? Our insight in the early days of Digital Green was that farmers learn well from other farmers in their local community and get excited by delivery of content via video. We therefore began to build the infrastructure to train extension agents to record and provide advisory content locally. The impact includes an estimated 24% increase in income across the 5.2 million farmers who’ve received our advisory services, 70% of whom are women, according to our own data. Over time, Digital Green has also built a repository of best practices across our work, which we now carry out in countries including India, Kenya, Ethiopia, and Nigeria. In recent years, we’ve been thinking about how we take that body of content and deliver it in new ways to more people. As part of this, our goal is to support governments to improve their extension systems and transfer the technology and processes we’re building. What has the company learned over time about delivering impactful content to farmers? The key is ensuring effective data collection and integration with existing content to allow closer contextualization of advice. One example is our work in Ethiopia with the International Center for Tropical Agriculture, which has developed models for making recommendations on how much fertilizer to apply based on the time of year and farmers’ location. Through our partnership with the Ethiopian government, we’ve also built a repository of over 40,000 of the country’s 70,000 government employees known as extension agents. Merging those two data sets through our data-integration and exchange tool FarmStack allows the provision of robust recommendations for fertilizer application rates. The upshot has been an increase in yields of up to 17% over a single season. We’re excited about what this indicates for the ability to build off the content we already have. What other methods has Digital Green been exploring for delivering content? We’ve tested various new mechanisms, including AI via a bot we’ve been developing with the startup Gooey.AI. We’ll be able to combine this with content we’ve created over many years, including the transcripts of our videos, and train the models based on that data to deliver advisory services in an even lower-cost way to increasingly large numbers of people. People can already sign up for the Farmer.CHAT system and start using it, but we plan to launch it officially by the end of the year. We haven’t yet pushed it out more broadly because a critical step for us is being able to say with confidence that when an extension agent asks a question, they’re getting appropriate, quality responses based on the local area. But in user tests so far, indications have been positive, with high volumes of requests and retention rates. How is the information conveyed to farmers? It will be multimodal. Some information will come to extension agents in the form of text, but also via text-to-voice and short-form video. The extension agents can then convey that information to the farmers they serve through the trusted channels they’ve already built. Extension agents can then act as great providers of feedback on the content we’re delivering. That two-way mechanism allows for the continued creation of a body of local knowledge, representing a step change in the speed at which we can do this. How well does this type of contextualization work? When people watch our videos, their most common response is to ask which village the person in the video comes from. Frequently, farmers watching the content will actually know the person, and are excited to see somebody from their community talking to them via that platform about best practices to apply. Making local farmers the stars of what we produce also aids trust in information-sharing. In the chatbot work, we’re investing heavily in training extension workers and leaders of local farmer organizations to be stewards of the information. Instead of just sending out a bot that farmers download and use on their own, we’re designing it to enhance the relationship they already have with local extension agents. How much do advances in innovation rely on collaboration between different organizations to ensure a real impact in practice? For the way we work, innovations that have real impact only do so through thoughtful collaboration between multiple organizations. For example, we’ve been working with the Kenya Agricultural and Livestock Research Organization, or KALRO, to build a digital community across the agricultural ecosystem via FarmStack. In Kenya, there are multiple entities — including government ministries, farmer organizations, and private companies such as fertilizer suppliers — that have country-level data sets, such as crop pricing and quantities sold by region. We have an encrypted data-sharing protocol built into our chatbot that enables all these organizations to opt in to data-sharing to improve KALRO’s research, ultimately improving the quality of agricultural extension. Such platforms, meanwhile, present capabilities that can be built on at a national level to allow a competitive landscape in agriculture to develop. We think the outcome of building strong content-driven interaction between extension agents and farmers is extremely relevant data that helps us to understand the role for private and public service providers in the agricultural ecosystem, and map where investments are needed. Visit Food Secured — a series that explores how to save the food system and where experts share groundbreaking solutions for a sustainable and resilient future. This is an editorially independent piece produced as part of our Food Secured series, which is funded by partners. To learn more about this series and our partners, click here.