Imagine a farmer in India growing millet on two acres of land. Rain has been scarce this year, and a new type of pest ate much of his crop. He’s concerned that his tried-and-true practices are not as effective as they used to be, and he’s wondering what might increase his odds of a better harvest next season. He pulls out his phone and asks an AI chatbot, which provides him with recommendations.

This scenario represents a new use of generative AI in agriculture and illustrates a new opportunity to bring innovations to the world’s farmlands.

Generative AI promises benefits for society at large as well as for the agriculture industry, from smallholder farmers to larger cooperatives to the giant players in commodity crops, seeds, fertilizers, and pesticides. The extent to which the industry has begun experimenting with targeted uses of generative AI offers multiple lessons learned for other business sectors.

More than ever, farmers need to make more efficient use of natural resources such as soil, seed, and water and lessen the toll that farming takes on the environment through methane generation, pesticide use, and other factors. The food system generates about one-third of greenhouse gas emissions globally, according to the United Nations. Consumers are demanding more sustainably grown products, as Harvard Business Review points out. At the same time, agriculture is part of the climate; for example, when farmers grow perennial crops, their soils store more carbon. But agriculture is also vulnerable to extreme weather patterns that are causing more floods, droughts, and wildfires.

In agriculture, AI is already being used to improve food production and sustainability. For example, Vista GmbH, a subsidiary of BayWa, Germany’s largest digital agriculture company, uses satellite data and AI models to forecast and improve yields in a project with SAP. Such tools can analyze data in ways that help farmers produce more food more sustainably for the earth’s growing population while also helping them adapt to a changing climate. They can also produce reliable data on sustainability for shareholders and government bodies.

Such data is increasingly important. Agriculture’s role in climate change was highlighted in November 2023 at COP28, the United Nations Climate Change Conference, where 159 countries signed the United Arab Emirates (UAE) Declaration on Sustainable Agriculture, Resilient Food Systems, and Climate Action. Through this declaration, countries committed to improving sustainable farming practices and pledged to include agricultural emissions in their next round of climate targets. In addition, the United States and the UAE committed about $17 billion toward agricultural innovations to address climate change.

With the added capabilities of generative AI, farmers and their advisors can more quickly and directly access agronomic data and analysis, enabling them to make near-real-time decisions about soil treatments, seed varieties, and market conditions. Using chatbots, farmers can ask natural language questions in their native languages (When should I plant this crop? How much and what kind of fertilizer should I apply?) and get answers that any farmer, regardless of their training or background, can understand.

For agricultural companies, generative AI can gather and crunch even more granular and timely information and analysis that help them improve their products, customer service, and processes and operations to better meet business goals—including sustainability targets.

While the promise is there, agricultural companies will need to carefully navigate four key areas as they implement generative AI:

1. Data quality and model training
2. Risks associated with the application of generative AI, such as false information
3. The need to ensure that IT systems and operational processes are set up to use generative AI
4. Developing trust in these systems and their results so that farmers use the technology

This article describes two examples of generative AI in use today and discusses how companies can reap the benefits and reduce the risks in each of these four areas.

Example 1: Generative AI chatbot provides fast advice to smallholder farmers

With the increased reliance on data in farming, every major player in agriculture is experimenting with generative AI—and the first application to find traction is a virtual agronomist chatbot. That’s because the ROI is clear: the cost of providing a chatbot is low in comparison to regularly sending human agronomists to remote areas, and the potential benefits are high. Specifically, a chatbot’s ability to access and analyze data on weather patterns, plant genetics, pesticides, soil chemistry, and other areas offers farmers direct, immediate help in improving their agricultural practices.

In India, agriculture is a major contributor to the gross domestic product, employing nearly half the population directly or indirectly, notes Analytics India Magazine. Yet these farms are very small; the average size is just 1.08 hectares (about 2.5 acres), reports Fortune India. Half lack basic farming equipment, and three-quarters are at risk of crop damage from pests and weather, according to a May 2023 McKinsey report.

Digital Green, a global nonprofit that builds technology to help farmers in developing nations, began rolling out an AI assistant in 2023 with farmers in India, Kenya, and Ethiopia. The assistant, an AI chatbot built into Telegram, initially targets agricultural extension agents, who are professionals providing training and information for farmers. Such agents can be few and far between—the ratio is as low as one agent for every 1,000 farmers in some areas—which makes it hard to provide timely, accurate, localized advice, says Jona Repishti, Digital Green’s head of global gender programs.

“Extension agents play a critical role in helping smallholder farmers get the information they need to be climate-resilient and get better yields at a time when there is soil degradation and new types of pests,” says Repishti. “But this requires them to be experts in plant genetics, soil health, new types of seeds, climatology, and market and government regulations.”

In some cases, agents may spend a week to research a problem and deliver an answer, which is often too late. “Agriculture is very time-sensitive,” says Repishti. “You need the right information now or pests quickly devastate your crop.”

Digital Green built the AI assistant to improve communication and advice to farmers, thereby increasing the speed and effectiveness of extension agents—which in turn will increase farmers’ faith in the agents’ advice.

Digital Green expects to have the chatbot in the hands of 10,000 agents by the end of 2024. In the early phases of the rollout, the agents used it in several ways: to research across volumes of scientific reports and papers, to quickly answer farmers' questions in the field, and to present information to groups at meetings. The agents have also begun sharing the chatbot with farmers to use directly, says Repishti. Eventually, Digital Green expects the chatbot to incorporate local market data so that a farmer could ask a question such as, “What is the most profitable crop to grow, given variables such as the costs for supplies like seeds and pesticides and local selling prices for certain crops or livestock?,” according to Repishti.

Example 2: Milking more data from cows

Datamars, a company based in Switzerland, also develops technology to help farmers become more sustainable. Datamars initially sold radio-frequency ID tags to track farm animals, but today it has a strategic focus on both hardware and software to enable smart farming, according to Kevin Coffey, the company’s global head of smart farming.

In early 2023, Datamars acquired the sensor business of a technology partner, Connecterra, which has built software that provides data to farmers, their advisors, and agricultural companies. Today, the two companies offer a service that uses data from sensor-chipped animals to create insights for farmers as well as the Connecterra software to crunch data and create insights for farm-supporting partners such as advisors and feed companies. The service helps farmers better manage their animals and in turn provides data that helps companies meet sustainability targets.

Through networked sensors—“like Fitbits for cows,” says Coffey—a farmer can see the individual characteristics and behavior of each animal, including eating, rumination, and even estrus, or fertility status. That helps farmers better manage their animals, and, says Coffey, “Good farming and sustainability are synergistic.”

For example, dairy farmers want the cows to be “in calf” as much as possible to maximize milk production, so they can improve collection rates by tracking estrus—which means they don’t need to buy more cows to maintain the same level of milk production. Fewer cows also means less methane, a significant contributor to global warming, Coffey explains. Large dairy processors and slaughterhouses need this information to show customers that their milk and beef is produced in a sustainable way, he notes.

Given how fast generative AI is developing, it may not be long before other agricultural apps are in the field. (See sidebars: Computer Vision in Farming Equipment and Developing New Crop Breeds Faster.)

Computer Vision in Farming Equipment

Some equipment, such as tractors and sprayers, already incorporate cameras and use AI to examine images to spot weeds and spray herbicide in just the right places. This reduces herbicide use and costs for farmers. For example, John Deere uses machine vision and machine learning in its agricultural equipment, Vision Systems Design reports.

Traditionally, training such AI requires a lot of images and tedious work, says Vikram S. Adve of the University of Illinois at Urbana-Champaign, where he is the Donald B. Gillies professor of computer science and director of the AIFARMS (Artificial Intelligence for Future Agricultural Resilience, Management, and Sustainability) Institute. “This supervised machine learning is very expensive and requires certain expertise,” says Adve.

By using the foundational image models now available, which have been trained on very large but general data on all sorts of images, companies could fine-tune agricultural vision apps using synthetic images that are quickly generated by generative AI. This dramatically reduces the time and expense required to train AI to spot particular things such as specific indications of disease.

Datamars is already evaluating such an application through a collaboration with Serket, a company that develops AI-based vision systems for livestock, according to Kevin Coffey, Global Head of Smart Farming at Datamars. AI trained on a foundation model might recognize a cow as distinct from a goat, for example. But generative AI could further train the model to distinguish between different behaviors of those animals and potentially identify the early onset of disease.

Key takeaways: Risks and rewards

Generative AI holds great promise for agriculture, but companies seeding new applications should pay special attention to the following factors.

Data and model training

All agriculture is local—so for generative AI to be useful, it requires specific, localized data. Climate, soils, diseases, growing practices, and many other factors differ significantly from one region to another. “The answers you get to the same question about the same crop and even the same climate conditions could be quite different in central Illinois versus Mexico versus other corn-growing regions of the world,” explains Vikram S. Adve of the University of Illinois at Urbana-Champaign, where he is the Donald B. Gillies professor of computer science and director of the AIFARMS (Artificial Intelligence for Future Agricultural Resilience, Management, and Sustainability) Institute. That means generative AI must be trained not only on domain-specific data but on location-specific data.

That presents a twofold challenge—first, being able to gather and use the data. Some data, such as weather data, is publicly available, but some, such as details on specific types of seeds developed by specific companies, might be proprietary, which raises legal risk. In addition, if companies are collecting and analyzing data from individual farms, they need to adhere to data privacy rules and make sure they have the express permission of farmers.

Second, such specificity makes model training more complicated. While ChatGPT—which is trained on information from the Internet—might be a start, agricultural chatbots need to be trained on trusted data sets, such as peer-reviewed agricultural research and best practices. But since farming is local, that training also requires specific data on regional variations in weather patterns, typical crops grown in that geography, and even market conditions in specific areas.

Digital Green’s Repishti says the organization has developed databases for select high-value crops. In Ethiopia and Kenya, the chatbot supports over 40 value chains, including potato and rice, which are among the major crops in those countries. It plans to expand the databases to additional crops and livestock.

But static databases may not be enough to enable the chatbot to answer highly specific questions in a changing context. So Digital Green is working to enable the chatbot to connect with dynamic databases through retrieval augmented generation (RAG) as well as fine-tuning the language models to keep up with farmers’ changing needs. In this way, the chatbot could access the most recent data on commodity markets and selling prices, for example.

With RAG technology, a model has access to the most current information, helping to ensure accuracy and helping it to sync up with hyperlocal data. This means the AI model can retrieve information while answering a specific question. RAG also reduces the chances of the model “hallucinating,” or making up erroneous answers based on data patterns on which it was trained. However, RAG requires specialized expertise and technology; for example, it depends on coding that enables relevant information to be drawn from specific databases.

Legal risks associated with data ownership

Any company developing a generative AI tool for farmers needs to make sure it owns the data it intends to use or at least has permission to use it. A company must also have knowledge of various laws around the world that govern data ownership and data mining. For years, large agricultural equipment vendors have collected data through sensors and cameras on tractors in addition to data from farmers, so they may have legal ownership of certain information, Adve notes.

However, there are legal opportunities as well. If trained using the right data, chatbots can help farmers comply with local regulations.

In addition, agriculture companies, technology suppliers, and software vendors should make sure that users understand the possibility of error. While it is improving, generative AI is still prone to hallucinating. One reason that Digital Green rolled out its chatbot only to extension agents is that the agents have the knowledge to spot such errors. “The agents are another layer of security because they are experts that can critically assess the information the chatbot offers,” Repishti says.

Adve notes that evaluating the accuracy of generative AI tools is difficult and time-consuming. There are good reasons that most of the large language models come from some of the biggest technology companies in the world, which have the vast resources required to reduce the chance of errors. For all the promise that generative AI holds, business leaders need to remember that it isn’t perfect.

IT systems and operation processes

Because farming is local, sophisticated technical tools were developed in isolation and were not widely adopted until recently. That means that much of the data is not standardized and that systems require integration, potentially leading to some heavy lifting for companies that want to gather and analyze agricultural data. For example, different countries use different codes to identify various cattle breeds, says Coffey.

Developing farmers’ trust

Companies need to pay special attention to gaining the trust of farmers. Traditional farming practices have been developed over many generations; often the same family has worked the same land for hundreds of years. So it is understandable that farmers are skeptical of new technologies like data analytics and generative AI. (For more on this issue, see “How AI Transforms Agriculture.”)

Clear explanations can help. State what generative AI can do, what it cannot do, and how it works. For example, describe how generative AI provides information expressed in probabilities and can make mistakes (just as people do) and also how it can take in much more data than people can. Reviewing these issues helps users understand the technology’s strengths and weaknesses—and how it benefits them.

Companies need to pay special attention to gaining the trust of farmers.

Digital Green focuses on gaining farmers’ trust. “It's notoriously hard to change farmer behavior and practices in the fields,” Repishti says. When farmers ask their extension agent questions, it can take time to get an answer if the agent must research it. With the AI chatbot, agents can immediately respond to farmer queries with relevant, useful, and accurate advice. In addition, Digital Green’s bot can handle multiple languages and respond to spoken questions rather than relying on one language and text. “Meeting farmers on their terms helps build trust between farmers and extension agents,” she notes.

The trust factor is important even in developed countries such as the United States because the use of digital technology in agriculture is still relatively new. The AI Institute for Climate-Land Interactions, Mitigation, Adaptation, Tradeoffs, and Economy (AI-CLIMATE), based at the University of Minnesota, is researching how AI can help U.S. farmers adopt and profit from more climate-smart practices.

“An important factor for this institute to be successful is to bring in the farmer,” says Zhenong Jin, assistant professor of digital agriculture at the University of Minnesota. “We are co-creating the solution with them. Farmers play a significant role in this process. It’s not just a scientist telling them what to do.”