We combine structured user research with rapid, feedback-driven iterations using Agile methodologies to balance deep user insight with adaptability.

Our process begins by engaging farmers and local extensionists to understand their needs and literature research to uncover weaknesses in existing advisory materials. Instead of building a perfect solution from the outset, we develop early versions for farmers to test. These are refined in short, iterative cycles based on real-world feedback. This blended approach allows us to create practical and user-friendly tools while remaining responsive to evolving needs and ideas while ensuring that the software has a real impact on farmers’ daily lives.

To help overcome language barriers, we currently use a lead farmer model. Lead farmers who understand English support other farmers in their communities by explaining and interpreting the technical advisories from Virtual Agronomist.

At the same time, we have begun developing versions in specific local languages, for example Kinyarwanda in Rwanda. French is already available.

Yes, plans for integrating multilingual support (including local African languages) are highlighted as a priority, recognizing that language is a fundamental barrier to scale, equity, and inclusivity.

Despite already providing a Kinyarwanda version for Rwanda, our plan is that when large language models (LLMs) are able to deliver robust outputs in Kiswahili and other widely spoken languages in Africa, various local languages will be incorporated to reach more farmers directly.

Currently, we are also piloting a voice-enabled feature with selected farmer groups. This emerging capability allows farmers to send voice notes in their own language and receive spoken responses, removing both language and literacy barriers. Early testing is underway in English, French, and Kinyarwanda.

The main limitation of IVR/USSD/SMS (Interactive Voice Response, Unstructured Supplementary Service Data, and Short Message Service) is restricted interactivity. These channels do not offer the same rich, two-way functionality or delivery of images and interactive content possible with chatbots.

Compared to chatbots, they do not provide real-time feedback to farmers who need immediate solutions to farming challenges. Nonetheless, these channels are under consideration as complementary methods, especially during the transition to higher digital inclusion.

We have been encouraged by the relatively balanced gender representation and broad age range among Virtual Agronomist users, especially given that it is a digital technology.

This is partly due to the accessibility of the WhatsApp chatbot, a platform already familiar to many farmers. Additionally, our lead farmer model plays a crucial role in extending the service to those who might otherwise face barriers to using digital tools. Our surveys have shown impacts of Virtual Agronomist are generally gender neutral.

Our decision models account for soil variability by using the iSDAsoil map, which is at a spatial resolution of 30 m.

Farmers are coached to go further and make their own visual soil assessments using our Know Your Soil tool. Our advisory models are tailored to individual crops and incorporate information on historic and current farmer practices through responses to Virtual Agronomist questions on WhatsApp.

The recommendations are based on the well-validated QUEFTS (Quantitative Evaluation of the Fertility of Tropical Soils) approach, which estimates balanced nutrient requirements (N, P, K) as a function of attainable yield and soil nutrient supply. QUEFTS integrates soil fertility parameters with crop nutrient uptake-yield relationships to derive site-specific, agronomically efficient fertilizer recommendations under balanced nutrition assumptions.

The model has been extensively tested across tropical cropping systems and shown to provide robust guidance for optimizing nutrient use efficiency and closing yield gaps under smallholder conditions. Our retrospective cohort surveys show an average 50% increase in yield when Virtual Agronomist is used, consistent across crops and countries.

See our paper for more detail: https://www.cabidigitallibrary.org/doi/10.31220/agriRxiv.2025.00332

While soil testing for an individual farm field may improve accuracy, these services are rarely affordable by or available to smallholder farmers. In cases where they are available, results are mostly delivered late, often after planting, and are most of the time not accompanied by advisory support to improve productivity.

Improved accuracy is not guaranteed either, as results from soil testing laboratories are often unreliable (see Hartmann and Suvannang, 2020: http://www.fao.org/publications/card/en/c/CA7091EN/). Soil test results are also only one input to nutrient recommendations, whereas the majority of samples used to train iSDAsoil were analyzed in one certified laboratory. Errors associated with soil property estimates from the iSDAsoil map are given as error layers in the online map itself.

Virtual Agronomist supports agronomic advice on irrigated and rainfed crops, but specific advice on irrigation management is not currently supported.

This may be included if there is sufficient demand from farmers.

We have made a simple weather forecast for the next 10 days available to farmers.

In addition, we have developed tools that use weather forecasts combined with farmer observations to support decisions on season start planning, planting timing, topdressing timing and method, spray timing, and wet harvests. We have not automated seasonal rain forecasts due to the unreliability of long-range forecasts.

Automated systems require high-quality photos, which is a limiting factor in actual use by farmers. We find that presenting the farmer with pictures, so the farmer’s eye is the camera, works best.

It is also educational because farmers can scroll through pictures and learn about different problems. We combine picture identification with other information, such as soil pH, to help distinguish disorders such as micronutrient deficiency. We can also rapidly develop picture libraries and disorder-management advice for new crops. When automated methods improve further, we may consider integrating them as an option.

We have designed the questions to allow advice to be tailored to local realities, and we continually improve how questions are asked to improve farmer understanding.

Where farmers cannot answer questions, we think using defaults that represent the most common (median) situation provides the safest outcome.

Virtual Agronomist uses QUEFTS principles, but not the QUEFTS soil nutrient supply equations.

How the nutrient management plan is derived is described in our paper: https://www.cabidigitallibrary.org/doi/10.31220/agriRxiv.2025.00332

We summarize QUEFTS parameters from the literature, represent them as probability distributions, and verify recommendations as far as possible against fertilizer recommendations reported in official research and extension manuals.

Virtual Agronomist adjusts for QUEFTS-derived nutrient requirements to account for soil and crop management history, target yield, soil properties, and planned manure use. Recommended fertilizer rates also depend on optimization of matching nutrient demand from available fertilizer products at the lowest cost, and the proportion of nitrogen applied basally versus at topdressing.

Yes, we will shortly launch a module on livestock production, including chickens, goats, and dairy cattle.

The advice covers different levels of intensification and guides farmers on how to move up the intensification ladder. The systems include tools for monitoring animal and economic performance and diagnosing health problems.

Farmer-reported data can be subject to both positive and negative bias. We aim to minimize these biases by:

• monitoring farmers’ understanding of questions and iteratively improving them
• putting in place safeguards (for example warnings when values go outside set ranges)
• measuring plot areas
• using random sampling in surveys
• conducting timely surveys of recall data
• designing questions to provide triangulation on responses
• running in-person field checks

We are currently doing crop cuts on samples of farmers and investigating the use of remote sensing to validate yield responses. We are also conducting surveys to establish causal pathways between adoption of Virtual Agronomist, improved agronomic practices, and yield increases.

To avoid selection bias, we use spatially stratified random sampling of farmers using Virtual Agronomist and select nearest-neighbor plots as farmer-practice controls.

We have achieved such a high degree of penetration in some counties that it is becoming increasingly difficult to avoid spillover effects. In such cases, we may need to consider selecting neighboring villages as controls, despite the challenge of matching biophysical and socioeconomic conditions. We are evaluating statistical methods for adjusting spillover-induced confounding effects through covariate inclusion and weighting. We are also designing randomized control trials whereby different villages receive different treatments.

We use Bayesian rule models to support specific farmer decisions. They synthesize and interpret knowledge from various sources, including scientific literature, expert knowledge, and farmer responses.

They provide a pragmatic solution and are evaluated through monitoring agronomic performance, yield surveys, and farmer feedback.

The impact of Virtual Agronomist is greatest when complemented by strong access to other agricultural services, and we are actively seeking partnerships with organizations that provide such support.

That said, we have found there is still significant potential to improve yields simply through better basic agronomic practices, such as diagnosing constraints to crop emergence, managing pests and diseases, and optimizing fertilizer types and application rates.

Across every other continent, the pathway to improved soil fertility has involved sustained investment in both organic and inorganic nutrient inputs over many years.

Achieving similar gains in Africa requires a comprehensive ecosystem-wide transformation that balances biophysical, social, economic, and institutional strategies. Sustainable progress depends on integrating nutrient management, diversifying farming systems, supporting farmers, advancing context-specific research, and fostering enabling policy environments. We see Virtual Agronomist as a vital component of this broader transformation.

We encourage organic inputs throughout Virtual Agronomist. In the nutrient management planner, farmers can specify how much organic manure they plan to apply and see the effect on reduced fertilizer amounts and costs.

We include advice encouraging farmers to apply organic manures with fertilizers to improve soil health and maximize response to chemical fertilizers. In the Plant Health Scout, increasing organic inputs is advised as a biological control measure for managing several plant diseases. In the Know Your Soil tool, organic additions are recommended to help tackle soil structural problems and micronutrient deficiencies. We are planning a composting module to encourage improved quality of organic inputs and advise on when to compost residues for control of several soil-borne plant diseases.

We are planning to do more work on market opportunities, as adoption of alternative crops depends primarily on what farmers can sell.

We already encourage rotation as part of management advice for many plant pests and diseases, and we plan to include more information on the economic impact of rotation. Encouraging flowering borders is also an important component of our integrated pest management strategy in the Plant Health Scout tool.

Because farmers are already familiar with WhatsApp, they find Virtual Agronomist intuitive and easy to use. This ease is further enhanced by the integration of ChatGPT, which allows farmers to ask questions and get clear explanations of unfamiliar terms.

Our lead farmer model also supports adoption by less literate users, ensuring that no one is left behind. We consistently receive positive feedback on user-friendliness and use this input to continuously refine the interface. Additionally, we are developing step-by-step guides to boost farmers’ confidence and help them get the most out of the system.

Virtual Agronomist uses ChatGPT primarily to aid communication, not as the source of agronomic knowledge, which is derived from a synthesis of global literature.

We also emphasize during training and in materials that ChatGPT is a supporting tool, not a replacement for farmer judgment or human agronomists. This helps maintain farmer agency and trust.

Our experience is that farmers highly value the nutrient management plan and the soil test results from the iSDAsoil map that come with it.

This makes sense because existing national agricultural extension advice does not cater for site-specific nutrient management. The next most popular tool is plant pest and disease diagnosis and management advice, which also aligns with major farmer constraints.

We build off existing networks of field agents and the lead farmer model rather than deploying large numbers of our own ground staff.

We foresee a mixed model of investments from input suppliers, off-takers, insurance and credit companies, and private and public extension services, all of which can benefit from supporting Virtual Agronomist. We also promote individual use through referrals, helping future adoption remain cost-free as farmers access the tool directly.

We are considering extending our advice to crop processing and storage.

We are also planning to provide farmers with more information on market opportunities to help them get better prices and support crop diversification.

Farmers provide consent by accepting our terms and conditions when signing up to the chatbot, and verbal consent is recorded during surveys.

During sign up, we provide caveats about the advice and encourage farmers to first try new practices on only part of their field. We explain these safeguards in simple, non-technical language during in-person meetings and group training. The Virtual Agronomist system is hosted in the EU and compliant with EU-based data protection laws. User data is stored in an encrypted database and is never shared with or made accessible to third parties.

The lead farmer model is intentionally designed as a pragmatic bridge to reach smallholder farmers who lack smartphones or English proficiency. Its purpose is not to replace direct empowerment, but to serve as an enabling mechanism during early stages of digital adoption.

Our system encourages farmer-to-farmer diffusion and learning, where lead farmers train others, often leading to spontaneous uptake and community-level knowledge transfer. Because farmers often look to neighbors and local leaders when trying new practices, the lead farmer model deliberately builds on existing social networks, making the digital tool feel trusted and community-owned rather than imposed from outside.

We recognize that for many smallholder farmers, digital tools can feel distant, risky, or difficult, especially when phones are shared within households, connectivity is unreliable, or past experiences with apps and SMS services have been disappointing.

Rather than assuming farmers will adopt technology simply because it exists, Virtual Agronomist is introduced through trusted local actors (lead farmers, extension staff, and partners) who already have strong community relationships. WhatsApp is used because it is familiar and widely used, and we pair it with in-person demonstrations, group training, and follow-up visits. This combination of trusted relationships, familiar channels, and clear practical benefits helps shift perceptions from “this is not for me” to “this is a useful tool that fits into how I already farm.”

Virtual Agronomist is designed to amplify farmer decision-making, not replace it. The WhatsApp chatbot asks farmers about their own fields, practices, and constraints, then uses that information to tailor recommendations, rather than pushing generic one-size-fits-all advice.

Farmer feedback is actively collected and used to refine both agronomic logic and question phrasing, so the tool evolves with farmers rather than being imposed on them. The lead farmer model also encourages peer-to-peer learning and discussion, supporting community ownership and reducing dependency on external experts. Over time, this builds confidence, knowledge, and autonomy at both farmer and community level.

We do not expect farmers to figure it out alone after a single training. For many users, especially those with lower literacy or less digital experience, human support remains essential.

Virtual Agronomist is embedded in a broader advisory ecosystem that includes lead farmers, extension agents, and partner organizations. These actors help farmers sign up, navigate the chatbot, interpret recommendations, and apply them safely in the field. We are also developing simple visual guides and step-by-step flows to reduce cognitive load, and we continually simplify questions and interface elements based on user testing. The goal is for technology to enrich human extension, not replace the relationships that enable learning and behavior change.

Experience from large-scale AgTech deployments in East Africa shows that the assumption that all farmers have phones and can use digital services independently is often a myth. Phones may be shared within households, data and airtime can be expensive, and network coverage is uneven.

Virtual Agronomist is designed with these realities in mind. The lead farmer model means one smartphone can support many farmers in a group. Trainings focus on basic digital skills and confidence, not only agronomy.

From the outset, we monitor who is using Virtual Agronomist and how it affects different groups. Early results show relatively balanced gender participation and broad age ranges among users, which is encouraging for a digital tool.

However, we know social norms, phone ownership patterns, and time constraints can still limit access for some groups. To address this, we work with partners to intentionally include women and youth in lead farmer roles, schedule trainings at convenient times, and create group-based usage patterns where one device can support multiple farmers. Where surveys, focus groups, or monitoring data show under-representation, we use that feedback to adapt outreach and training strategies.

In theory, fully digital systems scale quickly and cheaply. In practice, they often struggle with adoption, trust, and sustained use. Our approach is grounded in evidence that technology is most impactful when it supports, rather than substitutes for, human relationships and existing extension structures.

By equipping lead farmers and extension agents with tools like Virtual Agronomist, each person can serve more farmers with more consistent and data-informed advice while preserving face-to-face interactions that farmers value. This human-first, tech-enabled model recognizes that behavior change in farming is social and incremental: farmers watch what neighbors do, discuss new practices in groups, and test innovations on small areas before scaling up. Technology helps organize and personalize this process, but does not replace it.