Grounded Intelligence: Why Most AI Innovations Never Reach the Field

David Bergvinson

Most AI in agriculture gets built for the wrong people. As new technologies emerge at a rapid pace, many never move beyond pilot programs or demonstration projects. So what separates innovations that generate headlines from those that create real impact in the field? Grounded Intelligence is a new podcast from AGX AI hosted by David Bergvinson. Through candid conversations with researchers, founders, farmers, and funders, the series explores what it really takes to transform promising ideas into practical solutions.

Episodes

  1. 18h ago

    Participants, Not Data Sources: Building Farmer Data Systems That Give Back

    Thompson Ogunsanmi is a scaling specialist at the International Institute of Tropical Agriculture, and Ifeoma Umunna is director of agriculture at DevAfric. They reveal how AkiliMo built a 700,000-farmer database across three countries with 70% of data flowing through feature phones rather than smartphones; how national partner associations now control data flow back to IITA instead of the reverse; why replicating India's AgriStack model fails in African contexts where women lack land tenure and farming is nomadic; and why treating farmer data as reusable infrastructure rather than a disposable project input is the only path to returning tangible value to smallholder farmers. Thompson Ogunsanmi, Scaling Specialist at the International Institute of Tropical Agriculture, has spent seven years coordinating the scaling of AkiliMo, a digital agronomic platform, across Nigeria, Ghana, and Tanzania. His team built a database of more than 700,000 registered farmers and, after discovering that 70% of users rely on feature phones rather than smartphones, developed parallel delivery channels including SMS, USSD, IVR, and printable guides to reach farmers regardless of device type. In each country, partner organizations formed independent associations — Akilimo Nigeria Association, Akilimo Ghana Association, and Akilimo Tanzania Association — that now control data flow, with IITA receiving data from them rather than the reverse. His approach to farmer data governance centers on designing systems where national partners and farmer organizations hold decision-making authority over how data moves, treating interoperability and local control as prerequisites rather than afterthoughts. Ify Umunna, Director of Agriculture at DevAfric, has worked with farmers across multiple organizations over more than a decade, including roles as a gender specialist at Sahel Consulting, at Nourishing Africa, and at AFEX, where she led strategy working with more than a million farmers in Nigeria. Her career has moved across institutional design, gender inclusion, and agricultural market systems, building a perspective that spans how data infrastructure intersects with the structural realities farmers face. She approaches farmer data governance through the lens of gender and land tenure, identifying that replicating frameworks like India's AgriStack in African contexts would exclude women and nomadic farmers if digital identities remain tied to land ownership. Thompson explains: ◼️ Why seventy percent of farmer data submissions come from feature phones ◼️ How national partner associations rebuilt data flow away from IITA ◼️ What happens to farmer data when a project ends and funding disappears ◼️ Why replicating AgriStack in Africa fails women and nomadic farmers ◼️ How AkiliMo scales impact across three countries using one shared platform ◼️ What makes stewardship a better framework than ownership for farmer data ◼️ Why treating farmers as businesses changes how you design data systems ◼️ How SMS and IVR channels reach farmers that apps never will Ify explains: ◼️ Why seventy percent of farmer data submissions come from feature phones ◼️ How national partner associations rebuilt data flow away from IITA ◼️ What happens to farmer data when a project ends and funding disappears ◼️ Why replicating AgriStack in Africa fails women and nomadic farmers ◼️ How AkiliMo scales impact across three countries using one shared platform ◼️ Why treating farmers as businesses changes how you design data systems ◼️ How SMS and IVR channels reach farmers that apps never will To read the discussion papers, click the link below. https://agx.community/agx-ai/discussion-papers/

    Participants, Not Data Sources: Building Farmer Data Systems That Give Back
  2. Sep 3

    Beyond Translation: Why Localization Requires Trust, Governance, and People

    In this episode, we at AGX AI explore why localization for AI-based agricultural advisory services is not just a matter of translation or model adaptation. Audumbar Chavan, Prasaanth Balraj, and Dr. Harsh Vats of Athena Infonomics draw from implementation work with Indian state governments to show why state ownership, recurring budgets, human intermediaries, and accountability mechanisms determine whether smallholder farmers can trust and use these systems. Audumbar Chavan is the AI and DPI services practice lead for agriculture AI at Athena Infonomics, an AGX AI partner organization. He has conducted field assessments of AI advisory pilots in India, including a March 2026 visit to Baramati, Maharashtra, where he documented per-farmer costs reaching $100, and has co-authored a discussion paper on localization through the AGX AI community. He approaches AI localization as a governance commitment requiring state ownership, recurring budgets, and expert-verified labeling rather than a one-time technical deployment. Prasaanth Balraj is the associate director of AI and digital transformation at Athena Infonomics, having joined the organization two weeks before this recording. He previously worked at Wadhwani Institute of Artificial Intelligence Global, where he was involved in taking AI solutions into Africa and has identified strategies such as reusing open-source datasets, including pest identification data from Ethiopia applied to Indian contexts, to reduce localization costs. He frames localization around cost reduction and cross-context data reuse, treating adaptability as a design constraint shaped by institutional and resource realities. Dr. Harsh Vats is a program manager at Athena Infonomics with half a decade of experience serving as a liaison officer with multiple Indian ministries to deploy AI solutions. He has worked across healthcare, education, and agriculture, advising sub-national governments including Andhra Pradesh on integrating governance and accountability frameworks into AI deployments. He approaches AI implementation through the institutional layer, designing systems where state governments own accountability for model accuracy, human intermediaries remain embedded in the advisory chain, and recourse mechanisms exist when AI-generated advice fails. Audumbar explains: ◼️ What a $100-per-farmer pilot in Baramati revealed about AI cost realities ◼️ Why farmers in India trust extension workers over agricultural apps ◼️ How Andhra Pradesh built a government accountability layer for AI advice ◼️ What happens to food security when AI-generated crop guidance fails ◼️ Why proportional sampling could replace expensive full model retraining ◼️ How localization for women and tenant farmers changes system design ◼️ What reaching farmers without smartphones or internet actually requires ◼️ Why treating AI advisory as a public good demands recurring state budgets

    Beyond Translation: Why Localization Requires Trust, Governance, and People
  3. Aug 13

    The Benchmark Gap: Why Standard AI Metrics Fail Smallholder Agriculture

    F1 scores weren't built for farms with cottages, livestock, and bore wells on the same parcel. Praveen Pankajakshan and Deepa Karthykeyan (Athena Infonomics) join us live for real-time Q&A and first access to new insights rethinking AI benchmarking for smallholder agriculture. Praveen explains: ◼️ Why do F1 and BLEU scores miss what actually matters on smallholder farms ◼️ What happens when a crop identification model meets cottages, roads, and livestock in one parcel ◼️ How can agriculture borrow a gold standard from the health sector ◼️ Why does no one fund the localization smallholder AI actually requires ◼️ What metric could replace F1 once trust and equity enter the equation ◼️ How do you benchmark a model after deployment not just before ◼️ Why does building AI capacity in Indian states matter for agricultural benchmarking ◼️ What happens when human evaluators disagree on the same model output 00:00:05 | Make benchmarking matter for advisory tools 00:04:31 | Test models against farming diversity 00:08:00 | Move beyond controlled model certification 00:11:12 | Rethink benchmarking after real deployment 00:12:50 | Challenge crop models on messy parcels 00:17:09 | Ask who funds true localization 00:19:46 | Compare outputs across multiple evaluators 00:23:42 | Look for agriculture's gold standard 00:25:21 | Watch governments build AI capacity 00:30:54 | Invent metrics beyond F1 scores To read the discussion papers, click the link below. https://agx.community/agx-ai/discussion-papers/

    The Benchmark Gap: Why Standard AI Metrics Fail Smallholder Agriculture
  4. Jul 30

    Models: Why Translating the Interface Isn't Localization

    Femi Royal, Senior Advisor (governance and systems), Malcolm Durosaye, Consultant, and Ahmad Raji, Associate Consultant, are co-authors of the AGX AI discussion paper on localized Agri-LLMs for small-scale producers in Africa and India. The panel unpacks how frontier AI models fail without local crop, soil, and market data; how India's AgriStack and Bhashini contrast with Africa's fragmented data systems; how tools like PlantVillage Nuru and FarmerChat are reaching farmers through voice and SMS in low-connectivity environments; and why donor-funded agricultural AI collapses without blended public-private business models. Femi Royal is a Senior Advisor at Dev Afrique Development Advisors, a partner organization in the AGX AI initiative focused on responsible AI for smallholder farmers. He co-authored the AGX AI discussion paper "Localized Agri LLM: Exploring Low Power, Low Cost Models for Small-Scale Producers in Africa and India," which examines why frontier AI models trained on North American and European data fail to serve African and Asian agricultural contexts. Royal's work centers on governance and systems-level questions—how multi-stakeholder frameworks, digital public infrastructure, and policy alignment shape whether AI tools reach small-scale producers or remain donor-dependent prototypes. Malcolm Durosaye is a Consultant at Dev Afrique Development Advisors and co-author of the AGX AI discussion paper on localized Agri-LLMs for small-scale producers in Africa and India. His research for the paper examined the gap between language translation and true contextual localization, mapping how local crop, soil, weather, and market data determine whether AI-generated agricultural advice is relevant or misleading. Durosaye approaches localization as a data and context problem rather than a language problem, drawing contrasts between India's digital public infrastructure investments and Africa's fragmented agricultural data systems. Ahmad Raji is an Associate Consultant at Dev Afrique Development Advisors and co-author of the AGX AI discussion paper exploring low-power, low-cost AI models for small-scale producers in Africa and India. His contributions to the paper addressed how accessibility barriers—unreliable internet, lack of smartphones, and low literacy—determine whether even well-built AI models can reach the farmers they are designed to serve. Raji's work focuses on the infrastructure and delivery side of agricultural AI, examining how voice-based and SMS systems, telco partnerships, and blended public-private business models can sustain AI advisory tools beyond initial grant funding. Femi, Malcolm, and Ahmad explain: ◼️ Why translating a frontier model into local languages still fails smallholder farmers ◼️ What India's AgriStack reveals about Africa's missing digital infrastructure ◼️ How lead farmers and extension agents become the real AI delivery channel ◼️ Why promising agricultural AI tools collapse when donor funding ends ◼️ What hallucination and liability risks look like in farm advisory contexts ◼️ How African innovators already deploy voice and SMS tools in low-connectivity areas ◼️ Why telcos may shape agricultural AI the way M-PESA shaped mobile money ◼️ What putting farmers in control of their own data actually requires To read the discussion papers, click the link below. https://agx.community/agx-ai/discussion-papers/

    Models: Why Translating the Interface Isn't Localization
  5. Jun 17

    A Systems Approach to Agricultural AI: Stewart Collis on Data, Trust, and Farmer Access

    Delivering effective AI advisory services to smallholder farmers depends less on model sophistication than on building the underlying ecosystem—shared data infrastructure, consortium economics, localized digital public goods, and farmer trust—without which even the most capable AI repeats the scaling failures of the previous generation of digital agriculture. Stewart Collis, Senior Program Officer for Digital Solutions in Agricultural Development at the Gates Foundation, has spent over 25 years building and evaluating digital advisory systems for smallholder farmers — from co-founding AWARE's weather advisory service in emerging markets and advancing crop modeling at Texas A&M, to leading digital agriculture strategy at ICRAF before joining the foundation six years ago. At the Gates Foundation, he directs investments across the full advisory ecosystem — including the Institute for Agriculture and AI at Mohammed bin Zayed University, the CGIAR's Fairgrounds federated data-sharing infrastructure, and consortium-led public-private partnerships in Nigeria and India that have driven per-farmer service costs to measurable benchmarks such as 18 cents per farmer per year in Odisha, where farmers like Priya Sharma report that "the timely weather alerts helped me save my groundnut crop during unexpected rains." His approach to AI for smallholder agriculture is rooted in a systems lens — data infrastructure, model localization, delivery economics, farmer trust, and policy — because two decades of digital agriculture have demonstrated that technology reaches farmers only when the people, process, and institutional foundations are deliberately built first, as evidenced by smallholder farmer feedback from consortium partnerships indicating that "we trust the advisories because our local extension agents explain them in our language." Stewart explains: ◼️ Why do persistent infrastructure gaps—farmer registries, localized soil maps, granular weather forecasts—prevent AI advisory services from reaching smallholders at scale, and what consortium approaches are addressing the reality that no single organization can build these digital public rails alone? ◼️ How did the Gates Foundation pivot its entire digital agriculture strategy when ChatGPT launched, forcing abandonment of multi-year funding cycles—and what does that reveal about the human capacity and institutional readiness required before AI tools can function effectively? ◼️ What does Odisha, India's achievement of 18-cent-per-farmer digital advisory services for seven million farmers tell us about the institutional coordination and process standardization required to make AI-powered advice economically viable? ◼️ Why is the Gates Foundation building consortium models with private-sector partners like Indorama, OCP, and Flour Mills of Nigeria rather than funding standalone technology deployments—and what human systems and institutional partnerships must exist before digital tools can support farmer decisions? ◼️ What is the CGIAR's Fairgrounds project, and how does its federated data-sharing approach address the fundamental challenge that effective AI advisory requires both technical infrastructure and trusted local institutions to facilitate farmer adoption? ◼️ Why does farmer trust—built through consistent local extension networks and community validation processes—remain the non-negotiable design constraint for AI advisory services, and what institutional relationships must be established before technology deployment? To read the discussion papers, click the link below. https://agx.community/agx-ai/discussion-papers/

    A Systems Approach to Agricultural AI:  Stewart Collis on Data, Trust, and Farmer Access

Ratings & Reviews

5
out of 5
3 Ratings

About

Most AI in agriculture gets built for the wrong people. As new technologies emerge at a rapid pace, many never move beyond pilot programs or demonstration projects. So what separates innovations that generate headlines from those that create real impact in the field? Grounded Intelligence is a new podcast from AGX AI hosted by David Bergvinson. Through candid conversations with researchers, founders, farmers, and funders, the series explores what it really takes to transform promising ideas into practical solutions.