AI in Food and Agriculture: Transforming Systems from Farm to Fork

Contents

Executive Summary

This panel discussion examines the critical gap between AI pilots in agriculture and scalable, sustainable implementation—particularly for smallholder farmers in low- and middle-income countries. The core argument is that effective agricultural AI requires three pillars: model safety tailored to local contexts, trust-building through inclusive design and last-mile workers, and rigorous impact measurement on intermediate and final outcomes. The discussion emphasizes that success depends less on technological sophistication and more on addressing data interoperability, governance accountability, and farmer-centric incentive structures.

Key Takeaways

Pilots Don't Scale—Systems Do: The "pilot-to-policy" model is fundamentally broken. Effective AI deployment requires bottom-up problem definition paired with top-down governance frameworks that embed accountability, feedback loops, and dynamic responsiveness—not one-time policy documents.
Data Interoperability Precedes AI: Before deploying AI models, governments must first solve data governance: standardized formats, cross-department sharing protocols, and localized datasets reflecting regional variation. This is unglamorous but foundational infrastructure work.
Last-Mile Workers Are Co-Designers, Not Implementers: Extension workers, cooperatives, and local facilitators should jointly develop AI models (not receive them). When they own the model and can override/correct outputs, they become trust-builders rather than technology gatekeepers—enabling farmer adoption.
Impact Must Be Measurable and Intermediate: Define what constitutes success at farmer level (not researcher level). Track achievable intermediate outcomes (pest detection accuracy, input cost reduction) that indicate progress toward ultimate outcomes (yield, income). Align measurement timelines with technology iteration cycles.
Inclusivity Requires Solving Access, Not Just Awareness: Gender gaps, smartphone access, digital literacy, and land ownership are not peripheral issues—they determine who benefits from AI. Embedded AI (like WhatsApp payments) in existing trusted systems beats standalone applications requiring behavioral change.

Key Topics Covered

Evidence gaps in agricultural AI: Findings from a rapid systematic review commissioned by FCDO (2019-2024)
Adoption barriers: Trust deficits, behavioral change resistance, digital literacy, gender inequality, and asset ownership disparities
Data infrastructure challenges: Lack of localized datasets, data interoperability, and multimodal data integration across government departments
Scaling from pilots to public infrastructure: The failure of pilot-to-policy models and the need for systems-level thinking
Governance and accountability: Liability frameworks, data consent, ethical AI deployment, and last-mile responsibility
Embedding AI into existing systems: Aligning public digital infrastructure (DPI) with farmer needs rather than top-down technology deployment
Gender and inclusion: Systemic exclusion of women from asset ownership, decision-making, and technology design
Feedback loops and dynamic accountability: Moving beyond static policies to responsive, learning systems

Key Points & Insights

Pilot Sustainability Crisis: Over 51 analyzed studies show AI models in agriculture remain at pilot stage with minimal sustained adoption. Most supply-side AI designs fail post-pilot due to lack of trust, behavioral change resistance, and absence of ongoing support systems.
Trust as Primary Currency: Adoption is fundamentally a trust problem, not a technology problem. Farmers require human validation, accountability structures, and transparent decision-making processes. Extension workers and last-mile facilitators are critical trust bridges—not obstacles.
Multimodal Data Problem: Agriculture is uniquely complex because outcomes depend on soil type, weather, crop variety, cultural practices, region, and household-level decision-making. Current models trained on limited localized data fail to capture this variance; generic national datasets produce irrelevant outputs.
Language & Jargon Fragmentation: Unlike healthcare (where "typhoid" is universal), agriculture uses hyper-local terminology, dialects, and household-specific language for crops, pests, and practices. Models cannot work effectively without accounting for this linguistic diversity—a prerequisite before AI scaling.
Exclusion by Design: AI interventions often exclude the majority of intended users. Women are frequently locked out of asset ownership, smartphones, land rights, and formal financial systems. Cooperatives and advisory networks in some contexts show male-dominated membership, reducing female farmer participation despite women performing majority of agricultural labor.
Data Interoperability as Foundation: Government departments hold rich agricultural data but cannot "talk to each other" due to different formats and structures. Before scaling AI, Jammu & Kashmir and other regions must build data lakes and shared governance frameworks enabling integrated queries across departments.
Accountability Void: No frameworks exist to hold AI systems accountable when advisories fail. Farmers still bear full risk while technology providers avoid liability. Insurance products, transparent disclaimers, and legal responsibility must precede widespread deployment.
Three-Pillar Evaluation Framework: Scaling requires verification across (1) model safety: Is it tested with real users in real contexts, not just lab-validated? (2) trust & usability: Is it culturally and contextually relevant? Do last-mile workers own the model? (3) impact measurement: What intermediate outcomes (pest reduction, residue levels) and ultimate outcomes (yield, income) result?
Intermediate vs. Ultimate Outcomes: Model versions change every 4 months; agricultural outcomes take seasons. Tracking intermediate outcomes (pest incidents, pesticide residue) is measurable and actionable, allowing confidence in eventual yield/income impacts without waiting years for results.
Incentive Misalignment: Farmers in India respond to "what's in it for me" structures (evidenced by 98% WhatsApp adoption when incentivized). Current AI deployment lacks clear, farmer-facing reward mechanisms—whether financial, convenience-based, or social.

Notable Quotes or Statements

"We don't know if AI in agriculture has impact yet. Where there is some evidence, the impact hasn't lasted or sustained. It's very temporal, very projectized." — Deepa Menon (summarizing findings)

"Unless any AI model speaks the language of a farmer, upscalability will always fail. In government, when we talk of scale, we talk of trust. It's a service, not a small pilot you run in one district that fades off in a year." — Dr. Singla, Secretary Information Technology, Jammu & Kashmir

"Agriculture is a multimodal sector. You can't just look at a picture and say 'this is late blight of potato.' You need to know the month, variety, soil condition, weather, and state it's from. But we're developing models without collecting this multimodal data." — Prof. Cavia, IIT Delhi

"There is a stark difference: in healthcare, the jargon is consistent worldwide (typhoid is typhoid). In agriculture, the jargon sits with the farmer and is extremely diverse—language, dialect, household-level words for the same thing. This makes data pipelines critically important." — Prof. Cavia

"What if the AI system goes wrong? Even when the farmer stands in the field with traditional knowledge and experienced people, things go wrong—and there's no one to hold accountable. Yet we expect the farmer to trust a device over all his wisdom and experience." — Prof. Cavia

"Pilot is a controlled experiment. You hand-hold a small group and frame everything predictably. Policy must apply to the whole population. This is where the disconnect happens because we don't have multimodal data." — Prof. Cavia

"The number of studies that didn't even report what type of data they used for training, ground truthing, or ethical considerations was shocking. This calls for sector-wide reporting standards." — Zeba Sadiki, Athena Infonomics (on systematic review findings)

"We increased farmer income 4 to 8 times by converting information into income—not just knowledge. We built dynamic models with soil sensors and daily updates, not static government stacks." — Dr. Silender, CEO Agri Gati (audience comment)

Speakers & Organizations Mentioned

Panel Members:

Deepa Menon — Moderator; researcher/policy practitioner at intersection of research, policy, and practice
Zeba Sadiki — Evaluator, Athena Infonomics; led FCDO-commissioned rapid systematic review
Francis — Evaluator, Athena Infonomics; working on AI evaluation frameworks and horizon mapping
Dr. Singla — Secretary, Information Technology, Government of Jammu & Kashmir (public sector perspective)
Prof. Cavia — IIT Delhi; leading work on adoption, trust infrastructure, and human-AI interfaces

Funding/Commissioning Organizations:

FCDO (Foreign, Commonwealth & Development Office, UK Government)
University of Birmingham
International Initiative for Impact Evaluation (3ie)
RCC (unclear from transcript)

Institutions/Initiatives:

Athena Infonomics — Research organization conducting systematic reviews and case studies
IIT Delhi (Indian Institute of Technology)
IIT Jammu — Collaborating on center of excellence for AI use cases
Jammu & Kashmir Government — Piloting data lake concept and AI governance frameworks

Private Sector/NGO Representatives (Audience):

Agri Gati (Dr. Silender, CEO) — AgriTech company using DPI/UPI from Government of India; deployed in Pune
Sardar AI — AI readiness/training company; trained 10 million people globally
Tampere University Finland — Smart Agriculture & Entrepreneurship Program; 14 years working with agricultural universities in India
Various practitioners and coordinators from agriculture tech sector

Technical Concepts & Resources

Methodologies & Frameworks:

Rapid Systematic Review (mixed-methods): Rigorous quantitative review embedded with qualitative/narrative methodology to capture nuanced impact on smallholder farmers
Human-in-the-Loop Design: Last-mile workers/extension agents co-design models, validate outputs, can override AI recommendations, provide feedback for model improvement
Three-Pillar Evaluation Architecture (from Athena):
1. Model Safety (internal validity): Data source, localized training, consistency testing
2. Trust & Usability: Real-world testing, cultural/contextual relevance, ethical alignment
3. Impact Measurement: Intermediate outcomes (pest reduction, residue) and ultimate outcomes (yield, income)
Data Lake Concept: Centralized repository where multiple government departments contribute standardized data; enables unified AI queries across departments

Technical Problems Identified:

Model Drift: Models degrade when real-world conditions diverge from training data
Model Hallucinations: Common term for incorrect/misleading outputs
Data Interoperability: Same query produces different results when queried across different departmental datasets due to format/structure differences
Stress Testing: Testing models with varied input types/formats (e.g., diverse language/dialect expressions)

Data/Infrastructure Gaps:

Localized Datasets: Most models trained on generic national-level data; requires hyper-local (20km variance) soil, weather, pest data
Ground Truthing: Lack of documented validation data; insufficient reporting in literature
Consent & Data Governance: Absent mechanisms for farmer consent on data collection/usage
Digital Literacy: Barrier to smartphone/device-based AI access
Bandwidth Access: Uneven internet connectivity in rural areas

Case Studies & Evidence Base:

Systematic review scope: 55 journals, 19-20 repositories (academic + organizational), 100+ gray literature sources, 2019-2024
Studies analyzed: 450 initially screened; 51 included in final review
Geographic focus: Lower-middle-income countries (focus on Global South), with detailed case studies from Kenya and India (Andhra Pradesh, Jammu & Kashmir)
Crops/domains examined: Crop production, pest detection, precision agriculture, large language models, financial services advisory

Policy/Governance References:

PMKAAN (Pradhan Mantri Kisan Annadata Aay Sanrakshan Yojana) — Government of India agricultural scheme
DPI (Digital Public Infrastructure) — Government of India framework being adopted for agriculture
UPI (Unified Payments Interface) — Used as model for embedded, accessible financial infrastructure

Comparative Sector Lessons:

Banking sector: Focused on solving one problem at scale (fraud detection) before expanding; created ecosystem/confidence that enabled subsequent solutions
Healthcare: Narrowly focused (breast cancer diagnosis) before broad deployment; created trust infrastructure
Financial services: Embedded systems (WhatsApp Pay, Google Pay) achieve 98% adoption by integrating into existing trusted channels rather than requiring new apps

Gaps & Open Questions (Implied from Discussion)

What accountability and insurance structures should cover failed AI advisories?
How can government data lakes ensure data quality and prevent garbage-in-garbage-out scenarios?
What specific reward/incentive mechanisms will drive farmer adoption across diverse contexts?
How can AI systems be designed to override by last-mile workers without degrading model learning?
What is the optimal balance between standardized governance frameworks and hyper-local customization?
How can intermediate outcome measurement be standardized across different crops/regions?

Document Type: Conference Panel Discussion (AI Summit)
Duration: ~55 minutes (full session)
Key Theme: AI deployment scalability, governance, impact measurement, and inclusive design in agriculture