From Pilots to Impact: Evidence on Scaling AI for Farmers in LMICs

Abstract

The panel examined the evidence base for AI‑enabled agricultural interventions across low‑ and middle‑income countries, drawing on a rapid systematic review commissioned by the UK FCDO, the RCC, and the University of Birmingham. The presenters highlighted the thinness of impact‑oriented evidence, pervasive gaps in trust, gender inclusion, data governance and policy frameworks, and the difficulty of moving from isolated pilots to scalable public services. Government, academic and industry perspectives were shared on how to embed AI responsibly, with particular focus on the Jammu & Kashmir context, model safety, usability, and mechanisms for feedback and accountability. The session concluded with audience questions that sharpened the discussion around data interoperability, incentives for last‑mile adoption, and concrete steps toward a scalable, inclusive AI ecosystem for smallholder farmers.

Detailed Summary

1. Framing the Challenge

1.1 Opening Remarks (Ms. Deepa Karthekeyan)

• Highlighted three overarching “value‑for‑money” questions: cost‑effectiveness, safety, and inclusion.
• Stated that the panel would explore evidence, practice and policy, and invited the evaluator pair (Zeba Siddiqui and Dr. Francis Rathinam) to set the stage.

1.2 Systematic Review Overview (Ms. Zeba Siddiqui)

• The review was a rapid, rigorously designed systematic mapping (2019‑2024) commissioned by the RCC / FCDO and the University of Birmingham.
• Scope: Identify AI‑enabled interventions for agricultural problems and the associated welfare outcomes for smallholder farmers in LMICs.

Evidence‑base statistics

Element	Figure
Peer‑reviewed journals screened	> 55
Academic & organisational repositories examined	~ 20
Grey‑literature items screened	> 100
Total studies initially identified	~ 450
Studies finally included (full‑text)	51

Core findings

• Technology focus: Predominantly crop‑production, pest detection, precision agriculture, and large‑language‑model prototypes.
• Evidence gap: Very few studies reported poverty‑reduction, livelihood‑decision‑making, or who holds decision‑making power.
• Impact durability: Most interventions remained pilot‑scale, with little evidence of sustained outcomes or diffusion beyond the initial project.
• Data & reporting deficiencies: A large proportion omitted critical details on training data provenance, ground‑truthing methods, and ethical considerations.

1.3 Key Themes Emerging from the Review (Dr. Francis Rathinam)

1. Trust deficit – Farmers’ hesitation stems from traditional agricultural practices and perceived financial risk.
2. Gender disparity – AI solutions often ignore women’s roles (e.g., low membership in cooperatives, limited smartphone ownership).
3. Human‑in‑the‑loop omission – Lack of local stakeholder participation in model design and validation.
4. Capacity‑building shortfall – Digital‑literacy deficits hinder adoption.
5. Governance vacuum – Absence of agri‑specific AI policy frameworks, consent mechanisms, and localized data standards.
6. Financial constraints – Scale economies favour larger farms; smallholders lack the capital to adopt AI services.

---

2. Deep‑Diving into Adoption Barriers

2.1 Trust & Behavioral Issues (Dr. Francis Rathinam)

• Generational resistance: Older farmers view AI as a risky financial gamble.
• Behavioural inertia: Adoption requires cultural alignment and perceived reliability of the technology.

2.2 Gender & Social Inclusion (Ms. Zeba Siddiqui)

• Cooperative composition: In a Kenyan case study, cooperatives (SACROs) were male‑dominant, marginalising women.
• Design bias: AI models trained on male‑centric datasets perpetuate existing inequalities.

2.3 Human‑in‑the‑Loop & Capacity Building (Ms. Zeba Siddiqui)

• Stressed that local champions (extension workers, community leaders) must be embedded in the AI development cycle.
• Noted that digital‑literacy training is essential to move beyond mere deployment to meaningful usage.

2.4 Governance & Policy Gaps (Dr. Francis Rathinam)

• Missing agronomy‑specific AI regulations; existing AI frameworks are generic.
• Data consent: Farmers are rarely asked for consent before their farm data is harvested for model training.
• Local data paucity: Reliance on national averages rather than hyper‑local datasets reduces model relevance.

---

3. Government Perspective – Scaling at the State Level

3.1 Dr. Piyush Singla (Secretary, IT, J&K) – “From Pilot to Public Service”

• Contextual heterogeneity: Jammu & Kashmir contains saffron fields, apple orchards, and rice paddies, each with distinct climatic and behavioural requirements.
• Trust as a scaling metric: Government defines “scale” as trust‑driven service continuity rather than a one‑off pilot.

Practical Illustrations

• Weather advisory latency: Farmers need real‑time, accurate weather forecasts for timely sowing; any model drift that delays advice results in lost yields.
• Centre of Excellence (CoE) model: A joint initiative between IIT Jammu and the state government to develop use‑cases that are replicable and up‑scalable, focusing on minimum latency and regional specificity.

3.2 Data Strategy – “Data Lake” Concept

• J&K is piloting a state‑wide data lake to harmonise datasets across departments (agriculture, meteorology, finance).
• Goal: Enable interoperable AI models that produce consistent answers irrespective of the departmental source.

3.3 Policy Feedback Loops (Dr. Singla)

• Cited the book Seeing Like a State to argue that static policy documents hinder learning.
• Emphasised the need for dynamic feedback mechanisms that capture frontline experiences and adjust policies in near real‑time.

---

4. Academic Insight – Trust, Usability & Model Safety

4.1 Prof. Kavya Dashora (IIT Delhi) – “Human‑Centred Trust & Usability”

• Trust pillars: Model safety, real‑world usability, impact measurement.
• Described three‑column framework being piloted in Andhra Pradesh:

Pillar	Core Questions
Model Safety	Is the model internally valid? What data was used for training? Are localized micro‑level datasets (soil, weather, pest incidence) incorporated?
Trust & Usability	Is the model tested with the actual end‑users? Does the interface speak the farmer’s language (including dialects, local terminologies)? Are extension workers included as human‑in‑the‑loop?
Impact	What intermediate outcomes (e.g., reduced pesticide residues, pest‑incidence decline) can be measured before expecting yield gains?

• Argues that micro‑level data (e.g., a 20 km radius) is essential to avoid model drift and hallucinations.

4.2 Model‑Safety Checklist (Prof. Dashora)

1. Data provenance verification – source, granularity, consent.
2. Stress‑testing across input variations – e.g., dialect‑specific query forms.
3. Field validation – side‑by‑side comparison of model advice vs. farmer’s traditional knowledge.

4.3 Trust‑Building via Extension Workers (Prof. Dashora)

• Extension workers act as trusted intermediaries; empowering them to override AI advice and feed corrections back into the model enhances accountability.

---

5. Industry & Practitioner Views

5.1 Dr. Sailendra (CEO, AgriGati) – “Bottom‑Up Accountability”

• Described AgriGati’s dynamic soil‑sensor‑to‑pesticide delivery pipeline, which raised farmer income 4‑ to 8‑fold in a pilot block.
• Highlighted data precision issues (government datasets at block level often lack the resolution needed for autonomous spraying).
• Emphasised ownership: Development was native‑place‑centric, ensuring accountability to the community.

5.2 CTO, Sardar AI – “Reward Mechanisms”

• Noted that incentives are crucial; Indian farmers respond to “what’s in it for me”.
• Suggested government‑driven reward schemes (e.g., subsidies tied to AI usage metrics) to stimulate adoption.

5.3 Guru Raj Mahajan (Tampere University) – “Access to Last‑Mile Farmers”

• Raised the question of scalable soil‑testing for marginal farmers in remote areas.
• Proposed leveraging embedded AI devices (e.g., low‑cost soil sensors) combined with WhatsApp‑based delivery to reach the most underserved.

5.4 Additional Practitioner (UT Austin) – “Learning from Banking & Health AI”

• Suggested an incremental, sector‑focused approach: first solve a few well‑defined problems (e.g., credit risk, disease detection) before expanding to whole‑crop portfolios.

---

6. Panel Synthesis – A Framework for Responsible Scaling

6.1 Three‑Tiered Evaluation Model (Synthesised by the Panel)

Tier	Focus	Typical Metrics
1️⃣ Model Safety	Internal validity, data quality, stress testing	Data provenance, geographic granularity, bias audits
2️⃣ Trust & Usability	Real‑world testing, human‑in‑the‑loop, cultural relevance	Farmer satisfaction, extension‑worker adoption rates, language coverage
3️⃣ Impact Assessment	Intermediate and long‑term outcomes	Pest‑incidence reduction, pesticide‑residue levels, income change, yield variance

• Decision‑gate: A model must clear Tier 1 before proceeding to Tier 2; only after demonstrating trust and usability may impact evaluations be funded for large‑scale roll‑out.

6.2 Policy Recommendations (Consensus)

1. Create an agri‑specific AI governance framework that mandates data consent, localization standards, and accountability mechanisms (e.g., liability for erroneous advice).
2. Institutionalise data lakes with interoperable formats to enable consistent AI outputs across ministries.
3. Formalise human‑in‑the‑loop processes: extension workers should have the authority to validate or override AI recommendations and feed corrections back into the model.
4. Design incentive schemes (subsidies, performance‑based payments) to align farmer motivations with AI adoption.
5. Standardise a reporting checklist for all AI pilots (training data, validation protocol, ethical review) to close the evidence gap identified in the systematic review.

6.3 Open Questions (Raised by the Audience)

• How to ensure gender‑inclusive data collection when many women lack smartphones?
• What minimal data granularity is required for reliable micro‑level modeling?
• How to design rapid feedback loops that capture farmer experiences and feed them back into policy within weeks rather than years?

---

7. Closing Remarks

• Ms. Deepa Karthekeyan thanked the panelists and audience, emphasizing that the discussion highlighted a “sweet spot” between context‑specific design and broader policy frameworks.
• The moderator noted the need for participatory, not hand‑over approaches, and hinted that a follow‑up systematic review in a year would likely show markedly different results if the identified gaps are addressed.

Key Takeaways

• Evidence Gap: The rapid systematic review found very few robust impact assessments of AI in agriculture; most works remain pilot‑scale and lack data on poverty‑reduction, livelihood changes, or decision‑making power.
• Trust Is Central: Farmers’ adoption hinges on trust, which is eroded by lack of local relevance, language barriers, and perceived financial risk.
• Gender Inequity: AI solutions often exclude women (both as data subjects and beneficiaries), reinforcing existing agricultural gender gaps.
• Human‑in‑the‑Loop: Embedding extension workers and local champions in the AI lifecycle improves both trust and model safety.
• Governance Void: No dedicated agri‑specific AI policy exists; missing data‑consent, localized data standards, and liability frameworks impede scaling.
• Data Interoperability: State‑wide data lakes are essential to avoid inconsistent AI outputs across departments.
• Three‑Tiered Scaling Framework: Successful scaling requires sequential validation of model safety → trust & usability → impact, with clear go‑/no‑go checkpoints.
• Incentive Mechanisms: Government‑driven reward schemes (e.g., subsidies linked to AI‑usage metrics) can drive farmer adoption.
• Micro‑Level Data: Hyper‑local datasets (soil, weather, pest incidence at ≤ 20 km radius) are crucial to prevent model drift and ensure relevance.
• Continuous Feedback: Policies must incorporate dynamic feedback loops that capture frontline experiences and enable rapid policy adjustments.

---

End of summary.

See Also:

• ai-in-governance-revolutionising-government-efficiency
• effective-ai-assessments-verification-and-assurance-establishing-the-foundations-for-responsible-confidence-in-ai
• ai-and-the-new-frontier-for-economic-progress-linking-innovation-to-inclusive-growth