AI in Climate and Agriculture: Building Resilient Communities

Detailed Summary

(Delivered by Maulik Jagnani, with reference to co‑author Rohini Pandey)

1.1 Context & Motivation

Escalating flood risk: Global flood exposure has quadrupled since the 1980s; ≈ 1 in 4 people worldwide now live in flood‑prone areas.
India’s exposure: ~390 million people at risk; Bihar is the most flood‑prone Indian state (≈ 18 % of the nation’s flood‑prone area).
Historical trend (maps 1984‑1995, 1996‑2008, 2009‑2021): Both the spatial extent and frequency of flooding have increased dramatically, as shown by deepening blue shades on the maps.

1.2 Impact of Recent Floods (2019 data)

> 50 % of surveyed households reported reduced harvests and increased sickness.
≈ 25 % reported damage to houses, livestock, and goods.

1.3 Existing Early‑Warning Systems (Pre‑AI)

Central Water Commission (CWC): Manual gauge readings → radio → district bulletins (fax, press, radio, TV).
Limitations: No “last‑mile” delivery, heavy reliance on indigenous knowledge and fragmented information; low reach in rural Bihar.

1.4 AI‑Powered Early‑Warning Systems

Government AI system: Hourly water‑level data from thousands of gauges → real‑time modelling → 2‑5 day forecasts with > 95 % accuracy (scatter plots close to the 45° line).
Google AI system: Combines CWC water‑level data with digital elevation models → produces hyper‑local flood‑risk maps and targeted alerts.

1.5 The Core Problem: Delivery Gap

Survey evidence (2019‑2025):
- < 20 % of households receive any CWC or Google alerts, even when forecasts are generated.
- Gap persists across years: 5 % (2022) → 18 % (2024‑25) for CWC alerts; similar low rates for Google alerts.
Consequences: Alerts exist but do not reach the people who need them; a large “forecast‑generation → no‑receipt” gap.

1.6 Last‑Mile Delivery Experiments

Model	Description	Outcome
Panchayat Mukhiya Model (2019)	Alerts sent via WhatsApp/SMS to elected village leaders (Mukhiya) who were expected to disseminate within the community.	No impact – treatment and control villages reported similar alert‑receipt rates; Mukhiya largely ignored the alerts.
Community‑Agent Model (2022‑2025)	2‑3 trained local agents per Panchayat; use loudspeakers, flags, SMS, WhatsApp to broadcast alerts throughout the flood season (June‑Oct).	Positive impact – treatment households 14‑pp higher alert‑receipt (e.g., 64 % vs 50 % in 2025). More alerts per household (≈ 15 vs 5), earlier receipt (60 % before water arrival vs 45 %).

Willingness to Pay: In a high‑stakes elicitation, ≈ 60 % of unexposed households would pay ~ ₹66 per season for a delivery service (SMS/WhatsApp). Indicates a perceived delivery gap.

1.7 Broader Impacts of Effective Delivery

Reduced false negatives: Treatment households far less likely to miss an alert when flood occurs (10 % vs 28 % in 2025).
Increased false positives (alerts without flood): Treatment households report higher rates (12 % vs 4 % in 2025) – a trust trade‑off.
Perception & Trust: Treatment groups view alerts as more accurate (45 % vs 34 % perceiving high accuracy) and more trustworthy.
Adaptive Actions (2025):
- + 6 % stock food before floods.
- + 13 % secure drinking‑water supply.
- + 22 pp use sandbags.
- + 8 % take health‑precaution measures.
Spill‑over to Science Attitudes: Treatment respondents show higher trust in science/technology (up‑shift of ≈ 6 pp), suggesting that reliable AI alerts can reshape broader perceptions.

1.8 Concluding Remarks (Presenter)

AI forecasts are highly accurate, timely, and information‑rich, but impact hinges on last‑mile delivery.
Community agents provide a scalable delivery channel that leverages existing government structures.
Effective delivery not only improves immediate flood resilience but also boosts trust in scientific innovations.

2. Panel Discussion – “Impact Evaluation of AI for Climate & Agriculture”

The moderator introduced the panelists and opened the floor to thematic questions. The discussion revolved around four core strands: (a) investment and policy drivers, (b) evidence‑based programming, (c) barriers to adoption & scaling, and (d) data ownership & integration with indigenous knowledge.

2.1 Panelist Contributions

Panelist	Key Points Raised
Veena Srinivasan (Well Labs)	- UAE has a national AI strategy (Minister of State for AI, AI‑focused university, AI data centre). - Aim for Scale initiative (with Nobel laureate Michael Kremer) launches evidence‑based, scalable AI tools for agriculture (AI weather forecasts, digital agriculture services, AI livestock). - Emphasises government‑partnered scaling to reach millions.
George Richards (Community Jameel)	- Community Jameel is evidence‑centric, partnering with J‑PAL and AI‑focused donors. - Highlights the Agricultural Technologies Adoption Initiative (AT‑I) and stresses that trust and last‑mile delivery (as shown by Malik’s flood‑EWS study) remain pivotal. - Calls for rigorous impact evaluation even in “hard‑to‑reach” contexts.
Neerik Shah Shetty (Precision Development)	- Focuses on institutionalising AI‑driven advisory services (weather forecasts, pest predictions) through government channels to reach 45 million farmers. - AI is a vehicle for scaling existing evidence‑based interventions; three levers: (i) richer content, (ii) better targeting via behavioural data, (iii) faster, more engaging delivery (e.g., Gen‑AI voice advisories).
Fatima Al Mullah (UAE Presidential Court)	- UAE invests $1 bn in AI for Africa and 5‑GW AI data centre. - Discusses AI‑ecosystem for global agriculture in partnership with the Gates Foundation. - Stresses data ownership, the need to co‑design with farmers, and the importance of open‑source AI tools to foster trust.
Moderator (unspecified)	- Synthesised themes: trust, human‑in‑the‑loop, speed vs. evidence generation, scaling barriers, open‑data culture, and integration of indigenous knowledge. - Fielded audience questions on adoption barriers for subsistence farmers and data‑sharing ecosystems.

2.2 Thematic Highlights

Trust as the linchpin – All panelists agreed that technology adoption hinges on trust—built via local intermediaries, transparent evidence, and co‑design.
Evidence‑based scaling – Emphasis on randomised evaluations (e.g., Malik’s RCT, AT‑I) to inform policy decisions and government procurement.
Last‑mile delivery models – The community‑agent model (Malik) is cited as a successful, scalable approach; contrastingly, Panchayat‑Mukhiya model failed due to low engagement.
Open data & ownership – Fatima highlighted the need for farmer‑centric data governance and open‑source AI to reduce mistrust and enable broader adoption.
Barriers to scaling –
- Heterogeneous contexts: “One size does not fit all”; solutions must be context‑specific.
- False positives vs. false negatives: Trade‑offs affect trust; systematic communication of risk is essential.
- Institutional inertia: Need for government‑backed institutionalisation (e.g., Aim for Scale, Precision Development’s government partnerships).
Integration of Indigenous Knowledge – Panelists stressed that AI outputs should anchor on local priors (e.g., religious calendars, traditional flood signs) to avoid “black‑box” perceptions.

2.3 Audience Q&A – Representative Questions & Panelist Responses

Question	Summary of Responses
Why will subsistence farmers adopt AI tools when they’re barely surviving?	Neerik: AI offers risk‑mitigation information (e.g., advance rain forecasts, flood warnings) that can protect livelihoods; adoption is about enhancing existing decision‑making rather than replacing it.
How can we achieve open‑data sharing across governments, private sector, and civil society?	Fatima: Emphasised data ownership and co‑design; proposed open‑source AI platforms and policy frameworks that grant farmers control over their data while enabling aggregation for model improvement.
What are the biggest scaling bottlenecks for AI‑driven climate services?	Veena: Lack of institutional pathways to embed evidence‑based tools in development plans; need for government‑level buy‑in and budgetary allocations. George: Trust and last‑mile delivery remain the core constraints; scaling requires trusted intermediaries.
How do we manage the risk of false positives eroding trust?	Malik (presenter) (referenced): Highlighted that treatment groups still perceived higher accuracy despite higher false‑positive rates; transparent communication and continuous evidence collection can mitigate trust loss.
How can AI respect indigenous knowledge while delivering hyper‑local forecasts?	Fatima & George: AI systems should be co‑developed with communities, integrating local indicators (e.g., water colour, traditional warnings) as features or validation checks, ensuring a feedback loop between AI outputs and farmer experience.

2.4 Closing Reflections (Each Panelist’s “Hope” & “Concern”)

Panelist	Hope	Concern
Veena	Collaborative, government‑driven AI ecosystems can deliver massive scale (tens of millions).	Risk that AI tools fall into the “wrong hands” or become mis‑used without proper partnership safeguards.
George	Centering people (farmers, vulnerable communities) in AI design; expanding to post‑conflict agricultural recovery.	Need to push further to reach most vulnerable and ensure AI works in fragile, conflict‑affected settings.
Neerik	Rapid convergence of innovators, institutions, and NGOs to scale AI‑driven advisory services.	Inclusion gaps – who is left behind? Also, accountability when advice leads to adverse outcomes.
Fatima	Belief in human goodwill and ability to co‑create beautiful AI solutions.	Potential mis‑allocation of AI benefits to undesired actors; safeguarding against misuse.
Malik (presenter)	Demonstrated that simple, incentivised community agents can dramatically improve alert receipt and trust.	Sustaining trust over the long term, especially when false positives increase; need robust risk communication.

Key Takeaways

AI flood forecasts in Bihar are > 95 % accurate and can predict events 2‑5 days in advance, but without effective last‑mile delivery they have negligible impact.
A community‑agent delivery model (trained local agents using loudspeakers, flags, SMS/WhatsApp) significantly improves alert receipt, timeliness, and trust; it also leads to material adaptive actions (sandbags, food stockpiling, water security).
Trust is the decisive factor for adoption of AI‑based climate services; evidence‑generation, co‑design, and transparent communication are essential.
False positives are inevitable, but perceived accuracy and trust can still increase if delivery is reliable and risk communication is clear.
Willingness‑to‑pay surveys show a clear demand for a delivery service, confirming the delivery gap exists even when alerts are technically available.
Policy & scaling insights:
- National AI strategies (e.g., UAE) and government‑partnered programmes (Aim for Scale, Precision Development) are crucial for reaching millions.
- Institutionalising AI tools through development banks and ministries accelerates scale.
- Open‑source, farmer‑owned data and integration of indigenous knowledge foster trust and relevance.
Broader societal impact: Exposure to reliable AI alerts can increase trust in science and technology, suggesting AI interventions can have spill‑over effects beyond immediate climate resilience.

The above summary captures the substantive content of the session, attributing key statements to the respective speakers, surfacing data, insights, recommendations, and the panel’s synthesis of challenges and opportunities for AI‑driven climate and agricultural interventions.

See Also:

India AI Impact Summit 2026

Explorer

ai-in-climate-and-agriculture-building-resilient-communities