AI diffusion across India’s farms will hinge on earning the trust of farmers as much as infrastructure

The challenges were remarkable. The system must be able to understand farmers speaking Gujarati and local dialects, work amid intermittent connectivity in rural areas and give advice they can stake their livelihoods on. It draws from over 50 years of verified cooperative data, including 2 billion milk procurement transactions annually, veterinary treatment records of 30 million cattle and farmer-wise cattle census data.

Maharashtra’s Vistaar agricultural advisory system took nine months from commitment to deployment in 2025. The state wanted AI-powered advice for farmers in Marathi and local dialects, accessible via basic phones and available even when connectivity drops during the monsoon season. Ethiopia’s OpenAgriNet took three months to deploy earlier this year, addressing the same core challenge of agricultural advice at scale, but with a pathway already mapped by Maha Vistaar.

Amul’s Sarlaben took just three weeks. This time compression happened because deployment know-how became transferable. The technical architecture, governance frameworks, evaluation protocols and deployment playbooks that took nine months to build for a pioneer like Maharashtra could be re-used and adapted rather than rebuilt from scratch.

Most AI pilots succeed in controlled environments with clean data, engaged users and vendor support. Then organizations try to scale them and encounter systematic failures that pilots did not reveal. A farming advisory could work perfectly with 500 farmers during a pilot test. But when it expands to cover 50,000, farmers find it hard to connect during the rainy season when they need advice the most. The system struggles with dialects and recommendations often contradict local agricultural universities. So farmers quit using it.

These failures are visible in India because the scale required leaves no room for workarounds. A service in 22 official languages that must account for seasonality and intermittent electricity means every infrastructure gap becomes apparent. A Marathi-speaking farmer using a chatbot on a basic feature phone represents the actual AI frontier, far from the controlled demo environment where all of it works.

India’s approach focuses on ensuring people can adopt AI systems in real conditions. An adoption pathway is a reusable route that combines technical architecture, data and safety governance with evaluation benchmarks and deployment playbooks, maintained in a way that lets others adopt it without starting from scratch. This pattern enabled Amul’s three-week deployment.

The work began much earlier, though. AI4Bharat at IIT Madras spent years collecting speech data across 400 districts to build datasets that reflect India’s actual linguistic reality. That foundation enabled the government’s Bhashini language platform, which serves countless people, and EkStep Foundation’s AXL that personalizes learning for millions of students in government schools.

These systems have moved beyond pilot projects to reliable production infrastructure, serving populations larger than many countries. Vistaar provides agricultural advice in Marathi and local dialects because it can draw on Bhashini and AI4Bharat’s multilingual models. Without these shared building blocks, every agricultural system would need to rebuild language capabilities from the ground up. With them, deployment becomes repeatable, costs drop and timelines compress.

The real friction in the adoption of these tools may have more to do with risk exposure than technical capability. Institutions hesitate because adoption could fail publicly, disrupt workflows, result in compliance burdens or create accountability gaps should recommendations go wrong.

Every pathway includes specific mechanisms to address adopter risk. Evaluation and testing protocols ensure systems behave as expected before they are deployed. Human oversight makes space for escalation channels in special cases. Monitoring tracks performance in real conditions. Institutional backing creates accountability structures. These reduce reputational, political and compliance risks for adopters. Weak diffusion capacity creates a strategic risk: when only a few actors have the capability to deploy AI at scale, resilience and public trust are hard to acquire. India’s digital infrastructure avoided this trap.

Creating these pathways takes time and investment. The innovation is largely done; models work. What matters now is diffusion infrastructure to enable adoption across sectors: agriculture, education, healthcare and governance. The creation of this infrastructure requires governance mechanisms that can help generate institutional trust and interoperability, apart from the data pipelines, safeguards and accountability mechanisms that let AI work under real constraints.

AI diffusion demands that an invisible framework be trusted across India. The difference between controlled pilots and population-scale infrastructure will determine whether AI works for a billion people who face real constraints and challenges.

The author is CEO and co-founder of Ekstep Foundation.