As the global discourse on Artificial Intelligence (AI) matures beyond Large Language Models (LLMs) confined to digital abstraction, a paradigm shift is emerging from the Global South. This paper examines two pioneering AI deployments in India: ?AI Sarlaben? in Gujrat and ?Mahavistaar? in Maharashtra. I argue that generative AI inverts the cost structure of farm extension: the marginal cost of one more personalised approach zero, while fixed costs migrate to three scarce assets-verified data, institutional trust and language infrastructure. I distinguish two archetypes: a Library model (Mahavistaar), which retrieves from vetted public knowledge and cites its sources, and a Ledger model (Sarlaben), which reasons over transaction- level records generated by a cooperative market institution. This distinction determines what can be replicated elsewhere: architecture, guardrails and voice interfaces travel well: ledgers and trust do not. This article explores the architectural efficiency, historical data leverage, and socio-economic integration that make these indigenous AI models masterclasses in solving real-world,? last-mile? bottlenecks.