GenAI for smallholder farmers — what actually changes

Generative AI in agriculture is having a moment. Most of the demos are in English, on a laptop, answering questions a farmer would never ask. The harder — and far more useful — problem is generative AI that answers a real farmer's real question in the dialect the farmer actually speaks, at the moment the decision matters.

That last clause is the one almost nothing in market does.

The three brittleness points Building GenAI advisory for smallholder farmers fails at three places, in this order:

• Language and dialect. A farmer in eastern UP does not ask a question in textbook Hindi. They ask in Bhojpuri. A model that handles Hindi but not Bhojpuri is, to that farmer, an English model.
• Timing. Agronomy is a calendar problem. An advisory about top-dressing arrives useful in week 6 and useless in week 9. A GenAI layer without a scheduler is a chatbot, not an advisory.
• Provenance. The advisory has to be grounded in the partner institute's own research corpus. A generic LLM answer is liability, not value.

What a credible build actually contains - A multilingual voice + text front end (WhatsApp + IVR) - A semantic retrieval layer over the partner's agronomic corpus - A no-code logic builder so agronomists can author rules without engineering tickets - A cohort-aware scheduler that pushes time-sensitive nudges, not just answers questions - A feedback loop so the agronomist sees what the model is being asked and where it fails

The bigger pattern GenAI in development-sector contexts is mostly a packaging problem, not a model problem. The model is the easy part. Making it speak Bhojpuri, ground its answer in CIMMYT's published agronomy, push the right nudge in week 6, and route the hard question to a human — that is the work.