Digital Green's Farmer.Chat puts the human between the model and the farmer

Who the interface is actually for

Most agricultural chatbot pitches imagine a farmer typing questions and getting answers. Digital Green explicitly did not build that. Farmer.Chat is deployed as an assistant to extension agents, with the farmer on the other side of a human conversation.

CEO Rikin Gandhi is direct about why: the farmer-facing option was rejected as too risky.

To safeguard against the risk of the chatbot advising farmers in error, we carefully curated the chatbot's knowledge base and deployed the chatbot as an assistant to extension agents, rather than deploying the chatbot directly to farmers. This enables another stage of human review.Montana Labs

This reframes the product surface. The visible frontend is still the agent's conversation with the farmer, often face to face. The chatbot is a back-office tool the agent consults. That structural choice — a person as the final rendering layer — is what makes the deployment defensible in a domain where a wrong answer damages a crop and a livelihood.

Meeting agents where they already work

The frontend strategy has moved across three surfaces. The first pilot was a custom chatbot on GPT-4. Digital Green then shipped a Farmer.Chat GPT inside ChatGPT, adding multimodal photo input so an agent can photograph a crop and get a diagnosis, plus real-time weather and market data.

The most telling step is the third. Digital Green is now using the Assistants API to bring Farmer.Chat into WhatsApp and Telegram — interfaces extension workers already use. That is a decision to stop asking agents to adopt a new app and instead embed the capability in tools already open on their phones.

For a field of 4,500-plus agents across Kenya and India, that distinction matters more than any model upgrade. The best interface for this user is the one they haven't been told to install.

Language as a first-class frontend requirement

Farmer.Chat operates in Hindi, Swahili, and regional languages, integrating local translation datasets and services per country. Digital Green frames multilingual access as a driver of its claimed 100x cost reduction — from $35 to $0.35 per farmer — because a single agent can now serve more farmers per day in their own language.

The company is also testing a fine-tuned 'Agri-LLM' so questions and answers can happen in local dialects without a round trip through English. That is an admission that translation-as-middleware is a lossy frontend layer for agricultural nuance, and that the language handling belongs deeper in the stack.

Notably, the training data for that model would flow through a data trust so farmers keep oversight of their own contributions — a governance detail wired into how the interface sources its knowledge, not bolted on afterward.

The lesson: constrain the surface, then widen the reach

Farmer.Chat's design sequence is worth copying for any high-stakes advisory product. Digital Green narrowed the interface to a reviewed intermediary, curated and government-validated the knowledge base behind RAG, and only then expanded reach — to more languages, more states, and more messaging channels.

The specific implication is that in advisory domains where errors carry real cost, the frontend question is not 'how do we get the model in front of end users' but 'who should be standing between the model and the consequence.' Digital Green answered with the extension agent, and built every subsequent surface — ChatGPT GPT, WhatsApp, Telegram — around that person rather than around the farmer.