In the Río Lempa watershed in El Salvador, farmers and field staff need timely agronomic guidance, and in-person technical assistance can't reach every question. It isn't a matter of effort: nobody can be on every farm when a doubt comes up.
That's why we designed and launched Somos Río Lempa, an AI assistant that answers on WhatsApp. These are the lessons it leaves us, useful for any agricultural extension program considering something similar.
Start with the channel people already use
A new app has to be downloaded, learned and remembered. WhatsApp is already part of many people's daily conversations, so the assistant arrives somewhere familiar and a question is just another message.
Choosing the channel is a design decision as important as choosing the language model, and it should come first.
Experts provide the knowledge
A language model on its own can give answers that sound convincing and are wrong. In agriculture, that's a real risk for the person asking. That's why the assistant answers with expert-validated knowledge.
The technical piece is called RAG: the system first searches that knowledge base and answers from what it finds, so responses can be traced back to their sources. The time goes less into the model than into gathering, organizing and validating that knowledge.
Good advice depends on the climate
What makes sense on a farm changes with the weather. So the assistant integrates climate information: a general answer is worth less than one grounded in a place and a moment.
Every conversation teaches something
The assistant logs each conversation so experts can review it. That record shows what people ask, which answers fall short and what's missing from the knowledge base, and it feeds a continuous improvement cycle.
This work is often left out of budgets because it happens after launch. Without it, an assistant ages quickly. Launching is where the work starts, not where it ends.
It supports field staff instead of replacing them
An assistant provides an always-available channel for questions that would otherwise go unanswered. Field support is still needed, and field staff's judgment is part of what makes the system reliable.
If you're considering something similar
- Set the scope. Decide which topics the assistant covers and what it does when it doesn't know.
- Make time for the knowledge. Involve experts from the start.
- Plan for human review. Assign owners and budget for the period after launch.
- Agree on indicators. For example, queries resolved, response time, expert-validated accuracy and active users.
See how we approach AI agents and assistants for landscapes.