SMS Agricultural Advisory AI Assistant

The problem

A large share of smallholder farmers in Congo have access neither to a nearby technical advisor nor to an internet-connected smartphone — the two channels typically used to share advice on crop disease, planting schedules or market prices.

SMS and USSD, by contrast, remain accessible on nearly the entire installed base of phones, including the most basic devices — a channel largely underused for this kind of high-value service.

Our approach

We framed the problem around a hard technical constraint: an SMS fits 160 characters, and a useful answer has to fit that format without losing relevance. This rules out large language models outright, poorly suited to this channel and its per-message cost.

The analysis focused on a narrow but high-impact set of questions (identifying common diseases from symptoms described in natural language, planting-calendar reminders, localised weather alerts) rather than a general-purpose assistant.

Technical solution

The concept relies on a lightweight classification model, trained on a corpus of agricultural questions in French and local languages, able to route each message to a pre-validated answer reviewed by agronomists rather than generating free-form text — a deliberate choice to guarantee the reliability of the advice given.

Localised weather alerts are pushed proactively based on the area the farmer declares at sign-up, with no daily action required on their part.

Any question outside the covered scope triggers an honest response stating the service’s limits, rather than an approximate attempt at a sensitive topic like crop health.