AgriVoice is looking for one cocoa-sector partner to run a two-week pilot with 20 to 50 Asante-Twi-speaking farmers, backed by one named extension officer. The pilot will test whether farmers receive a useful reviewed answer or a clear human escalation, without unsafe pesticide guidance reaching anyone.
At 6:40 on a damp morning, Kojo stood beside a cocoa tree at the edge of his plot, holding a pod marked with pale patches. His younger brother had already sent him two voice notes with different advice: one said to spray immediately; the other said to cut away every affected pod. Kojo had work waiting and no confidence that either answer was safe.
This is an illustrative composite, but the pressure is real. A wrong instruction can cost a farmer time, money, crop health, or a worker’s safety. If the marks signal something that needs an extension officer’s judgment, a confident guess is the dangerous outcome.
AgriVoice is being prepared for that exact moment. A farmer sends a question by voice in Asante Twi. The system transcribes it, selects from reviewed guidance, speaks the response back, and routes uncertain or missing cases to a person. It does not invent agronomy advice. For pesticide questions, it must preserve required safety information, and it must escalate when reviewed dosage, re-entry, or pre-harvest details are unavailable.
A pilot needs a real person behind the escalation
Kojo’s question may be common, ambiguous, or outside the available guidance. In each case, the farmer needs to know what happens next.
That is why this pilot requires one named extension officer from the partner organisation. The officer receives escalations and owns the response path. A voice tool can help a farmer reach reviewed information quickly; it cannot replace accountable agricultural support when the answer is uncertain.
The partner would recruit the pilot cohort, help make consent and deletion behaviour understandable, and work with the named officer on a practical escalation routine. The goal is modest on purpose: one group, one language workflow, one route for questions that require human judgment.
When Kojo receives a response, the important test is not whether the technology sounds impressive. It is whether he understands what to do next. If the system cannot safely match his question to reviewed guidance, a clear handoff is better than a polished wrong answer. The pesticide detail that stops AgriVoice from answering explains why withholding an answer can be the correct safety decision.
The work before farmer exposure is part of the product
A pilot should begin only after the advice is ready to be heard and acted on.
The current content set needs spoken Asante Twi reviewed by a translator who understands farming vocabulary. Cocoa terms cannot be treated casually. Earlier machine translation work turned the Twi word for cocoa into “chicken,” a fluent error with no place in spoken farm guidance.
Three chemical-related content blocks also remain held until a Ghanaian agronomist verifies dosage, re-entry, and pre-harvest information. They stay unavailable until that review happens. The pilot also needs real farmer voice questions, recorded and transcribed with consent, so the team can measure speech recognition on the language people use in the field rather than relying only on cleaner read speech.
For Kojo, this preparation changes the outcome. Instead of receiving an answer that sounds certain because a system produced it, he can receive reviewed guidance when it fits, or an honest escalation when it does not. The distinction matters most when the decision cannot be safely undone.
Two weeks should produce a decision, not a publicity story
The pilot is designed as a learning gate, with clear measures rather than broad promises.
The team will track whether questions are answered or correctly escalated, whether farmers understand the spoken response, whether they return to ask another question in the second week, how quickly the extension path responds, and how long the full interaction takes. It will also measure the cost of a completed question by component. Any unsafe pesticide answer is a failure condition.
A useful result could support a tightly scoped next step. Repeated confident wrong answers, poor comprehension, broken escalation ownership, or conversation-breaking delays should lead to a redesign or stop decision. That discipline protects farmers and partners from a pilot that looks active while producing no dependable evidence.
By the following week, Kojo’s next question should leave a trace that the partner can inspect: what he asked, whether a reviewed block was selected, whether he understood it, and whether an officer took over when needed. That is the standard AgriVoice wants to test with one committed cocoa-sector partner.
Comments
No comments yet.