AgriVoice cocoa pilot: Why a Named Extension Officer Makes Escalation Credible

A two-week AgriVoice pilot needs one cocoa-sector partner to recruit 20 to 50 Asante-Twi-speaking farmers, appoint an extension officer for escalations, and help measure safety, understanding, repeat use and response time. The pilot can then answer a practical question: do reviewed Twi voice responses help farmers enough that they come back with another question?

The unanswered call needs an accountable person

At 6:40 one morning, Kwaku stands beside a cocoa plot with a damp voice note open on his phone. He has noticed marks on a pod and wants to know what to do before the next field visit. His neighbour has offered an opinion. Another farmer has named a chemical. Neither answer gives Kwaku a source he can check.

In this illustrative pilot scene, AgriVoice can recognize the question, select from reviewed guidance and speak the response in Twi. But the hard moment arrives when the question reaches a gap: a pesticide dosage has not been verified, the description is unclear, or the right answer depends on someone seeing the farm. A confident guess could put workers, crops or both at risk.

That is where a named extension officer matters. The system must be able to say, plainly, that it cannot safely answer and pass the question to a person who owns the next step. Without that person, escalation becomes a polite dead end. The pesticide detail that stops AgriVoice from answering explains why withholding an uncertain chemical answer is part of the product’s safety boundary.

For Kwaku, the useful outcome is not a voice assistant that sounds certain. It is a clear answer when one is available, and a reliable route to human help when it is not.

Recruitment turns a demo into evidence

A demo can show that a voice message moves through speech recognition, reviewed content selection and speech output. It cannot show whether farmers understand it in their own routines, whether field noise changes the transcription, or whether the first question is followed by a second.

A partner can make those questions measurable. The partner recruits a defined group of 20 to 50 farmers for two weeks, explains the pilot and consent process, and provides a real connection to extension support. That keeps recruitment and follow-up with an institution already close to the farming context, rather than asking a software team to invent trust from a distance.

The pilot should begin only after its remaining safety gates are met: the approved cocoa guidance needs spoken Asante Twi reviewed by a farming-aware translator; the held chemical blocks need a Ghanaian agronomist’s clearance on dosage, re-entry and pre-harvest intervals; and the WhatsApp route and escalation path need a complete staff rehearsal.

Those conditions protect the farmers and make the result interpretable. If a farmer receives silence because a block lacks Twi text, that is a content failure. If a chemical question is withheld because its safety details are unverified, that is the intended response. The pilot needs to distinguish those outcomes.

Measure the decision, not the novelty

By the second week, Kwaku may ask again, perhaps about a pod mark, pruning, or the meaning of advice he has heard elsewhere. Or he may never return. Both outcomes matter.

The pilot scorecard sets concrete signals:

• At least 70% of questions should receive either a useful reviewed answer or a correct escalation.
• No unsafe pesticide instruction should reach a farmer.
• At least 80% of participants should report understanding the response.
• At least 30% of activated farmers should ask another question in week two.
• Escalations need an accountable owner, with a median response under one working day.
• The median automated response should arrive in under eight seconds, excluding human escalation.
• The cost of each completed question needs to be measured by component.

These are decision measures, not promotional claims. A high first-week message count could reflect curiosity. Repeat use is harder to earn. A fast response that is misunderstood has not solved the farmer’s problem. An answer rate that hides uncertain chemical advice would be worse than a lower rate with correct escalations.

The same discipline applies to language. AgriVoice is designed to select reviewed content rather than author agronomy advice. The system should remain within that boundary when a question is ambiguous, outside the approved material, or unsafe to answer from a distance. What happens when a pesticide question has no named extension officer? is the operational question the partner helps answer.

One partner can make the next choice credible

Near the end of the two weeks, Kwaku is back at the edge of the plot, listening to a response he can understand or waiting for the extension officer who has his question. That is the point of the pilot: to see what holds up when the request comes from a real farmer, in real language, with a real consequence attached.

The outcome may be continue, one tightly scoped iteration, or stop. Each is useful if it comes from evidence. A partner who can recruit the cohort, name the escalation owner and review the scorecard gives AgriVoice the conditions to earn that decision.