A cocoa-sector partner can give one named extension officer a controlled, two-week view of what 20 to 50 Asante-Twi-speaking farmers actually ask. The pilot will measure whether farmers understand the answers, return in week two, receive timely human help, and avoid unsafe advice.
In 1854, physician John Snow faced a similar evidence problem in Soho, London. Cholera deaths were rising, but the dominant explanation blamed contaminated air. Snow went house to house, examined where deaths had occurred, and mapped them around the public water pump on Broad Street.
The pattern suggested contaminated water. Yet the map alone could not settle every case. Snow also investigated apparent exceptions, including nearby workers who seemed unaffected and a woman living elsewhere who had died after drinking water collected from the pump. Local officials agreed to remove the pump handle while the cause was still disputed.
Snow documented the investigation in his 1855 book, On the Mode of Communication of Cholera. His work became a lasting example of why decisions improve when one accountable person can inspect real cases instead of relying on assumptions from a distance.
That is the practical shape of this proposed AgriVoice pilot. The stakes and setting are different, but the mechanism holds: collect what people actually ask, record what the system does, investigate the exceptions, and let a named local professional judge the evidence.
A small pilot built around real farmer questions
Neuralis is inviting one cocoa-sector partner to recruit 20 to 50 Asante-Twi-speaking cocoa farmers for a two-week AgriVoice pilot.
Farmers will ask questions by voice. AgriVoice will convert the speech, select from reviewed cocoa guidance, return the answer in spoken Asante Twi, or escalate when the available guidance does not safely cover the question. The language model selects reviewed content blocks. It does not compose agronomy advice.
The partner will name one extension officer to receive escalations and help interpret what happens during the pilot. That ownership matters. A queue without a responsible person can leave an urgent question unresolved, even when the software correctly recognizes that it should stop.
The pilot also gives the partner a direct view of demand. Which questions recur? Where do farmers code-switch into English? Which recordings are difficult to recognize? Which topics fall outside the reviewed guidance? The answers can guide future content and training without guessing what farmers need.
Four measurements that determine what happens next
Comprehension comes first. The target is for at least 80 percent of participating farmers to report that they understood the response. A grammatically correct answer has little value when the voice, wording, or language prevents the farmer from using it.
Repeat use shows whether the service remains useful after the first encounter. The continuation signal is that at least 30 percent of activated farmers ask another question in week two. One-time curiosity will be recorded honestly as such.
Escalation time tests the human workflow. The partner must name the officer who owns the queue, and the target median response time is under one working day. Questions involving missing, uncertain, or out-of-scope guidance should reach that person instead of receiving a confident guess. What happens when a cocoa farmer’s question cannot be answered safely? explains that boundary in more detail.
Answer safety is the firmest gate. The target is zero unsafe pesticide answers. Three chemical guidance blocks remain held until a Ghanaian agronomist verifies dosage, re-entry, and pre-harvest intervals. AgriVoice currently refuses those questions, which is the required behavior until the missing details are approved. This example of a chemical question needing a named officer shows why escalation is part of the product rather than an afterthought.
What must be ready before farmers participate
Recruitment alone does not start the pilot. The 37 owner-approved guidance blocks need spoken Asante Twi reviewed by a farming-aware translator. An agronomist must clear or continue withholding the three chemical blocks. At least 20 real farmer questions must be recorded and transcribed to measure speech recognition under field conditions.
Consent and deletion behavior must also be tested. Staff must complete a full safety rehearsal, confirm that uncertain questions reach the named officer, and repeat a real WhatsApp conversation before farmer exposure.
These gates protect the farmers and the partner. They also protect the quality of the evidence. A pilot cannot tell us whether farmers value the workflow if missing translations produce silence or an unowned escalation queue produces delay.
An invitation to inspect the evidence together
The proposed partner does not need to endorse a broad platform or commit to a public launch. The immediate commitment is narrower: recruit a manageable cohort, appoint one extension officer, help complete the entry gates, and review the two-week scorecard.
The outcome may be to continue, make one tightly scoped change, or stop. Repeated confident errors would force a redesign. Poor comprehension would trigger a review of language and voice. Low repeat use would challenge the underlying demand. Costs will be measured by component so any later commercial discussion begins with evidence.
John Snow’s map mattered because each mark represented an observed case, and the exceptions received attention rather than being discarded. One extension officer can create the same discipline here: listen to the questions, examine every escalation, and leave the pilot with a measured account of what cocoa farmers asked and how safely AgriVoice responded.
Comments
No comments yet.