The cocoa sector loses good guidance not because good guidance is scarce, but because the questions young farmers actually ask rarely reach the people who could answer them. Over two weeks, AgriVoice will find out whether a voice workflow can close that gap for real users. We are inviting one cocoa-sector partner to recruit 20 to 50 farmers, name the extension officer who handles anything the system cannot answer, and review the evidence with us at the end.
Akua is twenty-six. She took over her father's four-acre plot in the Ashanti Region when he moved to Kumasi. Her black pod disease was spreading, she had heard three different dates for when to spray, and the shop where she buys inputs told her something different each visit. She mostly listens to farming advice on the radio and from neighbours, because she does not read English comfortably and she cannot reach an extension officer quickly. Akua is exactly who AgriVoice is built for, and exactly the kind of farmer our pilot needs.
Why a two-week pilot beats a grand launch
In 1969, a small team at a Palo Alto research lab thought they could build a network that let computers in different places talk to each other. They had a plan, a deadline, and almost no idea whether it would work. On the evening of October 29th, they tried to send the word "login" between two machines. The system crashed after the first two letters. The "l" and the "o" had gone through, so they tried again and it worked. That moment, documented in the team's own logs and retold in histories of the early internet, was the first successful message on what became the ARPANET. They did not start by connecting a hundred sites. They started with two machines, one message, and a willingness to look at what actually happened.
The same discipline applies to voice AI for farmers. A broad public launch might reach a thousand people and tell us nothing about whether the answers are safe, whether farmers understand the voice, or whether anyone comes back for a second question. A controlled pilot with 20 to 50 farmers tells us exactly those things in two weeks. We learn what breaks, what confuses, and what is worth building next. That is the point of the ARPANET story, scaled down: prove the mechanism on a small, real system before you claim it works at all.
What the partner provides, and what they get back
AgriVoice works like this. A farmer records a spoken question in Asante Twi on WhatsApp. The system understands it, selects a reviewed block of agronomy content, and speaks the answer back in the same language. It does not generate advice from scratch, it chooses from content a Ghanaian agronomist has reviewed. Pesticide responses carry mandatory safety information. And when the system is not confident, it does not guess. It escalates to a named person.
The partner's role is specific. They recruit the farmers and run the two-week timeline. They name the extension officer who receives escalations and commits to responding within one working day. They review the scorecard with us at the end: what percentage of questions got safe, useful answers, whether farmers said they understood the responses, and whether at least 30% of activated farmers came back to ask again in the second week. The partner sees the same evidence we do. If the pilot fails a threshold, we do not declare victory, we make one tightly scoped fix or we stop.
We are deliberately honest about what is ready. The voice that reads the answers has a measured intelligibility gap and sounds formal, because it learned from scripture recordings. It is acceptable for this pilot, not for reading news to someone's parent. Machine translation cannot do the content job, we saw "kokoo" come back as "chicken" on a live call, so the content is being translated and reviewed by a farming-aware Twi speaker. The three chemical blocks with no verified dosage stay locked until a Ghanaian agronomist clears them. Nothing reaches a farmer until the full team has run the app end to end and found no unsafe answer.
Farm funding news versus verified guidance
One of the hardest problems in this sector is separating what is true from what is current. A policy announcement about cocoa pricing is not the same as a verified recommendation about when to spray for black pod disease. AgriVoice handles this by routing: policy news and agronomy guidance are different classes of content, and only the latter is confirmed by an agronomist before it is spoken. The pilot measures whether that separation holds in practice, and whether farmers notice the difference when they hear it.
The 2026/2027 purchasing season is on the horizon, and young farmers are already making decisions they cannot easily undo. Spraying too early or too late costs money and yield. Following a wrong dosage can harm the crop or the person applying it. The questions they ask deserve answers that are reviewed, spoken in their language, and owned by someone accountable when the system does not know.
An invitation, not a promise
We are not promising a finished product. We are inviting one partner to join a controlled experiment and help us answer a narrow question: can a Ghanaian-language voice workflow deliver useful, safe answers to real farmers, through an institution that can distribute or pay for it? The ARPANET started with "l" and "o" and a crashed machine, then a working message, then a network. AgriVoice starts with a two-week pilot, a named extension officer, and a scorecard nobody gets to cherry-pick.
If your organisation works with cocoa farmers, and you can recruit a small cohort and name the person who takes the escalations, we want to talk. Bring the questions your farmers are already asking. We will bring the reviewed content, the voice, and the discipline to report what we find.
Comments
No comments yet.