A cocoa-sector partner can help test AgriVoice with 20 to 50 Asante-Twi-speaking farmers over two weeks by recruiting the cohort, naming the extension officer who owns escalations, and helping measure understanding and repeat use. The pilot will only send reviewed guidance, while uncertain or safety-sensitive questions go to a person.
At 4:40 p.m., Abena is standing beside her cocoa sacks near the edge of her farm, holding a phone with a cracked corner. Rain has darkened the path home. She has spotted something on the pods and wants to know what to do before it spreads.
In this illustrative composite, the danger sits in the question itself. If the advice is wrong, Abena could waste money on the wrong input, delay a response to a crop problem, or expose herself to unsafe pesticide use. If the system cannot answer safely, silence creates its own problem. Her question needs a clear route to someone accountable.
That is the work we want to do with one Ghanaian cocoa-sector partner.
A small pilot built around real farm questions
AgriVoice is preparing a focused two-week pilot for 20 to 50 cocoa farmers who speak Asante Twi. The purpose is to find out whether a voice workflow can give useful, understandable guidance for the questions farmers actually ask, rather than perform well only in a demo.
A partner would recruit the farmer cohort through an existing relationship and help set expectations before the first message arrives. Farmers need to know what the service can help with, what it cannot answer, and what happens when a question needs human follow-up.
The pilot does not ask a language model to invent agronomy advice. It selects from reviewed content blocks. Each answer must be available in spoken Asante Twi, and pesticide-related guidance carries mandatory safety content. Questions outside the reviewed material, unclear questions, and cases with missing safety information must escalate.
That boundary matters because fluent language can still be wrong. In earlier translation work, the Twi word for cocoa, “kokoo,” was rendered as “chicken.” The sentence sounded plausible. The domain noun had changed. What Happens When Cocoa Translates as Chicken? explains why reviewed local-language content must come before broad claims about automated advice.
An escalation owner turns a queue into care
Abena records her question. The system recognizes enough to identify that the request is uncertain, but no reviewed answer fits safely. It does not fill the gap with a confident guess. Instead, her question is passed to the named extension officer.
For a farmer, this is the point where confidence can disappear. She has already taken time from the farm to ask. The crop issue may worsen while she waits. If nobody owns the queue, the bad ending remains possible: her question stays unanswered until the decision she needed to make has already passed.
A pilot partner needs to name the person or team responsible for that next step before recruitment begins. We will measure whether escalations have an accountable owner and whether the median response is under one working day. A safe escalation path needs more than a forwarding rule. It needs a person who can receive, assess, and close the loop.
This requirement is especially important for chemical questions. Three AgriVoice content blocks remain held because dosage, re-entry, and pre-harvest intervals have not yet been verified by a Ghanaian agronomist. The service must keep refusing those instructions until expert review clears them. The Re-Entry Instructions AgriVoice Didn’t Have, and Why a Farmer Must Wait traces the practical reason for that restraint.
Understanding comes before repeat use
The existing Twi voice is ready for a pilot demonstration, though it has limits. It was trained on scripture recordings, so it can sound formal. A native speaker scored it at 55% on first hearing. That makes field comprehension a core test, not a detail to assume away.
After Abena receives a response, the important question is simple: did she understand what it told her to do? The partner can help gather that answer in language and circumstances that make sense to the farmers involved. We will track whether at least 80% report understanding the response.
Then comes the harder signal. Does she use it again?
A farmer may try a new voice service once out of curiosity. Repeat use in the second week shows something more useful: the service was understandable enough, relevant enough, and reliable enough to return to when another question arose. The pilot’s continue signal is at least 30% of activated farmers asking again in week two.
When Abena gets home, the phone goes back on the shelf beside a small torch and a folded cloth. The next time she sees a worrying pod, the service should earn another question only if the first answer made sense and the difficult question reached a real person.
What the partner helps us learn
This is a learning gate, with clear measures and clear reasons to stop or redesign. We will measure successful answers or correct escalations, unsafe pesticide answers, comprehension, repeat use, escalation operations, response time, and the cost of each completed question.
A partner does not need to promise that AgriVoice will work for every farmer or every cocoa problem. The value of the pilot is an honest answer about where the workflow helps, where the voice creates friction, and where reviewed content or human support must improve.
The immediate need is one Ghanaian cocoa-sector partner prepared to recruit a small cohort, appoint an escalation owner, and examine the evidence with us after two weeks. Abena’s next question should have a safe answer, or a clear human path, before it reaches her phone.
Comments
No comments yet.