Kwame’s spraying timing is uncertain. The chemical stays sealed.

A safe spraying decision should pause when the timing, product, dosage, re-entry interval, or pre-harvest interval is uncertain. The right next step is a clear handoff to a named extension officer while the chemical remains unopened.

At 6:12 a.m. on a Tuesday, Kwame stands at the edge of his cocoa plot outside Kumasi with a blue sprayer tank beside his boot and an unopened chemical sachet in his hand. Rain has left the leaves wet. He wants to know, in Asante Twi, whether he can spray now and return to the farm later that day to weed.

The question sounds short. The consequences are not.

If the answer is wrong, Kwame could expose himself or someone else entering the plot. If the interval before harvest is wrong, he could make a decision that affects the crop he has spent months tending. Waiting also has a cost. The marks he has noticed on the pods will still be there after breakfast, and he is worried that another wet day will make the problem worse.

He sends a voice question in Twi because that is how he can describe the situation quickly, using the words he uses on the farm. The important part is not getting a fast sentence back. It is preventing a confident answer from filling a gap nobody has verified.

A timing question can contain several safety questions

“Can I spray today?” can mean several things at once. Is the product name clear? Is the product suitable for the problem? What dosage applies? How long should a person stay out of the treated area? Is there a required interval before harvesting?

A voice system should separate those questions instead of pretending they are one simple yes-or-no request. AgriVoice is being prepared around reviewed cocoa guidance, where the language model selects an approved answer block rather than writing agronomy advice itself.

That distinction matters most for chemical questions. A polished reply can still be unsafe if it invents a dosage or misses a re-entry instruction. Neuralis has already held three chemical guidance blocks because their dosage, re-entry, and pre-harvest details have not been verified by a Ghanaian agronomist. Holding them back is a safety decision.

The same principle applies when speech recognition is unsure. A farmer may use an English product name inside a Twi question, speak over roadside noise, or describe a symptom in a way that could fit more than one problem. The system needs permission to say that it cannot safely determine the answer.

The useful answer may be a pause

In Kwame’s hypothetical case, the system recognizes that the question concerns spraying timing and field re-entry. It finds that the required product-specific details are uncertain or unavailable in the reviewed guidance.

So it does not tell him to spray.

Instead, it returns the safe part that has been reviewed: keep the chemical unopened, avoid spraying until the product and timing can be confirmed, and send the question to the extension officer responsible for escalations. The uncertainty stays visible. It is never translated into false confidence.

That moment can feel frustrating. Kwame wanted to start before the morning moved on. Yet a pause with a real next step is more useful than a vague warning such as “be careful,” and far safer than guessed instructions.

The handoff only works if somebody owns it. Before a farmer pilot can begin, AgriVoice needs a cocoa-sector partner that recruits the cohort and names the extension officer who receives escalations. The pilot scorecard also sets an operational expectation: the escalation owner should respond in less than one working day.

A named person, a visible queue, and a stated response window turn “ask an expert” from a dead end into a route forward. Why AgriVoice refuses to guess and escalates explains why that route matters when pesticide guidance is incomplete.

Voice access needs a human route behind it

Voice can lower the barrier to asking a question in the language a farmer speaks most comfortably. It does not remove the realities around access. Connectivity, electricity, device availability, and comfort with AI can still determine who gets help and who does not. IFPRI has warned that agricultural chatbots and voice assistants can deepen unequal adoption when those conditions are ignored.

That is why the current goal is a small, measured pilot rather than a public scale launch. The proposed two-week AgriVoice pilot would involve 20 to 50 cocoa farmers, with checks for safe answers, comprehension, repeat use, response speed, escalation operations, and cost per completed question.

A successful answer can be an approved response. It can also be a correct escalation. Both outcomes protect the farmer from an answer that should never have been given.

For Kwame, the change comes before he opens the sachet. He sets it back on the low wooden stool by the plot entrance and sends the product name and his question for review. The spraying decision has stopped, but the request has moved. That is the point.

Build trust by making uncertainty actionable

Agricultural advice carries weight when it arrives in a farmer’s own language, especially when it is spoken aloud. That makes the boundary around uncertainty even more important.

The system should say what it knows from reviewed material. It should identify what it cannot confirm. Then it should pass the unresolved question to a person with responsibility to answer it.

This is a practical design choice, not a refusal to help. A farmer who hears “I cannot safely confirm that yet, and this officer has your question” has a path. A farmer who receives an invented interval has a risk disguised as assistance.

Kwame’s morning remains unfinished for a while. His field still needs attention. But the chemical stays sealed until the timing is confirmed, and the next decision belongs with someone accountable for the advice.