Kwaku’s unfamiliar cocoa leaf damage. A wrong pesticide could put his crop at risk.

Unfamiliar cocoa leaf damage calls for identification and reviewed guidance before a farmer buys or applies a pesticide. AgriVoice is being prepared to help Asante-Twi-speaking cocoa farmers describe what they see, hear a reviewed response, and reach an extension officer when the answer is uncertain.

The decision before the pesticide store

At dawn, Kwaku stood beneath a cocoa tree outside Kumasi with three leaves in his hand. The newest leaves had pale marks he had not seen in that corner of the farm before. His daughter needed him back before school, and the pesticide store would open soon. A neighbour had already suggested a product by name.

The risk sat in the leaves, not in the length of the walk to town. If Kwaku chose the wrong product, he could spend money on a treatment that did nothing while the damage spread. If he sprayed without knowing the correct use instructions, he could put himself, other workers, and the crop at risk.

This is an illustrative pilot-day scenario, not a customer case study. It reflects the decision AgriVoice is designed to support: pause, explain the problem in Twi, and get only reviewed guidance or a clear escalation.

A spoken question can begin simply: what colour are the marks, where are they on the leaf, are pods affected, when did the change begin? Those details give a safer starting point than a guess based on one damaged leaf.

A voice question should lead to a safe next step

AgriVoice is being prepared as a constrained voice workflow for cocoa advice. It is designed to recognise a farmer’s question, select from reviewed content blocks, return the selected response in spoken Asante Twi, and escalate questions that cannot be answered safely.

That boundary matters. The system is not meant to invent agronomy advice. A confident-sounding answer with the wrong disease name can push a farmer toward the wrong action. The product’s content path is deliberately narrower: reviewed blocks for known questions, an extension officer for missing, uncertain, or higher-risk questions.

Pesticide questions require even more care. Product names, concentrations, dosage, re-entry time, and pre-harvest intervals cannot be filled in from memory or guessed from a partial description. AgriVoice currently withholds chemical guidance where those details have not been verified by an agronomist. A refusal followed by escalation protects the farmer better than an answer that sounds complete.

That choice can feel frustrating when Kwaku is ready to leave for the store. It is still the right interruption. The bottle should remain unopened until the instruction is tied to the actual product and has been reviewed. A different concentration can change the decision completely.

Uncertainty needs a person at the end of the line

Kwaku’s description may be incomplete. Wind, poor signal, code-switched speech, or a noisy recording can change what a system hears. Leaf damage can also share visible signs with other problems. The safe response is sometimes a request for another observation, a photo where the channel supports it, or a handoff to a named extension officer.

That handoff has to be real. A queue without an accountable person leaves the farmer with the same uncertainty, only later in the day. The planned AgriVoice pilot requires a cocoa-sector partner to recruit farmers and name the extension officer who receives escalations. It also measures whether escalations are answered within a working day.

This approach follows a practical truth in pest response: accurate guidance depends on getting the right information to people who can interpret it. FAO, IITA, and Ghana’s Ministry of Food and Agriculture have trained extension agents and radio personnel on pest guidance while introducing AI-driven tools for identifying, mapping, and reporting outbreaks. Voice tools can help bring a farmer’s first question into that chain. They cannot replace the expert at the point where uncertainty remains.

The morning after the right pause

Kwaku does not need a lecture before breakfast. He needs to know what to do next.

In the intended AgriVoice flow, he speaks in the language he uses on the farm. If his description matches a reviewed question, he hears the approved guidance. If it does not, the workflow says so plainly and routes the question onward. No invented diagnosis. No pesticide instruction without the required safety content.

Later, he returns to the same tree with a clearer next step than the one he had at dawn. He may be asked to watch a particular pattern on the leaves, avoid a particular action, or wait for the extension officer’s response before buying anything. The changed detail is small but important: the pesticide store is no longer the first decision. The question is.