Kwaku’s cocoa question is uncertain. A wrong pesticide answer could cause harm.

A voice system that cannot safely identify a cocoa question should say so and send it to a named extension officer. For code-switched WhatsApp voice notes, AgriVoice is designed to select reviewed advice only when the match is clear, then escalate uncertainty instead of generating agronomy.

When a voice note carries a high-stakes question

This is an illustrative composite: at 6:40 a.m., Kwaku stands beside a cocoa plot outside Kumasi with his phone held close to his mouth. Rain has left dark marks on several pods, and he records a WhatsApp note in Twi, switching into English for the product name he remembers from a shop.

He asks whether he should spray that afternoon. His eldest daughter is waiting at home with breakfast, but the farm decision cannot wait. If the disease is spreading, delaying could cost him pods. If he uses the wrong product, dose, or timing, he could expose himself and others in the field to harm.

A system that hears a familiar word and fills in the rest with confidence creates the most dangerous kind of answer: one that sounds useful.

Confidence must have a boundary

Twi speech often includes English words, especially for product names, farm inputs, and organisations. A farmer may say “fungicide,” then return to Twi. Audio may include wind, another person speaking nearby, or a word the speech recogniser transcribes poorly.

That uncertainty matters because the system is not there to improvise agricultural instruction. AgriVoice is built around reviewed content blocks. The language model selects among those blocks; it does not write new crop advice. Questions that do not map safely to a reviewed answer must remain open.

The same rule applies with translation. In an earlier test, “kokoo,” meaning cocoa, came back as “chicken.” Fluent output can still carry the wrong noun. That is why the project is testing direct Twi selection against translation-first selection with real farmer questions and real ASR transcripts. The comparison must measure unsafe selections and correct escalations alongside accuracy, latency, and cost. Read more in What Happens When Cocoa Translates as Chicken?.

The useful answer may be an escalation

Kwaku’s note contains enough uncertainty that the system cannot safely choose a pesticide block. Instead of replying with a guessed instruction, it records the question for the extension officer named by the cocoa-sector pilot partner.

That handoff needs to work in practice. A farmer should know that the question has reached someone. The extension officer needs the voice note, the available transcript, and the reason the system withheld an answer. The response path needs an owner and a measured response time.

For chemical questions, the safety boundary is even tighter. Three AgriVoice content blocks remain held because verified dosage, re-entry, and pre-harvest information is missing. Refusing to answer those questions is the correct current behaviour. A polished voice response cannot make absent safety information safe.

The pilot entry gates require a named extension officer before farmers are exposed to the workflow. That requirement protects both sides of the conversation. Kwaku has somewhere for an uncertain question to go. The officer receives a question that has been clearly marked as needing judgment, rather than being asked to clean up a confident mistake after the fact.

What changes after the handoff

Later that day, Kwaku sees that his message is still open for a person to review. He has not been told to spray. He has not been given a dosage the system cannot verify. The farm decision remains serious, but the uncertainty is visible instead of hidden behind a smooth-sounding answer.

That is a better starting point for a two-week farmer pilot than pretending every voice note can be solved automatically. The test is whether farmers receive useful reviewed answers or correct escalations, understand the spoken response, and return when they have another question. Any unsafe pesticide answer is a failure.

The most valuable output from a local-language voice service may sometimes be a careful pause. For cocoa advice, that pause should lead somewhere accountable: a reviewed answer when one exists, and a real person when it does not. See What Happens When a Cocoa Question Needs a Human Answer? for the operational work that makes that promise meaningful.