Safe escalation for Twi cocoa questions depends on work completed before the first farmer sends a voice note. A named extension officer, reviewed spoken answers, clear refusal rules, and a working handoff turn an unresolved question into a task someone can safely own.
At 2:17 p.m., Ama Mensah opens the afternoon queue on a borrowed laptop at a cocoa cooperative office outside Kumasi. Her tea has gone cold beside a paper notebook, and three voice notes wait above the one she has been expecting: a farmer asks about black pod after heavy rain; another asks whether a spray can be used again; a third message is muffled by wind and a passing motorbike.
The pesticide question stops her.
The farmer says the label is damaged. He wants to know what to mix before workers return to the farm. If Ama gives a dose from memory and it is wrong, someone could enter too soon, handle the wrong chemical, or harm the crop. If nobody responds, he may spray anyway because rain is coming and the pods are already at risk.
The queue has arrived. The safety decision was made weeks earlier.
The queue begins before the first message
An extension queue can look like an operations problem: messages come in, a staff member replies, the list gets shorter. For agricultural advice delivered by voice, the hard part happens before that list exists.
AgriVoice is being prepared for a two-week pilot with 20 to 50 Asante-Twi-speaking cocoa farmers. Its job is narrow. A farmer speaks in Twi, the system identifies a reviewed answer block when one fits, returns it as speech, and sends uncertain or sensitive questions to a human.
That boundary matters because fluent speech can create false confidence. A system may recognize enough of a question to sound helpful while still missing the word that changes the advice. In early translation work, “kokoo,” meaning cocoa, was rendered as “chicken.” The sentence could still sound complete. It could also send a farmer in the wrong direction.
The system therefore needs reviewed Twi content before exposure. It needs a farming-aware translator, because advice written for a person to read silently can sound unfamiliar or unclear when spoken aloud. It needs a controlled set of approved blocks, rather than a model freely composing agronomy guidance.
Ama’s queue is safer because the system has permission to stop.
Chemical questions need a person with a name
Three chemical-content blocks remain held because verified dosage, re-entry intervals, and pre-harvest intervals have not yet been supplied by a Ghanaian agronomist. That is an unfinished gate, not an inconvenience to work around.
For the farmer in Ama’s queue, the correct automated response is not a guessed mixture. It is a clear explanation that the system cannot verify the instruction, followed by escalation to the extension officer responsible for the pilot.
This is the difference between an escalation label and an escalation path. A label can mark a message “needs review” and leave it sitting in a dashboard. A path assigns ownership before the pilot begins: who receives it, what information they see, and what response time the partner has agreed to meet.
The pilot scorecard sets a practical standard: escalations should have a named owner and a median response time under one working day. That does not promise that every question receives an immediate answer. It makes responsibility visible when an answer cannot safely be automated.
For more on why a missing pesticide detail must trigger restraint, see What Should Voice AI Do When Pesticide Guidance Has Not Been Verified?.
A safe handoff carries context, not guesswork
Ama presses play again. This time she hears enough to understand the immediate risk: the farmer has a damaged label and workers nearby. The system has preserved the original voice note, its transcript where available, and the reason it escalated. She does not need to reconstruct the problem from a vague notification.
That context should include the farmer’s question, the answer category considered, and the uncertainty that blocked an automated response. It should also respect consent, retention, and deletion rules tested before field use. A voice note may contain a phone number, a location reference, or details a farmer did not expect to become part of a permanent record.
The human handoff needs rehearsal. Staff should run the full route before farmers do: consent prompt, incoming audio, reviewed answer, refusal, escalation, and a real response. A queue that works only in a product demo can fail when one garbled recording arrives beside a question where caution matters.
The morning after the queue is cleared
Ama calls the farmer through the agreed process. She does not invent a spray instruction. She tells him not to act on an unverified mixture, records the question for the responsible agronomy route, and makes sure the next step belongs to a person rather than an empty status field.
By the next morning, the message has an owner, a documented reason for escalation, and no unsafe instruction attached to it. That is a small outcome, but it is the one a pilot must prove repeatedly.
A useful voice service will sometimes answer quickly. Its more important test comes at 2:17 p.m., when it recognizes that the safest answer is to bring the right human into the conversation.
Comments
No comments yet.