The Queue Without an Extension Officer, and What It Could Cost Cocoa Farmers

A cocoa-sector institution can help run a measured two-week AgriVoice pilot by recruiting 20 to 50 Asante-Twi-speaking cocoa farmers and assigning one extension officer to own the escalation queue. The pilot uses reviewed Twi advice for known questions and sends uncertain or safety-sensitive questions to that named person.

A queue is only useful when someone owns it

In April 1970, Apollo 13’s crew faced a carbon dioxide problem after an explosion ended the planned Moon landing. The lunar module had round receptacles for CO2 scrubber cartridges, while the command module had square cartridges. Mission Control in Houston had to devise a way to use the available materials so the crew could keep breathing.

The outcome was uncertain at the point the problem appeared. Engineers on the ground worked from the same limited equipment available to Jim Lovell, Jack Swigert, and Fred Haise, then passed a practical procedure to the crew. Lovell and Jeffrey Kluger document the episode in Lost Moon, the account behind the film Apollo 13.

AgriVoice has a smaller version of that design problem. A farmer may send a voice note in Asante Twi about black pod, swollen shoot symptoms, pruning, shade, harvesting, or a pesticide. Some questions can receive a reviewed answer quickly. Others need a person with local authority, especially when a description is unclear, several problems may be present, or chemical guidance lacks verified safety details.

A queue without an accountable extension officer only moves uncertainty from a farmer’s phone to a database. The pilot needs one named person who can review escalations, decide when to ask a follow-up question, and make sure a farmer does not treat an incomplete answer as a spraying instruction.

Reviewed Twi comes before confident voice replies

AgriVoice is being prepared for cocoa farmers who prefer to ask and hear questions in Asante Twi. Its system can transcribe a voice question, select from reviewed content blocks, return an answer by voice, and escalate when the safe response is uncertainty.

That order matters. The language model selects reviewed block IDs. It does not write agronomy advice. Three chemical-content blocks remain withheld because dosage, re-entry, and pre-harvest information still require Ghanaian agronomist review. A refusal in those cases is a safety control.

Translation also requires care. Early testing exposed a serious example: “kokoo,” meaning cocoa, was translated as “chicken.” A fluent sentence with the wrong domain noun can sound persuasive while sending a farmer in the wrong direction. Before exposure to farmers, the approved advice blocks need spoken Asante Twi reviewed by a farming-aware translator.

The first escalation queue will therefore be part of the product, not an exception to it. It will show which questions the reviewed content handles well, which questions are ambiguous, and which gaps deserve expert attention. That is more useful than claiming broad coverage before the evidence exists.

For the underlying risk in this work, see Kwaku’s cocoa question is uncertain. A wrong pesticide answer could cause harm.

What the two-week pilot will measure

This is an invitation to one cocoa-sector institution that can recruit a small farmer cohort and designate the extension officer who receives escalations. The institution may be a cocoa-sector partner already supporting farmers, with enough local trust to explain consent, encourage realistic questions, and help assess whether replies are understood.

The pilot will test a narrow, visible set of outcomes:

• Whether at least 70% of farmer questions are answered safely or correctly escalated.
• Whether any unsafe pesticide instruction reaches a farmer. The target is zero.
• Whether at least 80% of participants report understanding the voice response.
• Whether at least 30% of activated farmers ask a second question in week two.
• Whether escalation has a named owner and a median response time below one working day.
• Whether the median automated response stays below eight seconds, excluding human escalation.
• What each completed question costs across WhatsApp, speech, language-model, hosting, and human-review components.

These measures make the decision after the pilot concrete. A good result supports a carefully scoped continuation. Repeated confident wrong answers, poor comprehension, or an unresolved escalation queue would call for redesign or stopping.

The institution does not need to promise scale. It needs to help create conditions where the evidence is trustworthy: real farmer questions, clear consent and deletion behavior, a reachable extension officer, and honest reporting when the system fails.

The first queue should teach us where the service ends

Before the pilot begins, staff will run the complete path, including a question that should be escalated. We will also collect and transcribe real farmer voice questions to measure speech recognition in field conditions. Read speech and a formal demonstration voice do not prove that a farmer’s natural voice note will be understood.

The pilot should begin only after WhatsApp works end to end, reviewed Twi content is loaded, chemical blocks are either cleared by an agronomist or kept unavailable, and the escalation path has been rehearsed.

Apollo 13 did not depend on a generic promise that Mission Control would “handle it.” The response worked because people had a defined problem, constrained materials, a shared procedure, and a crew who could act on it. AgriVoice needs the same discipline in its own setting: reviewed answers for what is known, a person for what is uncertain, and a record of what happened next.

An institution ready to assign that person and support a two-week cohort can help establish whether local-language voice advice earns a place in cocoa extension work.