A safe voice service needs a named extension officer before a farmer can use it. Without that person, an escalation button records uncertainty but leaves the farmer without accountable follow-through.
In 1854, during a cholera outbreak in Soho, London, physician John Snow gathered evidence linking deaths near Broad Street to the public water pump. The cause of cholera was still disputed, and removing the pump handle did not settle a scientific debate. It created an immediate public-health action by the St James parish authorities while the outbreak’s outcome remained uncertain. The John Snow Archive documents the investigation and the role of the local Board of Guardians.
That distinction matters for AgriVoice. A fallback only protects someone when a real person receives it, understands their responsibility, and can respond.
A pesticide question can reach its safety boundary quickly
Imagine a cocoa farmer sending a Twi voice note: they remember part of a pesticide product name, see a problem on their trees, and want to know how much to apply before returning to the farm.
AgriVoice can recognize the question, select from reviewed advice blocks, and speak a response in Twi. But some pesticide questions require details the product must never guess: the exact product, dosage, re-entry interval, or pre-harvest interval. The safe response is to withhold an instruction and escalate.
That is the point where software reaches its boundary. Saying “contact an extension officer” may sound responsible, but it creates a new burden for the farmer. Which officer? How will the question reach them? Who decides whether the product name was heard correctly? What happens if nobody replies?
The product must answer those questions before the first farmer sends a voice note.
The name turns an alert into a handoff
For the AgriVoice pilot, the cocoa-sector partner needs to own recruitment and name the extension officer who receives escalations. This is an entry gate because safety depends on operations as much as on the answer-selection model.
A named officer gives the escalation a destination. The partner can agree how questions are routed, what information the officer receives, and what response time they can realistically meet. The pilot scorecard sets a target of a median response in less than one working day. That target can be measured only when there is a clear owner.
It also gives the farmer an honest answer. Instead of receiving an authoritative-sounding instruction based on an uncertain product name, they can be told that their question has been passed to the responsible extension contact for confirmation. The system should preserve the uncertainty, not hide it behind fluent speech.
This is especially important because language systems can make errors that sound convincing. Neuralis has already found that machine translation changed the Twi word for cocoa, “kokoo,” into “chicken.” That experience is why reviewed Twi content and human escalation sit inside the workflow. Read more about that risk in The Twi Word That Became Chicken, and the Risk It Created.
Accountability needs a working route in both directions
John Snow’s evidence mattered because it reached people with the authority to act on it. A map left in a notebook would have changed nothing for the households using the pump.
The same principle applies to a farmer’s voice note. AgriVoice needs a route from WhatsApp to reviewed content or to a human, plus a route back with a response the farmer can understand. A queue without an owner becomes a record of unanswered risk.
Before pilot exposure, the team should test the whole handoff with staff: send an uncertain pesticide question, confirm that it reaches the named officer, record when it is accepted, and confirm that a safe answer reaches the original sender. The test should include the cases most likely to fail: background noise, code-switching, an unfamiliar product name, and a request for a chemical interval that has not been verified.
The three held chemical blocks remain withheld until a Ghanaian agronomist clears dosage, re-entry, and pre-harvest guidance. That restraint is a product capability. A system that refuses to invent chemical instructions can earn trust only when the next human step is real.
Build the institution before measuring the model
The two-week pilot is designed to learn whether AgriVoice delivers useful, safe answers for 20 to 50 Asante-Twi-speaking cocoa farmers. Its success cannot be reduced to whether the speech model transcribes a question or whether an LLM selects a likely content block.
The stronger measure is whether a farmer gets an appropriate outcome: a reviewed answer when the evidence supports one, or a timely human response when it does not. Repeated confident wrong answers are a stop or redesign signal. Any unsafe pesticide instruction reaching a farmer is also a stop signal.
Name the extension officer. Agree on their role with the cocoa-sector partner. Test the escalation route before recruitment begins. Then measure every handoff during the pilot.
Snow’s investigation in Soho reminds us that evidence needs an accountable route to action. For AgriVoice, the name beside the escalation button is where that route begins.
Comments
No comments yet.