A calendar can only tell a cocoa farmer whether it is safe to act when the advice includes the verified interval that connects treatment to today’s date. When that interval is missing, AgriVoice should stop and send the question to a named extension officer instead of turning an uncertain date into spoken pesticide advice.
At 5:40 one morning, Kwaku stood beside his cocoa plot outside Kumasi with a phone in one hand and an empty sprayer in the other. Dew had settled on the leaves. He could still see the patch where he had sprayed several days earlier, but he could not remember the product instructions clearly enough to know whether he could enter the plot, spray again, or pick pods.
He sends a short voice note in Asante Twi. The question sounds simple: “Can I go into the farm today?”
The danger sits inside the word “today.” If the waiting period after spraying is wrong, Kwaku could re-enter too soon or harvest before it is safe. The lost day matters, but the worse outcome is acting on advice that should never have been given.
A relevant answer can still be unsafe
AgriVoice is being prepared to help Asante-Twi-speaking cocoa farmers ask practical questions by voice. Its safety boundary matters as much as its language support.
For pesticide questions, the system is designed to select from reviewed guidance. It does not invent agronomy advice sentence by sentence. That gives a farmer a response based on content an expert has already examined, and it makes it possible to see exactly which guidance was selected.
But selecting the right topic does not complete the job.
A block might correctly identify that Kwaku is asking about re-entry after spraying. It might explain protective equipment, storage, or the need to follow the product label. Yet the calendar still needs a verified number: the interval for that particular product and use. Without it, “today” remains impossible to judge safely.
That is why three chemical guidance blocks are held back while a Ghanaian agronomist verifies dosage, re-entry, and pre-harvest intervals. A voice system can recognize the question in seconds. It cannot fill in a missing safety fact because the farmer is waiting beside a sprayer.
The safe response is sometimes an escalation
In Kwaku’s scenario, AgriVoice should do something less dramatic and more useful than guessing. It should say that the answer cannot be confirmed from the approved guidance and direct the question to the extension officer responsible for escalations.
That handoff has to work in real life. A named person must receive the question, know they own it, and respond quickly enough for the farmer’s decision. The active pilot measures whether escalations have an accountable owner and whether the median response comes in under one working day.
This is where a voice workflow earns trust. Farmers will ask questions in the moment, often when weather, labour, crop condition, and a previous instruction are all pressing at once. A refusal without a path forward leaves them with the same rumours and half-remembered directions they had before. An escalation turns uncertainty into a specific next action.
Kwaku puts the sprayer down. He does not enter the treated area. He waits for the extension officer to confirm the product-specific interval. The morning’s outcome is modest: a dangerous decision has been delayed until someone can answer it properly.
Reviewed content needs the facts that make it usable
Safety reviews can sound abstract until a missing field changes what a person does before breakfast. Re-entry and pre-harvest intervals are not administrative details. They tell a farmer how to relate an earlier treatment to the work that needs doing now.
The same principle applies to every part of local-language agricultural guidance. Translation needs farming-aware review because a fluent sentence can still carry the wrong domain meaning. Product names, symptoms, timing, and quantities need careful checking because the spoken answer may be treated as authority.
That is also why AgriVoice’s pilot entry gates include reviewed Asante Twi content, agronomist clearance for chemical blocks, a working WhatsApp path, real farmer questions for speech recognition measurement, and a staff safety rehearsal. Each gate addresses a different way a seemingly helpful answer can fail before it reaches a farmer.
The pilot is built to measure useful outcomes, including successful answers or correct escalations, comprehension, repeat use, response time, and the cost of each completed question. It also has a hard safety measure: no unsafe pesticide answers.
Better local-language access starts with a clear boundary
Local-language communication can bring essential agricultural information closer to the people making decisions in the field. Fourteen cocoa cooperatives across eight Ghanaian regions have been trained to create participatory local-language radio programmes on subjects including climate change, cocoa prices, input costs, and labour costs. The format matters because information is more useful when people can hear it in language that fits their daily lives.
A voice service can support that same goal only when it is honest about what it knows and what it does not.
For Kwaku, the useful answer is not a confident sentence built around an absent interval. It is a clear explanation that the record does not yet support a safe date, followed by a route to the person who can confirm one. That approach also matters for uncertain crop symptoms, as in The Pod Mark a Farmer Shouldn’t Treat on a Guess.
When the extension officer replies, Kwaku can return to the plot with a verified instruction rather than a guess spoken in a familiar voice.
Comments
No comments yet.