A blue tick should confirm that a farmer’s question arrived, not suggest that the answer is safe. For a spraying decision, the responsible promise is reviewed Asante Twi guidance or a clear handoff to a named extension officer, never an instant guess.
Imagine Kojo, a cocoa farmer outside Kumasi, standing beside a half-filled spray tank late in the afternoon. His gloves are tucked into one pocket. Rain clouds are gathering, and two workers are waiting for him to decide whether they should finish spraying before dark.
Kojo sends a voice note on WhatsApp. He asks about the correct dosage and how long workers must stay out of the field afterward.
The message turns blue.
Nothing about that colour tells him whether the advice has been checked. Yet the decision cannot remain open for long. If he guesses, workers could re-enter too soon. If he abandons the mixture, he may lose the chemical already prepared. If he sprays with the wrong amount, both the crop and the people handling it could be exposed to harm.
This is the moment when speed can become dangerous.
A fast answer can still be the wrong answer
A general AI system can produce a fluent response in seconds. Fluency creates its own kind of pressure: a complete sentence sounds settled, even when the underlying facts are missing.
Chemical guidance exposes that risk sharply. A useful answer may depend on the product, dosage, re-entry interval, pre-harvest interval, crop condition, and details omitted from the first question. Filling those gaps with plausible language would turn uncertainty into instruction.
Translation adds another failure point. During Neuralis testing, machine translation changed the Twi word for cocoa into “chicken.” The sentence could still sound confident. The subject had changed completely.
AgriVoice takes a narrower path. Its reasoning system selects from reviewed content blocks instead of composing new agronomy advice. Chemical blocks remain withheld when dosage or safety intervals have not been cleared by a Ghanaian agronomist. That refusal is a safety control.
The same principle appears in the half-filled spray tank AgriVoice must refuse to complete: when the verified answer runs out, the system must stop speaking with authority.
Reviewed Twi changes what the blue ticks mean
For Kojo, a useful response must do more than reach his phone. He needs to understand it well enough to act without translating technical English in his head while two workers wait beside the tank.
That requires spoken Asante Twi reviewed by someone who knows farming vocabulary. The wording must sound natural aloud. Terms farmers commonly use in English may stay in English. Advice that reads incorrectly must be queried, rather than quietly rewritten by a translator.
Neuralis is preparing 37 owner-approved guidance blocks for this process, with three chemical blocks still requiring agronomist clearance. The product should remain silent on those held instructions until dosage, re-entry, and pre-harvest details are verified.
This restraint matters because spoken advice feels personal. A voice in a familiar language can reduce the distance between information and action, but it can also make a bad answer easier to trust. Review must come before warmth.
The pilot therefore measures more than whether the system replies. It will look for correct answers or correct escalation, farmer comprehension, repeat use, response time, and any unsafe pesticide instruction. One unsafe pesticide answer is enough to trigger a stop or redesign decision.
Escalation needs a person’s name
A safe refusal still leaves Kojo beside the tank.
The next step cannot be a vague promise that someone will follow up. Before farmers enter the pilot, a cocoa-sector partner must name the extension officer who owns escalations. The path must work during a staff safety rehearsal, and unresolved questions must remain visible until a person handles them.
In Kojo’s imagined afternoon, the system finds no cleared block that matches his chemical question. It tells him in spoken Twi that it cannot provide the instruction safely and sends the case to the named officer.
The uncertainty remains. The rain is closer, and the mixture is still waiting.
Then the officer receives the question, checks the missing details, and gives Kojo a verified next step before anyone returns to the field. The turn comes from accountable human judgment, not a more adventurous model response.
That handoff is the practical difference between “I don’t know” and “this question now belongs to someone.” Human escalation makes AgriVoice safer only when the owner, route, and response expectations are defined before the first farmer asks for help.
The pilot promise is deliberately narrow
AgriVoice is being prepared for a two-week pilot with 20 to 50 Asante-Twi-speaking cocoa farmers. It is a learning gate, not a public scale launch.
Before farmer exposure, the project still requires a committed cocoa-sector partner, reviewed Twi content, agronomist clearance for the held chemical blocks, a working WhatsApp path, real farmer-speech evaluation, tested consent and deletion behaviour, and a complete safety rehearsal.
Those gates keep the promise honest. Farmers should receive reviewed guidance they can understand, or reach a person responsible for the question. Anything broader would run ahead of the evidence.
Back beside the tank, Kojo puts the chemical down while the question is checked. His workers do not enter the sprayed area on the strength of a sentence assembled from uncertainty. The blue ticks mark the beginning of responsibility, and a named person carries it to the answer.
Comments
No comments yet.