Kwabena’s Pesticide Question. His Crew Is Ready to Spray.

When a pesticide question arrives minutes before spraying, the safest voice system should escalate uncertainty to a named extension officer instead of offering a plausible instruction. A fast, accountable handoff can prevent unverified dosage, re-entry, or pre-harvest guidance from reaching the field.

At 6:12 a.m., Kwabena stands beside his cocoa farm with a cracked phone in one hand and a product container in the other. He is a hypothetical composite, but the decision in front of him is concrete: workers are already filling their knapsack sprayers, and he cannot remember how long everyone should stay out of the treated area.

He records the question in Asante Twi. The crew waits, but the water is already mixed. If the answer sounds confident, they may begin spraying before anyone checks whether the instruction fits that product and use.

The dangerous comfort of a plausible answer

A language model can produce advice that sounds complete. That is precisely the risk.

Agricultural safety depends on details such as the named product, dosage, crop, application conditions, re-entry interval, and pre-harvest interval. A smooth sentence cannot replace a missing fact. Translation can also introduce quiet errors. During Neuralis testing, the Twi word for cocoa, “kokoo,” came back as “chicken.” The sentence remained fluent while its meaning failed.

For Kwabena, the bad ending is already on the table. The crew could spray using an unverified rate, return to the field too soon, or treat cocoa too close to harvest. Nobody standing beside the sprayers has time to inspect a model’s reasoning trace.

AgriVoice therefore uses a constrained path. The language model selects from reviewed content blocks; it does not write agronomy advice from scratch. Chemical blocks remain unavailable when a Ghanaian agronomist has not verified dosage, re-entry, and pre-harvest information.

That refusal may feel less impressive than an instant answer. It is also the more useful response when the system does not know enough.

A handoff needs a person, a clock, and context

At 6:13 a.m., AgriVoice cannot find an approved block that safely resolves Kwabena’s question. It says the advice requires human review and routes the case for escalation.

This is the turn, but only if the escalation reaches someone accountable.

“Ask an expert” leaves Kwabena exactly where he started. A working handoff needs a named extension officer, the farmer’s original question, the selected topic, and a clear operating expectation for the response. Neuralis requires the cocoa-sector pilot partner to name that officer before farmers are exposed to the system. The pilot scorecard also measures whether escalations receive a response in less than one working day.

That target still does not make every urgent question answerable before a crew begins work. The safe immediate instruction is to pause the uncertain activity until reviewed guidance arrives. Speed matters, but speed cannot turn an unknown pesticide interval into a known one.

This principle sits at the centre of why AgriVoice tells a farmer to wait before spraying. Safety comes from controlling what the system may say, preserving mandatory warnings, and giving uncertainty somewhere responsible to go.

The pilot must test the whole response chain

A laboratory can confirm that speech becomes text and text selects a content block. Kwabena’s morning exposes the harder questions.

Did the system understand field-recorded Twi rather than clean read speech? Did the spoken response make sense over farm noise? Did the escalation reach the assigned officer? Could the officer see enough context to answer safely? Did the reply arrive while it could still affect the decision?

Neuralis plans to test that full chain with 20 to 50 Asante-Twi-speaking cocoa farmers over two weeks. Before exposure, at least 20 real farmer questions must be recorded and transcribed for speech-recognition measurement. The 37 owner-approved content blocks need spoken Asante Twi reviewed by a farming-aware translator. Three held chemical blocks require a Ghanaian agronomist’s decision. Staff must also rehearse the complete workflow without an unsafe answer.

During the pilot, one unsafe pesticide response is enough to trigger a stop or redesign decision. Successful handling includes both correct answers and correct escalations. That distinction matters. A system that recognizes its boundary can protect a farmer more effectively than one that answers every question.

The related six safety gates behind a pesticide response show why channel access, reviewed language, expert content, speech measurement, escalation operations, and staff rehearsal belong in the same release decision.

What changes at the edge of uncertainty

Kwabena lowers the product container. The crew stops filling the next sprayer and waits for reviewed guidance. There is no dramatic diagnosis and no invented dosage, only a clear boundary and a question placed with the person responsible for answering it.

That changed state is modest by design. The uncertain instruction did not reach the field.

For teams building voice AI in agriculture, the practical work starts before the first farmer sends a message. Name the human who receives uncertain cases. Define what context reaches them. Hold every chemical answer that lacks verified dosage and interval guidance. Then rehearse the 6:12 a.m. moment, when a plausible sentence would be quick, useful-looking, and unsafe.