A safe agricultural voice service needs an accountable person ready to receive unresolved questions before the first farmer uses it. If nobody owns the escalation queue, uncertainty can sit unanswered while a farmer decides whether to spray, wait, or seek advice elsewhere.
Imagine Adwoa, a composite extension officer in Ghana’s cocoa belt, standing beside her motorbike at dawn. Her helmet hangs from one hand. In the other, her phone shows a farmer’s voice question about a chemical treatment: the product is named, rain may be coming, and the farmer wants to enter the field again afterward.
AgriVoice cannot find a reviewed answer that covers the dosage, re-entry interval, and pre-harvest interval. The system refuses to fill the gaps and sends the case to Adwoa.
She has not reached the first farm yet. Her queue has already opened.
The first unresolved question defines the real service
The easy demonstration ends with a spoken answer. The harder and more important test begins when the system cannot answer safely.
Adwoa listens again. The farmer sounds ready to act. If she stays silent, he may spray using an assumption carried over from another product. If she answers from memory and gets the interval wrong, someone could return to the field too soon. The bad ending remains possible because the clock now belongs to a human operation, not a model.
This is why Neuralis is preparing AgriVoice around reviewed content and explicit escalation. The language model selects approved content blocks. It does not compose agronomy advice. Questions that fall outside those blocks, contain uncertainty, or require missing pesticide details must reach a named extension officer.
That named owner matters more than a generic promise of “human support.” Someone must know that the queue is theirs, when they are expected to check it, and how the farmer receives the eventual response.
Before any farmer joins the planned pilot, a cocoa-sector partner must name that person.
Safe refusal still creates work
A refusal can prevent a confident wrong answer. It cannot complete the farmer’s job.
The farmer still needs to decide what to do with the chemical container beside him. Adwoa still needs enough context to respond: what product was mentioned, what question was asked, and why the automated path stopped. The partner still needs a way to see whether cases are waiting too long.
The system’s responsibility is to preserve uncertainty instead of disguising it. The institution’s responsibility is to resolve that uncertainty.
This division becomes especially important for pesticide questions. Three AgriVoice content blocks remain held because verified dosage, re-entry, and pre-harvest information is incomplete. AgriVoice currently refuses those questions by design. A Ghanaian agronomist must review the missing details before those blocks can be released.
The same reasoning appears in why AgriVoice tells a farmer to wait before spraying. Waiting protects the farmer only when a qualified person takes responsibility for what happens next.
Measure the queue before celebrating the voice
In the imagined scene, Adwoa calls the farmer with one instruction: do not proceed until the product guidance has been checked. She then routes the question to the agronomy contact responsible for the held content.
The immediate danger pauses. The case remains open.
That distinction belongs in the pilot scorecard. Neuralis plans to count a question as successfully handled when it receives an approved answer or reaches the correct escalation path. The pilot also requires a named escalation owner and targets a median human response time of less than one working day.
Those measures expose institutional performance. A voice can sound clear while unresolved cases pile up. An answer can arrive quickly while omitting a mandatory safety detail. A dashboard can show activity while no one knows who should call the farmer back.
The planned two-week pilot with 20 to 50 farmers is meant to surface those failures on a small, inspectable scale. It will track safe handling, comprehension, repeat use, response time, latency, and cost per completed question. Any unsafe pesticide instruction reaching a farmer triggers a stop or redesign decision.
The relevant unit of value is a completed question, including the human work required when automation reaches its boundary.
Open the queue during rehearsal
Later that morning, Adwoa reaches the farm. The farmer has left the container closed. She confirms that the case is still being checked, then marks it for follow-up rather than treating the pause as a completed answer.
That changed detail is modest: nothing was sprayed while the advice remained uncertain. It is also the result the whole workflow must protect.
Before field exposure, staff should rehearse this exact moment with ambiguous, out-of-domain, pesticide, code-switched, and speech-recognition-corrupted questions. Each test should reveal who receives the case, what context they see, how quickly they acknowledge it, and how the final response returns to the farmer.
If any case can sit ownerless, the pilot gate remains closed.
The first morning should begin with Adwoa checking an empty, working queue. Then the farm gate can open.
Comments
No comments yet.