Kojo’s Unconfirmed Cocoa Price. His Farm Input Money Is at Risk.

A petition signed before the cocoa price was announced reveals a costly information gap: farmers may have to make seasonal decisions before policy details are clear. AgriVoice should route those questions to an accountable institution, then deliver the verified response in Asante Twi, rather than generate an answer that merely sounds plausible.

Consider Kojo, an illustrative composite of a young cocoa farmer near Kumasi. At 6:20 one morning, he stands beside a stack of empty fertilizer sacks, holding a phone with a voice note from another farmer. The Young Cocoa Farmers Association has petitioned President John Dramani Mahama for urgent clarity on key policies, but the message circulating among Kojo’s friends presents one rumour as settled fact.

He needs to decide whether to commit scarce money to hired labour before the season advances. If the rumour is wrong, he could spend money he needs for farm inputs. If he waits too long, the workers may accept another job.

The price has not been announced. The policy details remain uncertain. Yet his decision cannot sit untouched forever.

A fluent answer can still be the wrong answer

Kojo asks his phone in Asante Twi: “What will the new price be, and when will farmers receive it?”

A general AI system could produce a tidy reply. It might infer an amount from an earlier season, repeat an unverified report, or turn a political expectation into a definite statement. Spoken aloud in Kojo’s language, that answer could feel more authoritative than the evidence behind it.

That is precisely where AgriVoice must stop.

Its reasoning system is designed to select reviewed content blocks, not compose agricultural guidance from scratch. The same discipline should govern policy questions. When no approved answer exists, the safe result is an escalation: the question goes to the cocoa-sector partner or named extension officer responsible for answering it.

The response to Kojo should be plain: the current information does not confirm the price. His question has been sent to the responsible person.

That sentence may sound less impressive than a confident prediction. It gives him something more useful: a clear boundary between verified information and speculation.

Accountability belongs inside the workflow

Routing a question to an institution means more than displaying “please contact support.” Someone must own the queue, know what evidence counts, and respond within an agreed period.

For the planned AgriVoice pilot, that owner must be named before farmers take part. The pilot scorecard calls for a median escalation response below one working day. If nobody is accountable, an escalation becomes a polite dead end.

Policy questions make this requirement visible. A language model cannot announce a government decision, interpret an unpublished policy, or promise that a farmer qualifies for support. The institution with responsibility for the programme must provide the answer. AgriVoice can carry that answer back through a familiar voice channel and preserve the distinction between approved guidance and an unresolved question.

This is the same principle explored in The Cocoa Policy Answer AI Must Not Write, and What Farmers Risk: uncertainty should remain visible until an accountable source resolves it.

The farmer needs a usable answer, not a clever one

Later that morning, Kojo hears a short reply: the price remains unconfirmed, so he should avoid treating the circulating figure as official. His question has been recorded for follow-up by the designated officer.

His labour decision is still difficult. The system has not removed the uncertainty, and it should never pretend otherwise. But Kojo now knows which part of the message is fact, which part is rumour, and who owes him the next answer.

That changes his next move. He postpones the full commitment and arranges only the work he can afford under either outcome. The feared bad ending, spending his input money on the strength of an invented price, remains avoidable.

A useful escalation also creates evidence for the institution. If several farmers ask versions of the same question, the partner can see where communication has failed and prepare one reviewed response for future callers. The voice workflow becomes a listening channel as well as an answering channel.

The planned two-week pilot will test whether that process works with 20 to 50 farmers. Success includes correct escalation, repeat use, understandable speech, and a queue that an identified person actually clears. What Can One Extension Officer Learn From a Two-Week AgriVoice Pilot? examines what those questions can reveal.

Build the handoff before inviting the question

Before the first farmer joins the pilot, Neuralis and its cocoa-sector partner need to agree on the policy boundary. They should list which questions have reviewed answers, which require institutional confirmation, and which fall outside the service entirely.

They also need a simple operating record: when Kojo asked, what the speech system understood, why the question was escalated, who accepted it, and when an approved response returned. That record makes delays and unsafe shortcuts visible.

Back near the empty fertilizer sacks, Kojo plays the reply once more before calling the labour organiser. He has no invented price in his ear. He has a cautious decision he can explain, and a named path to the answer still owed.