A cocoa farmer who hears at 6 a.m. that the season may start late should not receive advice built on that rumour. AgriVoice must identify what has been verified, refuse to guess at missing facts, and send uncertain questions to a named extension officer before suggesting any next step.
In August 1854, physician John Snow faced a similarly dangerous gap between what people believed and what the evidence could support. Cholera was spreading through Soho, London. The prevailing explanation blamed foul air, but Snow suspected contaminated water. Suspicion alone was not enough to justify public action.
He mapped deaths around the Broad Street water pump and investigated where affected households obtained their water. The evidence pointed toward the pump, though Snow could not yet identify the biological agent responsible. He presented his findings to local authorities, who removed the pump handle.
Snow later documented the investigation in the 1855 edition of On the Mode of Communication of Cholera. The outbreak was already subsiding when the handle was removed, so historians still debate how much that specific intervention changed its course. The lasting lesson is narrower and more useful: before recommending an action, separate the claim, the available evidence and the uncertainty that remains.
That is exactly the discipline a season-start rumour requires.
A rumour can become a costly instruction
At 6 a.m., the message may sound simple: the cocoa season could begin late.
By breakfast, it may have travelled through calls, voice notes and conversations between farmers, buyers and households. Each retelling can make the claim sound firmer. “May start late” becomes “will start late.” A possibility becomes a planning assumption.
That shift matters. A farmer might delay preparing sacks, rearrange labour, hold cocoa longer than intended or make commitments to a buyer. The appropriate decision depends on facts the rumour does not provide. Who made the announcement? Does it apply nationally or only in one purchasing area? Is it official policy, an industry proposal or one person’s expectation? When was the information issued, and has it changed?
A voice assistant can make the danger worse if it responds fluently without resolving those questions. Spoken advice sounds immediate and personal. In Asante Twi, it may also be easier to understand and act on than a written notice. That increases the responsibility to verify the substance before speaking.
The same concern appears when a buyer may leave before a farmer receives an answer. Kwaku’s buyer may leave. His cocoa sale still needs a verified answer. Speed matters, but a quick unsupported answer can create a larger loss.
AgriVoice must keep three things separate
The first item is the farmer’s report: “I heard the season may start late.” That is a useful question, not confirmed policy.
The second is reviewed guidance. AgriVoice is designed to select from approved content blocks rather than ask a language model to compose agricultural or purchasing advice. If a reviewed block addresses the verified announcement and its scope, the system can select that block and speak it in Asante Twi.
The third is everything still unknown. If the available content does not confirm the rumour, AgriVoice should say that it cannot verify the claim and escalate the question to the partner’s named extension officer. It should preserve the original question so the human receives the issue the farmer raised, without converting uncertainty into fact.
This is constrained selection in practical terms. The model chooses reviewed material. It does not fill gaps with plausible wording.
The approach also protects against a subtler failure: answering a nearby question instead of the actual one. General guidance about preparing for the season does not establish when the season starts. Advice about storing cocoa does not prove that a delay has been announced. A response can be accurate in isolation and still mislead the farmer if it implies confirmation.
Verification must come before the next step
A safe exchange should establish the claim’s status before moving to action.
AgriVoice may need to distinguish an official announcement from a community report, check whether a reviewed answer exists, and identify whether location or timing changes the answer. If those checks do not resolve the question, escalation is the answer the system can safely provide.
The current AgriVoice pilot is being prepared around that boundary. Before farmer exposure, a cocoa-sector partner must own recruitment and name the extension officer who receives escalations. The spoken Asante Twi content must be reviewed by a farming-aware translator. Staff must also rehearse the complete workflow and confirm that uncertain questions reach a person without an unsafe answer slipping through.
The pilot scorecard treats a correct escalation as a successful answer. That choice matters. A system that says “I cannot verify this yet; your question has been sent for review” may feel less impressive than an instant recommendation. It is more useful than confident fiction.
For a closer look at this operating model, see Ghana Cocoa Advice Pilot: Why Human Escalation Makes AgriVoice Safer.
The reply should preserve uncertainty
Snow acted from evidence while remaining clear about what he had not proved. AgriVoice needs the same restraint when a morning rumour arrives before a verified update.
The first response should preserve the distinction between hearing and knowing. It can acknowledge the report, explain that the system has no reviewed confirmation, and route the question to the responsible extension officer. Only after verification should it suggest a season-specific next step.
That sequence gives the farmer something concrete at 6 a.m.: no false confirmation, no invented deadline and no advice resting on a rumour. The next action belongs to the verification path.
Comments
No comments yet.