NeuralisNeuralis
← All posts

Abena’s bottle has a different concentration. Guessing could harm her cocoa.

A pesticide dosage must be checked against the exact product and concentration on the bottle. When those details are uncertain, AgriVoice is designed to hold back the instruction and route the question to a named extension officer instead of repeating remembered guidance.

At first light, Abena stood beside her cocoa plot with a blue sprayer at her feet and a newly bought bottle in her hand. The brand name looked familiar. Her neighbour had used that name last season, and Abena still remembered the mixing advice he gave her beside the road.

Then she turned the bottle around.

The label carried a different concentration from the one she expected. The crop could lose more pods if the disease spread, but guessing the mixture could harm Abena, anyone returning to the plot, and the cocoa she hoped to harvest. Her neighbour’s remembered answer suddenly had a dangerous gap in it.

This is an illustrative composite, but the problem is real enough to design around: a familiar product name can conceal a changed formulation, a different concentration, or instructions that require a qualified review.

A familiar name can hide a different instruction

Farmers build useful knowledge over years. They remember which treatment helped after heavy rain, which shop stocked it, and how a previous bottle was mixed. That experience matters.

A product name alone cannot settle a dosage question. Labels can vary by formulation and concentration. The safe answer depends on what is actually in the farmer’s hand, along with the approved use instructions and the conditions that matter for safe application.

That is why a voice system should not treat “I used this product before” as enough information to provide a mixing instruction. A confident answer based on an incomplete match creates the worst kind of error: one that sounds familiar enough to follow.

For pesticide questions, missing information is information. It should change the next step.

The moment a safe conversation has to stop

Abena’s question sounds simple: “How much should I add to my sprayer?”

A system can recognize the intent. It can also recognize that the answer depends on details it does not have. If the concentration on the bottle differs from the earlier product, remembered guidance may be wrong. If the label is unclear, the uncertainty grows.

In AgriVoice’s planned cocoa-farmer workflow, the language model selects reviewed advice blocks rather than writing agronomy advice on its own. Chemical guidance has an even stricter boundary. Three held blocks remain unavailable because dosage, re-entry, and pre-harvest details have not been verified by a Ghanaian agronomist.

That restraint is deliberate. The system should never fill a safety gap with fluent Twi.

Instead, a question that cannot be answered safely should be escalated to the partner’s named extension officer. The pilot requires that person to be accountable for the escalation queue before farmers are exposed to the service. The goal is a useful answer or a correct handoff, never a quick answer at any cost.

For a related example of why an uncertain bottle must remain unopened, see Pesticide Escalation: Why Ama’s Bottle Remains Unopened Until a Safe Answer.

What the escalation needs to capture

A handoff only helps if the extension officer receives enough context to assess the question. “What dosage?” leaves too much work for a follow-up conversation, especially when a farmer is already standing in the field ready to spray.

The useful details are practical:

  • The product name and the concentration shown on the bottle.
  • What problem the farmer is trying to treat.
  • Whether the label is readable and available to share.
  • The crop stage and the timing pressure the farmer is facing.
  • Any missing safety information, including when it is safe to re-enter the field and what interval applies before harvest.

The point is to make the uncertainty visible. A farmer should not have to translate a label into technical language before asking for help. The voice workflow can collect the question in Asante Twi, preserve the relevant details, and direct the case to the person responsible for resolving it.

A safer next step than memory

Abena did not pour anything into the sprayer that morning. She set the bottle beside the label, kept the cap sealed, and asked for review of the exact formulation she had bought.

That choice may feel slower in the moment. It protects against a mistake that is much harder to undo once spraying has started.

AgriVoice is being prepared for a two-week pilot with cocoa farmers, reviewed Twi content, a cocoa-sector partner, and a named extension officer for escalations. Until the required chemical guidance is cleared by an agronomist, pesticide questions that depend on unverified dosage must remain unanswered by the system.

The practical rule is simple: when the bottle changes, check the label again. A remembered dose belongs to the product it came from, not automatically to the one in your hand.

Neuralis

AgriVoice helps Asante-Twi-speaking cocoa farmers ask farming questions by voice and receive answers assembled only from agronomist-reviewed content, with human escalation when the system is unsure.

Try Neuralis

Comments

No comments yet.